Dynamic adaptive operation and maintenance data intelligent prediction system, method and server

Abstract

The application relates to a dynamic self-adaptive operation and maintenance data intelligent prediction system, method and server, and belongs to the technical field of computers, which comprises a data acquisition module, a model training module, a model evaluation module, a period adjustment module and a prediction execution module.The data acquisition module is used for determining a time range parameter of to-be-predicted operation and maintenance data, and dynamically acquiring target operation and maintenance data based on the time range parameter.The model training module is used for adopting multiple candidate regression algorithms, and performing model training on the target operation and maintenance data to obtain multiple trained candidate models.The model evaluation module is used for performing performance evaluation on the multiple trained candidate models based on preset evaluation indexes, so as to determine a best prediction model.The period adjustment module is used for adaptively adjusting a training period of subsequent model training according to a time range parameter corresponding to the training of the best prediction model.The prediction execution module is used for using the best prediction model to predict new operation and maintenance data received in real time to obtain a prediction result.The system can improve the accuracy of a data prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology, and in particular to a dynamic self-adaptive operation and maintenance data intelligent prediction system, method and server. BACKGROUND

[0002] In the information age, the scale and complexity of information systems are increasing, and the operation and maintenance data generated during their operation presents the characteristics of massiveness, multidimensionality and dynamic change. Timely and accurate prediction of these operation and maintenance data is of great significance for early detection of potential system failures, optimization of resource allocation and guarantee of stable system operation.

[0003] In the prior art, a common operation and maintenance data prediction scheme is as follows: operation and maintenance data of a specific time period is extracted from a database according to a preset fixed period (for example, daily), and then one or more pre-selected algorithms (for example, a time series model or a simple regression model) are used to train the data, and the trained model is used to predict future data trends.

[0004] However, in actual application, the inventors have found that the above prior art at least has the following technical problems:

[0005] Fixed and rigid data acquisition method: most existing methods extract operation and maintenance data according to a preset fixed period and time range, and it is difficult to flexibly adjust the time granularity of data (such as monthly, daily, weekly) according to actual analysis requirements. When operation and maintenance personnel need to analyze operation and maintenance trends of different time spans, they must manually modify the data extraction rules, which is tedious and inefficient, and cannot quickly respond to diversified prediction requirements.

[0006] Poor model training adaptability and low prediction accuracy: existing methods often use a single algorithm model for training, while operation and maintenance data often has complex nonlinearity, periodicity and dynamics, and a single algorithm cannot fully fit all data patterns, resulting in limited prediction accuracy. Even if some schemes use multiple algorithms, they usually lack a mechanism for dynamically selecting the best model according to the current data characteristics, thereby affecting the final prediction effect.

[0007] Fixed training period and unreasonable resource allocation: the model training period of existing methods is usually a preset fixed value, such as daily or weekly. This fixed strategy cannot be dynamically adjusted according to the performance of the model on different time granularity data. For example, if a model performs best on monthly data, but the system still trains it daily, unnecessary computing resources will be wasted; conversely, if the model performs better on daily data, but the system trains it weekly or monthly, the model update may be lagging, resulting in decreased or failed prediction accuracy.

[0008] Therefore, how to realize an operation and maintenance data prediction method capable of dynamically obtaining data of different time granularities, adaptively selecting an optimal prediction model, and intelligently adjusting a model training period to improve prediction accuracy and resource utilization is a technical problem to be solved by those skilled in the art. SUMMARY

[0009] The present application aims to provide a dynamic and adaptive operation and maintenance data intelligent prediction system, method and server to solve the problems existing in the prior art. The technical problems to be solved by the present application are solved by the following technical solutions.

[0010] According to a first aspect of the present application, a dynamic and adaptive operation and maintenance data intelligent prediction system is provided, comprising:

[0011] A data acquisition module is configured to determine a time range parameter of to-be-predicted operation and maintenance data according to prediction requirements, and dynamically acquire target operation and maintenance data based on the time range parameter.

[0012] A model training module is configured to use a plurality of candidate regression algorithms to perform model training using the target operation and maintenance data, thereby obtaining a plurality of trained candidate models.

[0013] A model evaluation module is configured to perform performance evaluation on the plurality of trained candidate models based on a preset evaluation index, thereby determining an optimal prediction model.

[0014] A period adjustment module is configured to adaptively adjust a training period of subsequent model training according to the time range parameter corresponding to training of the optimal prediction model.

[0015] A prediction execution module is configured to use the optimal prediction model to predict new operation and maintenance data received in real time, thereby obtaining a prediction result.

[0016] Preferably, in the data acquisition module, a connection is established with an ES system through a preset interface, and a data query statement is dynamically generated according to the time range parameter.

[0017] The ES system extracts target operation and maintenance data of the corresponding time range from the stored operation and maintenance data according to the data query statement.

[0018] Preferably, the operation and maintenance data includes, but is not limited to, server CPU usage, memory occupancy, network traffic, disk I / O, service interface response time, service interface throughput, and service interface success rate.

[0019] Preferably, the model training module comprises:

[0020] The data preprocessing submodule is configured to preprocess the target operation and maintenance data, including removing outliers, filling missing values, and normalization processing.

[0021] The algorithm selection and initialization submodule is configured to select multiple regression algorithms as candidate models and set initial parameters of each candidate model.

[0022] The training submodule is configured to divide the preprocessed target data into a training set and a validation set according to a preset ratio, and use the training set to cross-train each candidate model.

[0023] Preferably, the candidate models include, but are not limited to, linear regression, ridge regression, Lasso regression, decision tree regression, random forest regression, gradient boosting regression, and support vector regression.

[0024] Preferably, in the model evaluation module, the preset evaluation indicators include at least one of mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R²).

[0025] The determination of the best prediction model includes:

[0026] According to the evaluation indicators, the plurality of trained candidate models are ranked according to a preset priority, and the model ranked first is determined as the best prediction model.

[0027] Preferably, in the cycle adjustment module, the method of adaptively adjusting the training cycle of subsequent model training is:

[0028] If the best prediction model is trained based on monthly data, the training cycle is adjusted to once a month.

[0029] If the best prediction model is trained based on daily data, the training cycle is adjusted to once a day.

[0030] If the best prediction model is trained based on weekly data, the training cycle is adjusted to once a week.

[0031] If the best prediction model is trained based on X-day data, the training cycle is adjusted to once every X days.

[0032] Preferably, in the prediction execution module, the best prediction model processes real-time operation and maintenance data belonging to the best prediction model, and the generated prediction result includes the numerical value and trend of the operation and maintenance data in a future specific time period. The prediction result is displayed or triggers a corresponding early warning mechanism.

[0033] According to a second aspect of the present application, a prediction method using the above-mentioned dynamic adaptive operation and maintenance data intelligent prediction system is provided, including the following steps:

[0034] According to the predicted demand, a time range parameter of to-be-predicted operation and maintenance data is determined, and target operation and maintenance data is dynamically acquired based on the time range parameter;

[0035] A plurality of candidate regression algorithms are adopted, and the target operation and maintenance data is used for model training respectively to obtain a plurality of trained candidate models;

[0036] Performance evaluation is performed on the plurality of trained candidate models based on preset evaluation indexes to determine a best prediction model;

[0037] The best prediction model is used to predict new operation and maintenance data received in real time to obtain a prediction result;

[0038] The time range parameter corresponding to the training of the best prediction model is used to adaptively adjust a training period of subsequent model training.

[0039] According to a third aspect of the present application, a server is provided, comprising a processor and a memory connected with the processor in communication; the memory stores computer program instructions; when the processor executes the computer program instructions stored in the memory, the above-mentioned prediction method is realized.

[0040] Compared with the prior art, the technical solution provided by the present application has at least the following beneficial effects:

[0041] Data acquisition is flexible and efficient: by dynamically connecting with data sources such as ES, operation and maintenance data of different time ranges (month, day, week) can be flexibly acquired according to actual needs without manual intervention, greatly improving the flexibility and efficiency of data acquisition, and being able to adapt to different granularity operation and maintenance analysis requirements.

[0042] High prediction accuracy: the present application uses a plurality of regression algorithms for dynamic training, and selects the best model that best matches the current data characteristics through strict evaluation indexes (such as R², MSE, MAE), which can more fully fit the complex rules of operation and maintenance data, thereby significantly improving the accuracy of the prediction result.

[0043] Strong model adaptability and reasonable resource allocation: the present application innovatively links the data time range based on the best model with the training period to realize intelligent adaptive adjustment of the training frequency. This enables the model to keep up with the changing trend of operation and maintenance data in a timely manner, while avoiding the problem of outdated models or waste of computing resources caused by fixed training period, ensuring the long-term effectiveness and economy of the model.

[0044] The application value is remarkable: by providing more accurate and timely operation and maintenance data prediction, it can help operation and maintenance personnel to discover and locate potential system risks in advance, provide decision support for preventive maintenance, thereby effectively improving the stability and reliability of the information system, and reducing operation and maintenance cost and false alarm rate. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a schematic diagram of the architecture of a dynamic adaptive operation and maintenance data intelligent prediction system of the present application;

[0046] Figure 2 is a flowchart of a dynamic adaptive operation and maintenance data intelligent prediction method of the present application;

[0047] Figure 3 is a schematic diagram of the architecture of a server of the present application. DETAILED DESCRIPTION

[0048] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0049] It should be noted that the above detailed description is exemplary and is intended to provide further description of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0050] The dynamic adaptive operation and maintenance data intelligent prediction system in an embodiment of the present application has a whole architecture including a hardware layer, a software layer and a data flow path, wherein:

[0051] The hardware layer includes servers for deploying ES (at least with 8-core CPU, 16GB memory, 1TB storage, supporting network bandwidth above 100Mbps), model training servers (with GPU acceleration cards such as NVIDIA Tesla T4 and above models), prediction engine servers (4-core CPU, 8GB memory) and terminal devices (for users to configure parameters and view results).

[0052] The software layer is composed of a data acquisition module, a data preprocessing module, a model training module, a model evaluation module, a prediction execution module, a periodic control module and a user interaction interface, each module performs data interaction through TCP / IP protocol, Java language is used to develop the core logic, and Python is used to implement the regression algorithm model.

[0053] Data flow path: the operation and maintenance data stored in the ES is sequentially passed through the data acquisition module, the data preprocessing module, the model training module, the model evaluation module, the prediction execution module, and the monitoring device, and the periodic control module and the data acquisition module and the model training module establish a linkage control link.

[0054] In the software layer, as shown in Figure 1 the roles of each module are as follows:

[0055] The data acquisition module is configured to determine a time range parameter of to-be-predicted operation and maintenance data according to a prediction requirement, and dynamically acquire target operation and maintenance data based on the time range parameter.

[0056] In this module, the time range parameter of the to-be-predicted operation and maintenance data is determined based on a prediction requirement set by a user or triggered automatically by the system. The parameter can be specified as monthly, daily, or weekly.

[0057] The data acquisition module is configured to determine a time range parameter of to-be-predicted operation and maintenance data according to a prediction requirement, and dynamically acquire target operation and maintenance data based on the time range parameter.

[0058] The ES extracts target data in the corresponding time range from the stored operation and maintenance data according to the query statement, and transmits the target data to the data processing module. The operation and maintenance data includes, but is not limited to, server CPU usage, memory occupancy, network traffic, disk I / O, business interface response time, business interface throughput, and business interface success rate.

[0059] The model training module is configured to use a plurality of candidate regression algorithms to train a plurality of trained candidate models using the target operation and maintenance data.

[0060] In this module, the following are included:

[0061] The data preprocessing submodule cleans the target data obtained from the ES, removes outliers (such as by the 3σ principle or the box plot method), fills in missing values (such as by using the mean filling, median filling, or linear interpolation method), and performs standardization or normalization processing to convert the data into a format suitable for model training.

[0062] The algorithm selection and initialization submodule selects a plurality of regression algorithms as candidate models, including but not limited to linear regression, ridge regression, Lasso regression, decision tree regression, random forest regression, gradient boosting regression (GBDT), support vector regression (SVR), and sets the initial parameters of each candidate model.

[0063] The training submodule divides the preprocessed target data into a training set and a validation set according to a preset ratio, and uses the training set to cross-train each candidate model.

[0064] a model evaluation module configured to evaluate performance of the plurality of trained candidate models based on preset evaluation indicators to determine a best prediction model;

[0065] In this module, after training, the prediction performance of each candidate model is evaluated using the validation set. The evaluation indicators include but are not limited to mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R2), etc. The model with the optimal evaluation indicator is determined as the best prediction model for the current time range.

[0066] a prediction execution module configured to use the best prediction model to predict new operation and maintenance data received in real time to obtain a prediction result.

[0067] In this module, the best prediction model is deployed to the prediction engine, and newly generated operation and maintenance data is received in real time as input.

[0068] The prediction engine calls the best prediction model to process the input real-time operation and maintenance data to generate a corresponding prediction result. The prediction result can include the numerical value and trend of the operation and maintenance data in a future specific time period.

[0069] The prediction result is transmitted to a display module or an associated alarm system for viewing by operation and maintenance personnel or triggering of a corresponding early warning mechanism.

[0070] a cycle adjustment module configured to adaptively adjust a training cycle of subsequent model training according to a time range parameter corresponding to training of the best prediction model;

[0071] In this module, the time range based on which the best model is trained in the training process is recorded (i.e., whether the best model is trained by monthly data, daily data, or weekly data).

[0072] If the performance of the best model is optimal when trained based on monthly data, the system automatically adjusts the model training cycle to once a month; if it is optimal based on daily data, it is adjusted to once a day; if it is optimal based on weekly data, it is adjusted to once a week; if the best prediction model is trained based on X-day data, the training cycle is adjusted to once every X days.

[0073] When the set training cycle is reached, the system automatically triggers the next round of dynamic data acquisition and model training process to update the best model, ensuring that the prediction performance of the model always remains at an optimal level.

[0074] As shown in Figure 2 The embodiment discloses a prediction method of the dynamic adaptive operation and maintenance data intelligent prediction system, comprising the following steps:

[0075] S10: According to the predicted demand, determine the time range parameter of the to-be-predicted operation and maintenance data, and dynamically acquire target operation and maintenance data based on the time range parameter;

[0076] In this step, ES connection parameters need to be configured, including ES server IP address (such as 192.168.1.X), port number (default 9200), access username and password (stored in the configuration file through AES encryption), target index name (such as "operation_logs_2025");

[0077] Then, data filtering is performed: when "monthly" is selected, the time filtering condition {"range": {"timestamp": {"gte": "{yyyy}-{mm}-0100:00:00", "lte": "{yyyy}-{mm}-last_day23:59:59"}}} is automatically generated.

[0078] When "daily" is selected, {"range": {"timestamp": {"gte": "{yyyy}-{mm}-{dd}00:00:00", "lte": "{yyyy}-{mm}-{dd}23:59:59"}}} is generated.

[0079] When "weekly" is selected, {"range": {"timestamp": {"gte": "{Monday date}00:00:00", "lte": "{Sunday date}23:59:59"}}} is generated (the start and end dates of each week are calculated through the Calendar class of Java).

[0080] The data acquisition module sends a query request through RestHighLevelClient (ES official Java client), sets the timeout time to 30 seconds, and sets the maximum number of retries to 3.

[0081] After receiving the JSON format data returned by ES, the "timestamp" (timestamp), "metric_name" (metric name), and "metric_value" (metric value) fields are parsed and stored in a local temporary file (in CSV format, fields are separated by commas, and each row records one data).

[0082] Then, data preprocessing is performed, including:

[0083] Outlier treatment: read the CSV file data, perform 3σ treatment on the numerical column of each index: calculate the mean μ and standard deviation σ of the column data, and remove records with values > μ + 3σ or < μ - 3σ.

[0084] For the index whose data volume after removal accounts for more than 10%, use the box plot method for secondary verification: calculate the quartiles Q1 and Q3, determine the upper limit as Q3 + 1.5×(Q3 - Q1), and the lower limit as Q1 - 1.5×(Q3 - Q1), and remove records outside the range again.

[0085] Missing value filling: count the missing fields of each record, if the missing rate < 5%, use the mean of the first 3 and last 3 valid data of the index to fill; if the missing rate is between 5%-20%, use linear interpolation method (through the interpolate function of pandas library in Python, method=linear); if the missing rate > 20%, mark it as an invalid index and notify the user.

[0086] Standardization: perform Z-score standardization on the processed data: convert each numerical value to (x - μ) / σ, where μ is the mean of all data of the index, and σ is the standard deviation, and the processed data is stored as a Numpy array format (dimension [sample number, feature number]).

[0087] S20: use multiple candidate regression algorithms to train models using the target operation and maintenance data, obtaining multiple trained candidate models;

[0088] In this step, first establish a candidate algorithm library:

[0089] Integrate 6 regression algorithm models, each model pre-configured with 3 groups of initial parameters:

[0090] Linear regression: set the regularization coefficient λ to 0.01, 0.1, and 1.0;

[0091] Random forest regression: number of decision trees is 50, 100, 200, maximum depth is 5, 10, None;

[0092] GBDT: learning rate is 0.01, 0.1, 0.2, iteration number is 100, 200, 300;

[0093] The remaining algorithms (ridge regression, Lasso regression, SVR) are configured with multiple groups of parameters according to the same logic.

[0094] All models are packaged as callable Python classes, providing uniform train, predict method interfaces.

[0095] Then data set split: the normalized Numpy array is randomly divided into training set and validation set in the ratio of 8:2 (through the train_test_split function of the sklearn library, random_state=42 to ensure reproducibility).

[0096] Finally, model training: the model training module calls the candidate algorithm library to perform training for each algorithm and parameter combination, including:

[0097] Linear regression is optimized by gradient descent method, with 500 iterations and learning rate 0.001;

[0098] Random forest uses bootstrap sampling to construct decision trees, and randomly selects a subset of features (the square root of the total number of features) when training each tree;

[0099] All model training processes record loss value changes, and training is terminated when the loss value decreases by <0.0001 for 10 consecutive rounds.

[0100] The trained model is saved to the specified directory, along with metadata such as training duration and iteration count.

[0101] S30: Based on the preset evaluation index, the performance of the plurality of trained candidate models is evaluated to determine a best prediction model;

[0102] In this step, the model evaluation module loads the validation set data, calls each trained model to output the prediction result, and calculates three core indicators:

[0103] Mean Squared Error (MSE): MSE=1 / nΣ(y_true - y_pred)², where n is the number of validation set samples, y_true is the true value, and y_pred is the predicted value;

[0104] Mean Absolute Error (MAE): MAE=1 / nΣ|y_true - y_pred|;

[0105] Determination coefficient (R²): R²=1 - Σ(y_true - y_pred)² / Σ(y_true - y_mean)², y_mean is the mean of the true value.

[0106] The evaluation results of all models are prioritized in descending order of R², ascending order of MSE, and ascending order of MAE. The model ranked first is selected as the best model.

[0107] If there are multiple models with the same indicators, the model with the shortest training time is selected as the "current effective model" and recorded in the configuration file (model_config.json), including the model path, corresponding time range (month / day / week), and training completion time.

[0108] S40: Using the best prediction model, predict the newly received operation and maintenance data in real time to obtain the prediction result;

[0109] This step specifically includes the following sub-steps:

[0110] Predict data access: The prediction execution module listens to the Socket port (default port 8080) to receive real-time operation and maintenance data (in JSON format, including:

[0111] "timestamp" "metric_name" "metric_value" fields), and processes up to 1000 data per second.

[0112] Real-time standardization processing is performed on the received data (reuse the μ and σ values during training) to convert it into model input format.

[0113] Prediction output: The "current effective model" is called to perform prediction on the standardized data, and the predicted values of the operation and maintenance data in the next hour (every 5 minutes for 12 data points) are output.

[0114] The prediction result is converted into JSON format, including:

[0115] "predict_timestamp" "metric_name" "predict_value" "confidence" (confidence, based on the R² value of the validation set), sent to the user interface of the terminal device through HTTP POST request, and stored in the prediction result database MySQL. The table structure includes: id, predict_time, metric, value, confidence.

[0116] S50: According to the time range parameter corresponding to the training of the best prediction model, adaptively adjust the training period of subsequent model training;

[0117] In this step, the cycle control module reads the time range corresponding to the best model in model_config.json:

[0118] If it is "monthly", set the training trigger time in the configuration file to 23:00 on the last day of each month;

[0119] If it is "daily", set it to 00:00 every day;

[0120] If it is "weekly", set it to 23:00 every Sunday.

[0121] Use the Linux crontab timing task mechanism to implement periodic triggering, and send instructions containing time range parameters to the data acquisition module when triggered.

[0122] After each new model training is completed, automatically retain the last 3 versions of historical models, and automatically delete more than 3 versions to release storage space.

[0123] Alarm through the ratio of predicted value and actual value, abandon the traditional single threshold trigger alarm, greatly improve the operation and maintenance efficiency.

[0124] In a specific example, the system is used for server memory usage prediction (best model determined after multi-time range training), the specific method is as follows:

[0125] 1. Multi-time range dynamic data acquisition:

[0126] Monthly data acquisition: The user sets the prediction target as "server memory usage", first selects the time range "monthly", and acquires data from January to June 2025. Generate ES query statement: {"range":{"timestamp":{"gte":"2025-01-01 0:00:00","lte":"2025-06-30 23:59:59"}},"query":{"term":{"metric_name":"memory_usage"}}}, connect ES (IP: 192.168.1.100, port: 9200) through RestHighLevelClient, acquire 182 days of 17472 records (1 per hour), and store as a CSV file.

[0127] Weekly data acquisition: Next, select the time range "weekly" to obtain data for the last two weeks of June 2025 (June 17 - June 30). The ES query statement: {"range": {"timestamp": {"gte": "2025-06-17 00:00:00", "lte": "2025-06-30 23:59:59"}}, "query": {"term": {"metric_name": "memory_usage"}}} obtains a total of 336 records for 14 days, stored as a CSV file.

[0128] Daily data acquisition: Finally, select the time range "daily" to obtain data for June 30, 2025. The ES query statement: {"range": {"timestamp": {"gte": "2025-06-30 00:00:00", "lte": "2025-06-30 23:59:59"}}, "query": {"term": {"metric_name": "memory_usage"}}} obtains 24 records, stored as a CSV file.

[0129] 2. Multi-time range data preprocessing

[0130] Monthly data processing: The 3σ principle is used for outlier processing, with a mean μ = 72% and a standard deviation σ = 10%. Remove 156 records (0.89%) with values greater than 102% and less than 42%. There are 28 missing values (0.16%), which are filled with the average of the previous and next three data points. After Z-score standardization, μ = 71.5% and σ = 9.8%. Store as a Numpy array of [17316, 1].

[0131] Weekly data processing: Remove 8 records (2.38%) with values greater than 105% and less than 45% using the 3σ principle. There are 5 missing values (1.49%), which are filled with the average. After standardization, μ = 70.2% and σ = 9.2%. Store as a Numpy array of [328, 1].

[0132] Daily data processing: There are no outliers. There are 2 missing records (8.33%), which are filled using linear interpolation. After standardization, μ = 69.8% and σ = 8.5%. Store as a Numpy array of [22, 1].

[0133] 3. Multi-time range model training and evaluation

[0134] Monthly model training: The dataset is divided by 7:3, and 3 groups of parameters are trained for each of the 6 algorithms. Random forest regression (decision tree 150, depth 15) performs best, with R²=0.91, MSE=9.2, MAE=2.3, and training time 150 seconds.

[0135] Weekly model training: After dividing the dataset, gradient boosting regression (learning rate 0.15, iterations 250) has R²=0.87, MSE=11.5, MAE=2.8, and training time 95 seconds.

[0136] Daily model training: After training, ridge regression (regularization coefficient 0.5) has R²=0.79, MSE=15.3, MAE=3.5, and training time 40 seconds.

[0137] Best model determination: Comparing the three time range models, the monthly data trained random forest regression has the highest R², the lowest MSE and MAE, and is determined as the best model, recorded in model_config.json.

[0138] 4. Real-time prediction: On July 1, 2025, 10:00, the real-time memory usage rate is 75%, after standardization (75-71.5) / 9.8≈0.36, the best model is called to predict the memory usage rate from 10:05 to 11:00 on July 1, with a prediction value of 76.5% at 10:40 and a confidence of 0.91, and the result is synchronized to the terminal and database.

[0139] 5. Training period adjustment: Since the best model is based on monthly data, set the last day of each month at 23:00 to trigger the next round of training.

[0140] In another specific example, the system is used for database connection prediction (best model determined after multi-time range training), with the following specific methods:

[0141] 1. Multi-time range dynamic data acquisition

[0142] Monthly data: Get database connection data from March to August 2025, and the ES query statement corresponds to the time range, resulting in 184 days and 17664 records (1 per hour).

[0143] Weekly data: Get data from the first week of August 2025 (August 1 to August 7), a total of 168 records.

[0144] Daily data: Get data for August 8, 2025, a total of 24 records.

[0145] 2. Multi-time range pre-processing and training

[0146] Monthly data post-processing and training, support vector regression (kernel function rbf, C=10) R²=0.89, MSE=8.7, MAE=2.2.

[0147] Weekly data post-training, random forest (decision tree 100, depth 10) R²=0.88, MSE=9.1, MAE=2.4.

[0148] Daily data post-training, decision tree regression (depth 8) R²=0.81, MSE=13.6, MAE=3.2.

[0149] 3. Best model determination: compared with the three time range models, the random forest model trained by the weekly data is slightly better than the monthly data, and is selected as the best model. The best model is used for data prediction.

[0150] 4. Real-time prediction and periodic adjustment: based on the weekly best model prediction, set the trigger training at 23:00 every Saturday.

[0151] In a third specific example, the system is applied to predict the application response time (multi-time range training to determine the best model), and the method is as follows:

[0152] 1. Obtain the application response time data of February to July 2025 (monthly), the last two weeks of July 2025 (weekly), and July 31, 2025 (daily).

[0153] 2. After pre-processing and training, the gradient boosting regression (learning rate 0.1, iteration 200) trained by the weekly data has R²=0.89, the monthly model has R²=0.88, and the daily model has R²=0.78. The weekly model is the best.

[0154] 3. Trigger training according to the weekly period (every Sunday 00:00), and predict the application response time in real time.

[0155] As shown in Figure 3 The server in an embodiment of the present application includes a processor 202 and a memory 201 connected to the processor 202 in communication; the memory 201 stores computer program instructions; when the processor 202 executes the computer program instructions stored in the memory 201, the above-mentioned prediction method is realized.

[0156] Compared with the prior art, the technical scheme provided by the present application has at least the following beneficial effects:

[0157] Data acquisition is flexible and efficient: through dynamic connection with data sources such as ES, the operation and maintenance data of different time ranges (month, day, week) can be flexibly obtained according to actual needs without manual intervention, greatly improving the flexibility and efficiency of data acquisition, and adapting to different granularity operation and maintenance analysis requirements.

[0158] High prediction accuracy: the present application adopts multiple regression algorithms for dynamic training, and selects the best model that best matches the current data characteristics through strict evaluation indicators (such as R², MSE, MAE), which can more fully fit the complex rules of operation and maintenance data, thereby significantly improving the accuracy of the prediction results.

[0159] Strong model adaptability and reasonable resource allocation: the present application innovatively links the data time range based on the best model with the training period, realizing intelligent adaptive adjustment of the training frequency. This enables the model to keep up with the changing trend of operation and maintenance data, while avoiding the problem of outdated models or waste of computing resources caused by fixed training period, ensuring the long-term effectiveness and economy of the model.

[0160] Significant application value: by providing more accurate and timely operation and maintenance data prediction, it can help operation and maintenance personnel to discover and locate potential system risks in advance, provide decision support for preventive maintenance, thereby effectively improving the stability and reliability of information systems, reducing operation and maintenance costs and false alarm rate.

[0161] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise. In addition, it should be understood that when the terms "comprise" and / or "include" are used in the specification, they refer to the presence of a feature, step, operation, device, component, and / or combinations thereof.

[0162] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the terms used in this way can be interchanged as appropriate, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0163] In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0164] For purposes of the description hereinafter, the terms "upper", "lower", "right", "left", "rear", "front", "vertical", "horizontal", and derivatives thereof shall relate to the application as oriented in the drawing figures. However, it is to be understood that the application can assume various alternative orientations and, accordingly, such terms are not to be taken as limitations of the present application. Additionally, terms such as "first", "second", "third", etc. are used herein for purposes of description and do not necessarily have an ordinal meaning unless specifically stated.

[0165] In the detailed description above, reference is made to the accompanying drawings, which form a part hereof. In the drawings, similar symbols typically identify similar components, unless context dictates otherwise. The illustrative embodiments described in the detailed description, drawings, and claims are not meant to be limiting. Other embodiments can be used, and other changes can be made, without departing from the spirit or scope of the subject matter presented herein.

[0166] The above description is embodied in the form of preferred embodiments only and is not intended to limit the present application. The present application can be variously changed and substituted, without departing from the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of the present application.

Claims

1. A dynamically adaptive intelligent prediction system for operation and maintenance data, characterized in that, include: The data acquisition module is used to determine the time range parameters of the operation and maintenance data to be predicted according to the prediction requirements, and dynamically acquire the target operation and maintenance data based on the time range parameters. A connection is established with the ES system through a preset interface, and a data query statement is dynamically generated according to the time range parameters. The ES system extracts the target operation and maintenance data corresponding to the time range from the stored operation and maintenance data according to the data query statement. The operation and maintenance data includes server CPU utilization, memory usage, network traffic, disk I / O, business interface response time, business interface throughput, and business interface success rate. The model training module is used to train the model using the target operation and maintenance data with multiple candidate regression algorithms to obtain multiple trained candidate models; the candidate models include linear regression, ridge regression, Lasso regression, decision tree regression, random forest regression, gradient boosting regression, and support vector regression; The model evaluation module is used to evaluate the performance of the multiple trained candidate models based on preset evaluation metrics to determine the best prediction model; wherein the preset evaluation metrics include at least one of mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). Determining an optimal prediction model includes: According to the evaluation metrics, the multiple trained candidate models are sorted according to a preset priority, and the model ranked first is determined as the best prediction model. The priority is: The model ranked first in descending order of R², ascending order of MSE, and ascending order of MAE is selected as the best prediction model. If multiple models have the same metrics, the model with the shortest training time is selected as the best prediction model. The period adjustment module is used to adaptively adjust the training period of subsequent model training based on the time range parameters corresponding to the training of the best prediction model. The prediction execution module is used to use the optimal prediction model to predict the new real-time operation and maintenance data and obtain the prediction results. The optimal prediction model processes the new real-time operation and maintenance data and generates prediction results including the values ​​and trends of operation and maintenance data within a specific future time period. The prediction results are then displayed or a corresponding early warning mechanism is triggered.

2. The dynamic adaptive intelligent prediction system for operation and maintenance data according to claim 1, characterized in that, The model training module includes: The data preprocessing submodule is used to preprocess the target operation and maintenance data, including removing outliers, filling in missing values, and normalization. The algorithm selection and initialization submodule is used to select multiple regression algorithms as candidate models and set the initial parameters of each candidate model. The training submodule is used to divide the preprocessed target data into training and validation sets according to a preset ratio, and to use the training set to perform cross-training on each candidate model.

3. The dynamic adaptive intelligent prediction system for operation and maintenance data according to claim 1, characterized in that, The specific method for adaptively adjusting the training cycle of subsequent model training in the cycle adjustment module is as follows: If the optimal prediction model is trained based on monthly data, then the training cycle will be adjusted to once a month. If the optimal prediction model is trained based on daily data, then the training cycle will be adjusted to once a day. If the optimal prediction model is trained based on weekly data, then the training cycle will be adjusted to once a week. If the optimal prediction model is trained based on X days of data, then the training cycle will be adjusted to once every X days.

4. A prediction method using the dynamic adaptive intelligent prediction system for operation and maintenance data as described in any one of claims 1 to 3, characterized in that, Includes the following steps: Based on the predicted needs, determine the time range parameters of the operation and maintenance data to be predicted, and dynamically obtain the target operation and maintenance data based on the time range parameters. Multiple candidate regression algorithms are employed, and the target operation and maintenance data are used to train the models respectively, resulting in multiple trained candidate models; Based on preset evaluation metrics, the performance of the multiple trained candidate models is evaluated to determine the best prediction model. Using the aforementioned optimal prediction model, new operation and maintenance data received in real time are predicted to obtain prediction results; The training cycle of subsequent model training is adaptively adjusted based on the time range parameters corresponding to the best prediction model during training.

5. A server, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer program instructions; When the processor executes the computer program instructions stored in the memory, it implements the prediction method as described in claim 4.

Images