Database intelligent monitoring and early warning method based on large model and multi-dimensional perception data

Abstract

This application discloses a database intelligent monitoring and early warning method based on large models and multi-dimensional sensing data, belonging to the field of power data technology. The method includes: combining historical global sensing monitoring data sequences to output a predicted fault risk index; configuring an adaptive monitoring and early warning sliding window within a preset time period; optimizing multi-dimensional sensing strategies to obtain an adaptive sensing factor cluster; and performing intelligent monitoring and early warning of cascading avalanche faults on the power database within the preset time period. This solves the technical problem of existing database monitoring and early warning methods relying on threshold judgments of a single fixed indicator and failing to dynamically adjust according to actual operational risk conditions, resulting in lagging risk perception and low early warning accuracy and efficiency.

Description

Technical Field

[0001] This application relates to the field of power data technology, specifically to a database intelligent monitoring and early warning method based on large models and multi-dimensional sensing data. Background Technology

[0002] With the rapid development and digital transformation of the power system, the power database, as a core infrastructure, is experiencing exponential growth in the amount of data it stores, increasingly complex business relationships, and a more dynamic and changeable operating environment.

[0003] Traditional database monitoring and early warning methods often rely on threshold judgments of single or a few fixed indicators, making it difficult to comprehensively capture the complex characteristics of database operation. This limitation leads to a lag in the perception of potential risks, especially when facing faults with chain reactions and explosive outbreaks, failing to provide early warnings.

[0004] Meanwhile, existing methods lack the ability to deeply mine historical data and predict future risks. Monitoring windows and perception strategies are usually fixed and cannot be dynamically adjusted according to the actual operational risk status of the database, resulting in unreasonable allocation of monitoring resources and low accuracy and efficiency of early warning. Summary of the Invention

[0005] This application provides a database intelligent monitoring and early warning method based on large models and multi-dimensional perception data. This solves the technical problem that existing database monitoring and early warning methods rely on threshold judgment of a single fixed indicator and cannot be dynamically adjusted according to the actual operational risk situation, resulting in lagging risk perception and low early warning accuracy and efficiency.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows: On the one hand, this application provides a database intelligent monitoring and early warning method based on large models and multi-dimensional perception data, the method comprising: By combining historical global perception monitoring data sequences, the fault risk of cascading avalanche in the power database within a preset time period is predicted, and the predicted fault risk index is output. Configure the adaptive monitoring and early warning sliding window within the preset time period based on the predicted fault risk index; Based on the adaptive monitoring and early warning sliding window, multi-dimensional perception strategy optimization is performed to obtain an adaptive perception factor cluster. Based on the adaptive monitoring and early warning sliding window and the adaptive sensing factor cluster, intelligent monitoring and early warning of cascaded avalanche faults are performed on the power database within the preset time period.

[0007] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides a database-based intelligent monitoring and early warning method based on large models and multi-dimensional sensing data. First, it predicts fault risks within a preset time period by combining historical global sensing monitoring data sequences, thus proactively sensing potential cascading avalanche fault trends. Second, it configures an adaptive monitoring and early warning sliding window based on the predicted fault risk index, achieving dynamic monitoring window matching with risk levels. When the predicted risk is high, the window can be narrowed to increase monitoring density and response speed; when the risk is low, the window can be appropriately expanded to optimize resource usage, avoiding the problem of delayed early warnings in high-risk situations or resource waste in low-risk situations caused by fixed windows. Third, based on the adaptive monitoring and early warning sliding window, it optimizes multi-dimensional sensing strategies and obtains adaptive sensing factor clusters. It selects the sensing indicator combination that best matches the current risk status and window characteristics from the global sensing monitoring indicator thresholds, ensuring the comprehensiveness and relevance of the sensing dimensions and avoiding the insufficient ability of single or fixed indicator sets to capture complex fault characteristics. Finally, intelligent monitoring and early warning are achieved based on the adaptive monitoring and early warning sliding window and the adaptive sensing factor cluster. Through the dynamically adjusted window and optimized sensing factor cluster, early warning of cascading avalanche faults in the power database is realized, which effectively improves the timeliness of risk perception and the accuracy of early warning, thereby ensuring the stable and reliable operation of the power database.

[0008] Through the above technical solutions, this application effectively solves the technical problems of existing methods that rely on a single fixed indicator threshold for judgment, have a lag in risk perception, have unreasonable allocation of monitoring resources, and have low early warning accuracy and efficiency, and provides an intelligent and dynamic monitoring and early warning solution for the safe and stable operation of the power database. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the intelligent database monitoring and early warning method based on large models and multi-dimensional perception data provided in this application embodiment. Detailed Implementation

[0011] This application provides a database intelligent monitoring and early warning method based on large models and multi-dimensional perception data. This method addresses the technical problem that existing database monitoring and early warning methods rely on threshold judgments of a single fixed indicator and cannot be dynamically adjusted according to actual operational risk conditions, resulting in delayed risk perception and low accuracy and efficiency in early warning.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0014] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0015] Examples, such as Figure 1 As shown in the embodiments of this application, a database intelligent monitoring and early warning method based on large models and multi-dimensional perception data is provided, including: S10: Combining historical global perception monitoring data sequences, predict the risk of cascading avalanches in the power database within a preset time period and output the predicted fault risk index. In this embodiment of the application, the historical global perception monitoring data sequence covers multi-dimensional monitoring data generated by the power database during its past operation, including the database server's CPU utilization, memory usage, disk I / O read / write speed, network bandwidth utilization, number of database connections, query response time, transaction throughput, frequency of log error messages, data backup completion rate, and success rate of application system interface calls associated with the database.

[0016] By integrating and analyzing historical data, a time-series dataset reflecting the long-term operational status of the database is constructed. When predicting fault risks, this historical global awareness monitoring data sequence is first preprocessed, including data cleaning to remove outliers and missing values, data standardization or normalization to eliminate the impact of differences in the units of different indicators, and data smoothing to reduce the interference of short-term random fluctuations on the prediction results.

[0017] Subsequently, a predictive model, such as a deep learning-based Long Short-Term Memory (LSTM) network model, is employed. This model is capable of capturing long-term dependencies and nonlinear features in time-series data. Preprocessed historical global perception monitoring data sequences are used as input to train the LSTM model. By adjusting hyperparameters such as the number of network layers, the number of hidden units, and the learning rate, the model learns the patterns of database operational status changes over time and the potential patterns of failure risks.

[0018] After training, the model is used to predict the operating status of the power database within a preset time period and outputs a predicted fault risk index that can quantify the probability of cascading avalanche faults. The index can be set from 0 to 1, and the larger the value, the higher the risk of cascading avalanche faults occurring within the preset time period.

[0019] Specifically, step S10 in the method includes: Configure the global perception monitoring index threshold for the power database, wherein the global perception monitoring index threshold includes an internal index set, a business index set, an environmental index set, and a log index set; According to the global perception monitoring index threshold, monitor and acquire the global perception monitoring data sequence within the historical time zone; Using a pre-built fault risk prediction plugin, the fault risk of cascading avalanche in the power database within a preset time period is predicted based on the historical global perception monitoring data sequence, and the predicted fault risk index is output.

[0020] The internal category indicator set includes performance indicators, capacity indicators, health indicators, and availability indicators; the business category indicator set includes business load indicators, critical transaction indicators, data flow indicators, and business rule indicators; the environment category indicator set includes infrastructure indicators, related system indicators, natural factor indicators, and time factor indicators; and the log category indicator set includes error log indicators, security log indicators, audit log indicators, and slow query log indicators.

[0021] In this embodiment of the application, firstly, a multi-dimensional and multi-level global perception monitoring indicator system is constructed, that is, a global perception monitoring indicator threshold is configured.

[0022] This metric threshold is divided into four main categories: internal metrics focus on the core operating status of the database itself, including performance metrics such as CPU utilization, memory hit rate, and disk I / O throughput; capacity metrics such as data file size, tablespace utilization, and remaining storage space; health metrics such as index fragmentation rate, table lock wait count, and transaction rollback rate; and availability metrics such as the percentage of uptime for database services, master-slave switchover success rate, and automatic fault recovery time.

[0023] The business metrics set focuses on the operational efficiency of the database in supporting business operations, covering business load metrics, such as the total number of query requests and concurrent users per unit time; key transaction metrics, such as core business transaction response time and transaction success rate; data flow metrics, such as data write rate, data read rate, and data synchronization latency; and business rule metrics, such as data integrity constraint satisfaction rate and business logic verification pass rate.

[0024] Environmental indicators consider the external support conditions for database operation, including infrastructure indicators such as server room temperature, humidity, power supply stability, and UPS battery life; related system indicators such as network latency between application servers and database servers, and middleware operation status; natural factor indicators such as the frequency of lightning strikes and earthquake intensity warnings in the area; and time factor indicators such as holiday peak business period indicators and daily operation and maintenance period indicators.

[0025] Log metrics reflect potential problems by recording information, including error log metrics such as the frequency of different levels of error messages and the number of times specific error codes appear; security log metrics such as the number of abnormal login attempts and the number of permission change records; audit log metrics such as the number of sensitive operation execution records and data access audit coverage; and slow query log metrics such as the frequency of slow query statements and the percentage of queries whose execution time exceeds the threshold.

[0026] Secondly, after configuring the global perception monitoring indicator thresholds, the system continuously and systematically monitors and collects the operational data of the power database within the historical time zone according to the definitions and collection frequency requirements of each indicator in the indicator set. For example, it collects the operational data of the power database over the past 12 months to obtain a complete historical global perception monitoring data sequence. Each data point in this sequence corresponds to the value of each indicator at a specific timestamp, forming a multivariate time series dataset.

[0027] Secondly, a pre-built fault risk prediction plugin is used to perform the risk prediction task. The feature set obtained by preprocessing and feature engineering the historical global sensing monitoring data sequence is input into the prediction model. The model learns the mapping relationship between the change patterns of various indicators before the occurrence of faults in historical data and the fault risk, thereby predicting the fault risk of cascading avalanche in the power database within a preset period. Finally, it outputs a predicted fault risk index between 0 and 1, which intuitively reflects the probability of fault occurrence within the preset period.

[0028] Furthermore, the construction process of the fault risk prediction plugin includes: Based on the historical operation and maintenance monitoring records of the power database, and constrained by the historical time zone, several sample global perception monitoring data sequences are collected, and the proportion of cascading avalanche faults that occur in the power database within a preset historical time period is collected as the sample fault risk index, thus obtaining several sample fault risk indices. Using the global perception monitoring data sequences of the aforementioned samples as input data and the fault risk index of the aforementioned samples as supervision data, a long short-term memory network is trained until convergence to generate a fault risk prediction plugin.

[0029] In this embodiment, firstly, multiple sets of historical data corresponding to the global perception monitoring index threshold are selected from the historical operation and maintenance monitoring records of the power database. Specifically, the historical time zone is used as the time range constraint, for example, operation and maintenance data from the past 3 years are selected, and several non-overlapping or partially overlapping sample global perception monitoring data sequences are collected from them. The time length of each sample sequence can be set according to actual needs, such as 1 month or 1 quarter.

[0030] Meanwhile, for each sample sequence, the number of cascading avalanche faults that actually occurred in the power database during the historical preset period, i.e. a specific duration after the end time of the sample sequence, such as one week, is counted, and the proportion of the actual occurrence of such faults in the power database during that period is calculated. This proportion is used as the sample fault risk index corresponding to the sample sequence.

[0031] For example, if a preset time period corresponding to a sample sequence is 7 days, and the total duration of cascading avalanche failures is 12 hours, then the sample failure risk index is 12 / (7×24)=0.071. Using the above method, a training sample set is constructed, containing the sample global perception monitoring data sequence as input data and the sample failure risk index as supervision data.

[0032] Subsequently, the long short-term memory network was trained using the global perception monitoring data sequence of the samples as input data and the corresponding sample fault risk index as supervision data.

[0033] Furthermore, during training, each sample sequence is divided into a training set, a validation set, and a test set, allocated in a 7:2:1 ratio. The training set data is input into the LSTM model, which calculates the predicted fault risk index through forward propagation and compares it with the supervised data. The difference between the predicted value and the true value is calculated using a loss function. The input layer has a node count equal to the dimension of the input features; for example, if the global perception monitoring data sequence has 4 features, the input layer contains 4 nodes. One to three hidden layers are set, with the number of nodes in each layer adjusted experimentally (e.g., 64, 32, etc.). The ReLU activation function is used. The output layer generally does not use an activation function; if the output takes 2 nodes, a continuous value is directly output.

[0034] Then, the model's weights and biases are adjusted using the backpropagation algorithm and optimizers, such as the Adam optimizer, to minimize the loss function. During the training iterations, the batch size is set to 32, the total number of training epochs is 50, and an early stopping mechanism with a patience of 5 is introduced. When the validation set loss does not decrease for 5 consecutive epochs, the training process is automatically terminated, resulting in a trained fault risk prediction plugin. This effectively avoids model overfitting while ensuring that the model reaches a convergent state.

[0035] Finally, the converged model is evaluated using a test set to ensure it has good generalization ability. Following this training process, a fault risk prediction plugin is generated that can accurately predict the risk of cascading avalanche faults in power databases.

[0036] S20: Configure the adaptive monitoring and early warning sliding window within the preset time period according to the predicted fault risk index; In this embodiment of the application, the parameters of the monitoring and early warning sliding window within a preset time period are dynamically configured according to the different value ranges of the predicted fault risk index, including the window size and sliding step size.

[0037] Specifically, when the predicted fault risk index is in a low range, it indicates that the database is operating relatively stably and the risk of cascading avalanche failures is low. In this case, a larger window size and a larger sliding step can be configured to reduce monitoring frequency and lower system resource consumption. When the predicted fault risk index is in a medium range, it indicates that the database has certain potential risks and the sensitivity of monitoring needs to be increased. In this case, the window size should be adjusted to a medium level and the sliding step should be adjusted to a smaller value to achieve more intensive monitoring of the database's operating status. When the predicted fault risk index is in a high range, it indicates that the database faces a high risk of cascading avalanche failures and needs to enter a high-intensity monitoring mode. In this case, a smaller window size and a very small sliding step should be configured to quickly capture subtle changes in the database's operating status and buy time for subsequent risk assessment and early warning.

[0038] Specifically, step S20 in the method includes: The ratio of the preset standard fault risk index for cascading avalanches in the power database to the predicted fault risk index is used as the window time limit compensation coefficient. The product of the preset standard monitoring and early warning time window and the window time limit compensation coefficient is used as the adaptive monitoring and early warning sliding window within the preset time period.

[0039] In this embodiment, firstly, a preset standard fault risk index for a cascading avalanche in the power database is set. This index can be determined comprehensively based on factors such as the power system's requirements for database stability, the extent of losses caused by historical faults, and industry safety standards, for example, set to 0.6. The ratio of this preset standard fault risk index to the predicted fault risk index output in step S10 is calculated to obtain the window time limit compensation coefficient.

[0040] For example, if the predicted fault risk index is 0.3 and the preset standard fault risk index is 0.6, then the window time limit compensation coefficient is 0.6 / 0.3=2; if the predicted fault risk index is 0.7 and the preset standard fault risk index is 0.6, then the window time limit compensation coefficient is 0.6 / 0.7≈0.857. This coefficient reflects the relative relationship between the current predicted risk and the standard risk. A coefficient greater than 1 indicates that the current risk is lower than the standard, and a coefficient less than 1 indicates that the current risk is higher than the standard.

[0041] Subsequently, a preset standard monitoring and early warning time window is set. This window is the monitoring and early warning sliding window parameter used under the standard risk level, that is, when the predicted fault risk index is equal to the preset standard fault risk index. For example, it includes a standard window size of 30 minutes and a standard sliding step of 5 minutes.

[0042] Furthermore, by multiplying the window size and sliding step of the preset standard monitoring and early warning time window by the window time limit compensation coefficient calculated above, the actual window size and sliding step of the adaptive monitoring and early warning sliding window within the preset time period can be obtained.

[0043] For example, when the window time limit compensation coefficient is 2, if the standard window size is 30 minutes and the standard sliding step is 5 minutes, then the actual window size is adjusted to 30 × 2 = 60 minutes and the actual sliding step is adjusted to 5 × 2 = 10 minutes; when the window time limit compensation coefficient is 0.857, the actual window size is adjusted to 30 × 0.857 ≈ 25.7 minutes, rounded to 26 minutes, and the actual sliding step is adjusted to 5 × 0.857 ≈ 4.285 minutes, rounded to 4 minutes.

[0044] The above method enables the dynamic adaptation of the monitoring and early warning sliding window parameters based on the predicted fault risk index, so that the monitoring intensity can match the risk level, and optimize system resource allocation while ensuring the effectiveness of monitoring.

[0045] S30: Optimize the multi-dimensional perception strategy according to the adaptive monitoring and early warning sliding window, and obtain the adaptive perception factor cluster; In this embodiment, the multi-dimensional perception strategy optimization aims to dynamically adjust the monitoring priority, collection frequency, and data processing accuracy of each dimension perception indicator based on the parameters of the adaptive monitoring and early warning sliding window, so as to construct a perception factor cluster that adapts to the current risk level and achieve efficient resource allocation and accurate risk perception.

[0046] Specifically, step S30 in the method includes: Using the adaptive monitoring and early warning sliding window as a time limit constraint, the sensing indicators are randomly selected within the threshold of the global sensing monitoring indicators to generate several sensing factor clusters. Based on the historical operation and maintenance monitoring records of the power database, several sample perception factor clusters are collected, and the historical early warning accuracy of different sample perception factor clusters is statistically analyzed as the sample early warning accuracy, thus obtaining several sample early warning accuracies. Using the aforementioned clusters of sample perception factors as input data and the aforementioned sample early warning accuracy as supervision data, the deep learning model is trained under supervision until convergence, thereby generating a fault accuracy predictor. Using the fault accuracy predictor, fault warning accuracy is predicted for each of the several sensing factor clusters, and several predicted fault warning accuracies are output. Using the adaptive monitoring and early warning sliding window as a time limit constraint and the global perception monitoring index threshold as the optimization space, a multi-dimensional perception strategy optimization is performed based on the several perception factor clusters and several predicted fault early warning accuracies to obtain the adaptive perception factor clusters.

[0047] In this embodiment, the window size and sliding step size of the adaptive monitoring and early warning sliding window are first used as the core time limit constraints. For example, the window size is 26 minutes and the sliding step size is 4 minutes. This means that within a preset time period in the future, the monitoring data window with a length of 26 minutes will be continuously analyzed at 4-minute intervals.

[0048] Under this constraint, within the established global perception monitoring indicator threshold, which includes four major categories of indicator sets (internal, business, environment, and log) and their respective specific indicators, different combinations of perception indicators are randomly selected to generate a large number of initial perception factor clusters.

[0049] Specifically, each perception factor cluster consists of several specific indicators from different dimensions. For example, a perception factor cluster may include CPU utilization of internal classes, core transaction response time of business classes, server room temperature of environmental classes, and frequency of slow query statements in log classes.

[0050] Secondly, based on historical operation and maintenance monitoring records of the power database, several sample perception factor clusters with structures similar to the randomly generated perception factor clusters mentioned above are collected. For each sample perception factor cluster, its early warning status for cascading avalanche faults in the power database is reviewed within a monitoring period corresponding to similar window sizes and sliding step sizes in history. The ratio of the number of successful early warnings to the total number of early warnings is calculated as the historical early warning accuracy of that sample perception factor cluster, i.e., the sample early warning precision.

[0051] Secondly, sample perception factor clusters are used as input data, and the corresponding sample early warning accuracy is used as supervised data to supervise the training of a selected deep learning model, such as a convolutional neural network (CNN). During training, the model learns the mapping relationship between different combinations of perception factors and early warning accuracy. For example, given a sample perception factor cluster containing a specific combination of indicators, the model outputs a predicted value for its early warning accuracy, which is compared with the actual sample early warning accuracy. The model parameters are adjusted through backpropagation until the model converges, generating a fault accuracy predictor that can predict the fault early warning accuracy based on the perception factor cluster.

[0052] For example, the fault accuracy predictor is trained by the following steps: First, data preparation involves collecting and generating several sample perception factor clusters.

[0053] Secondly, in model construction, the number of nodes in the input layer is equal to the dimension of the input features. If several samples perceive a cluster of factors with 2 features, then the input layer contains 2 nodes. Set 1-3 hidden layers, and adjust the number of nodes in each layer through experiments, such as 64, 32, etc. The activation function is ReLU. The output layer generally does not use an activation function. If the output takes 2 nodes, directly output continuous values.

[0054] Next, the model is trained, and the predicted fault accuracy is used as the output. Several sample warning accuracies serve as supervision data. The Adam optimizer and mean squared error loss function are used to construct the training framework. The batch size is set to 32, the total number of training epochs is 50, and an early stopping mechanism with a patience of 5 is introduced. When the validation set loss does not decrease for 5 consecutive epochs, the training process is automatically terminated, resulting in a trained fault accuracy predictor. This effectively avoids model overfitting while ensuring that the model reaches a convergent state.

[0055] Then, using the trained fault accuracy predictor, the fault warning accuracy is predicted for each of the several sensing factor clusters initially randomly generated in step S30, and the predicted fault warning accuracy for each sensing factor cluster is obtained.

[0056] Finally, taking the time limit constraint of the adaptive monitoring and early warning sliding window as the hard boundary and the global perception monitoring index threshold as the range of all possible optimizations, we comprehensively consider the prediction and early warning accuracy of each perception factor cluster, as well as its monitoring feasibility and resource consumption under the current window time limit constraint, to optimize the multi-dimensional perception strategy.

[0057] For example, given a 26-minute window size and a 4-minute sliding step, the cluster of sensing factors with the highest predictive fault warning accuracy, while keeping data acquisition and processing costs within acceptable limits, is prioritized. Through an optimization process, one or more combinations of sensing factors that best fit the current predicted fault risk level and monitoring window parameters are ultimately determined, thus obtaining the suitable sensing factor cluster. This suitable sensing factor cluster can achieve accurate perception of the cascading avalanche fault risk in the power database within a given time limit, using the optimal combination and configuration of indicators.

[0058] Specifically, using the adaptive monitoring and early warning sliding window as a time constraint, perception indicators are randomly selected within the global perception monitoring indicator threshold to generate several perception factor clusters, including: Within the threshold of the global perception monitoring index, perception indicators are randomly selected to generate a first initial perception factor cluster. Based on the historical operation and maintenance monitoring records of the power database, the average duration of the first historical monitoring and early warning required for the first initial sensing factor cluster is calculated. If the average duration of the first historical monitoring and early warning is greater than the adaptive monitoring and early warning sliding window, then the first initial perception factor cluster is discarded. If the average duration of the first historical monitoring and early warning is less than or equal to the adaptive monitoring and early warning sliding window, the first initial perception factor cluster is added to several perception factor clusters and iteratively selected randomly until the preset selection quantity is reached.

[0059] In this embodiment, firstly, within the threshold of global perception monitoring indicators, perception indicators of different dimensions are selected and combined using a random algorithm to form a first initial perception factor cluster. For example, memory utilization and disk I / O response time are randomly selected from internal category indicators, and transaction success rate and average query latency are randomly selected from business category indicators, together forming a first initial perception factor cluster containing four specific indicators.

[0060] Subsequently, historical operation and maintenance monitoring records from the power database were retrieved, and all historical monitoring and early warning cases containing all indicators of the first initial sensing factor cluster were selected. For each case, the total time taken from initiating indicator data collection to completing the early warning judgment and outputting the result was recorded, and the average time was calculated to obtain the average first historical monitoring and early warning time required for the first initial sensing factor cluster.

[0061] Furthermore, the mean is compared with the current window size of the adaptive monitoring and early warning sliding window. If the mean duration of the first historical monitoring and early warning is greater than the window size, it indicates that the sensing factor cluster cannot complete a complete monitoring and early warning process within the current window time limit, which may lead to data lag or untimely early warning. Therefore, the first initial sensing factor cluster is discarded. If the mean duration of the first historical monitoring and early warning is less than or equal to the window size, it indicates that it can effectively complete the monitoring and early warning task within the window time limit. It is retained and added to the sensing factor cluster set.

[0062] For example, suppose the first initial sensing factor cluster has been applied to 10 similar monitoring and early warning scenarios in history, with corresponding single monitoring and early warning durations of 8 minutes, 10 minutes, 7 minutes, 9 minutes, 12 minutes, 8 minutes, 11 minutes, 9 minutes, 10 minutes, and 8 minutes, respectively. The average duration of these durations is (8+10+7+9+12+8+11+9+10+8) / 10 = 9.2 minutes. If the current adaptive monitoring and early warning sliding window size is 26 minutes and the sliding step size is 4 minutes, since 9.2 minutes is less than 26 minutes, it indicates that the first initial sensing factor cluster is feasible under the current window time constraint and can complete data collection and early warning judgment within one window period. Therefore, it is retained and added to the candidate set of several sensing factor clusters.

[0063] Repeat the above method to randomly select perception indicators, calculate and screen historical average values, until the number of generated perception factor clusters reaches the preset selection amount, so as to ensure that there are abundant candidate solutions in the subsequent optimization process.

[0064] Furthermore, using the adaptive monitoring and early warning sliding window as a time constraint and the global perception monitoring index threshold as the optimization space, a multi-dimensional perception strategy optimization is performed based on the several perception factor clusters and several predicted fault early warning accuracies to obtain the adaptive perception factor clusters, including: The clusters of sensing factors are arranged in descending order of the accuracy of the predicted fault warning, and the clusters of sensing factors are used as the initial solutions to generate an initial solution sequence. The first Q solutions of the initial solution sequence are selected as optimal solutions and the last K solutions are selected as inferior solutions. Based on the Q optimal solutions, the K inferior solutions are clustered to generate Q solution sets. The higher the prediction fault warning accuracy of the optimal solutions in the solution set, the more inferior solutions there are in the solution set. Within each solution set, with the optimal solution as the adjustment direction, the inferior solutions within the solution set are adjusted once according to the preset optimization step size to obtain Q one-time updated solution sets. If the updated inferior solution does not satisfy the adaptive monitoring and early warning sliding window and / or the global perception monitoring index threshold, then any solution that satisfies the adaptive monitoring and early warning sliding window is randomly selected from the global perception monitoring index threshold and replaced. Identify the Q sets of one-time updated solutions. If, within the same set, the prediction fault warning accuracy of a substandard solution is greater than that of a superior solution, then the substandard solution replaces the superior solution. Based on the Q one-time update solution sets, continue to optimize the multi-dimensional perception strategy to obtain the appropriate perception factor cluster.

[0065] In this embodiment, the generated clusters of sensing factors are first arranged in descending order of their corresponding fault prediction accuracy to form an ordered initial solution sequence. For example, assuming there are 100 clusters of sensing factors with fault prediction accuracies of 0.92, 0.90, 0.88, ..., 0.65, the initial solution sequence will start from the cluster of sensing factors with accuracies of 0.92 and proceed sequentially up to the cluster of sensing factors with accuracies of 0.65.

[0066] Secondly, the first Q clusters of sensing factors are selected as optimal solutions and the last K clusters as suboptimal solutions from the initial solution sequence, where Q and K are preset positive integers, and Q + K ≤ the total number of sensing factor clusters. For example, if Q = 10 and K = 30, then the top 10 solutions with the highest fault prediction accuracy in the initial solution sequence are selected as optimal solutions, and the bottom 30 solutions with the lowest accuracy are selected as suboptimal solutions.

[0067] Subsequently, a clustering operation is performed on the K inferior solutions based on the Q superior solutions to generate Q solution sets. The clustering operation assigns each inferior solution to the solution set containing the superior solution that is most similar to it in terms of the composition of perception indicators.

[0068] Specifically, by calculating the cosine similarity or Euclidean distance of the index vectors between inferior solutions and each superior solution, inferior solutions are assigned to the solution set corresponding to the superior solution with the highest similarity or the closest distance. In this process, to ensure that superior solutions can guide the optimization of more inferior solutions, the higher the prediction accuracy of the fault warning for superior solutions within a solution set, the more inferior solutions are included in that solution set. For example, a superior solution with a prediction accuracy of 0.92 might be assigned 5 inferior solutions; a superior solution with a prediction accuracy of 0.90 might be assigned 4 inferior solutions, and so on, so that the total number of inferior solutions K can be distributed across Q solution sets.

[0069] Then, within each solution set, using the optimal solutions within that set as the adjustment direction and target template, an adjustment is made at the index level for each inferior solution within the set according to a preset optimization step size. The preset optimization step size can be adjusting a certain proportion of the perception indices in the inferior solutions, such as replacing 20% ​​of the low-weight indices in the inferior solutions with high-weight indices in the optimal solutions, or fine-tuning parameters such as the collection frequency and processing accuracy of the indices in the inferior solutions to align them with the optimal solutions. After the adjustment, Q one-time updated solution sets are obtained.

[0070] Furthermore, during the adjustment process, if an updated suboptimal solution no longer meets the time constraint of the adaptive monitoring and early warning sliding window, or if the perception indicators it contains exceed the range of the global perception monitoring indicator threshold, it needs to be replaced. The replacement method is to randomly select a new perception factor cluster within the global perception monitoring indicator threshold, and this new perception factor cluster must meet the time constraint of the adaptive monitoring and early warning sliding window, that is, its first historical average monitoring and early warning duration is less than or equal to the window size. The unqualified updated suboptimal solution is replaced with this new perception factor cluster.

[0071] Next, all perceptual factor clusters in the Q updated solution sets are identified and compared. If, within a certain solution set, the accuracy of a fault prediction prediction of a suboptimal solution after adjustment or a new perceptual factor cluster after replacement exceeds the accuracy of the original optimal solution in that solution set, a replacement mechanism is triggered. The suboptimal solution or the new perceptual factor cluster with the better performance replaces the original optimal solution, becoming the new optimal solution in that solution set.

[0072] Finally, based on the Q updated solution sets obtained above, the process of "clustering and assigning inferior solutions - adjusting inferior solutions according to the direction of superior solutions - checking and replacing unqualified solutions - comparing and replacing superior solutions" is repeated for multiple rounds of iterative optimization. Each iteration aims to improve the quality of inferior solutions in the solution set and may generate new, better solutions. This process is repeated until the preset upper limit of the number of iterations is reached, or the prediction fault warning accuracy of the optimal sensing factor cluster no longer improves significantly in multiple consecutive iterations (e.g., the improvement is less than 0.001), at which point the optimization process stops. At this point, the sensing factor cluster with the highest prediction fault warning accuracy and the best overall resource consumption is selected from the final solution set and determined as the suitable sensing factor cluster.

[0073] Specifically, based on the Q one-time updated solution sets, multi-dimensional perception strategy optimization is continued to obtain an adapted perception factor cluster, including: Based on the Q one-time updated solution sets, continue to perform multi-dimensional perception strategy optimization, and determine whether the current number of optimization iterations meets the preset merging condition. The preset merging condition is K consecutive optimization iterations, the initial value of K is 100, and K decreases as the number of merging increases. If satisfied, calculate the average of the Q predicted fault warning accuracies of the Q current updated solution sets, and arrange the Q current updated solution sets in descending order of the average of the Q predicted fault warning accuracies to generate a current updated solution set sequence; Select the top 10% of the updated solution sets in the current updated solution set sequence as the elite solution set, select the bottom 10% of the updated solution sets in the current updated solution set sequence as the sub-solution set, and add the sub-solution sets into the elite solution set in order to complete the solution set merging; Continue iterative optimization and merging until the number of remaining solutions is less than or equal to the preset number of solutions. Then the optimization converges, and the optimal solution corresponding to the maximum predicted fault warning accuracy in the remaining solution set is output as the adaptive sensing factor cluster.

[0074] In this embodiment, firstly, during each round of iterative optimization, the current number of optimization iterations is continuously tracked, and it is determined whether a preset merging condition is met. This preset merging condition is triggered after K consecutive optimization iterations, where the initial value of K is set to 100. To balance the breadth and depth of optimization, the value of K is dynamically reduced as the number of merging iterations increases. For example, the value of K is halved after each merging iteration or decreased by a fixed proportion, thus maintaining a broad search in the early stages of optimization and gradually focusing on local regions with better solutions as the iterations deepen.

[0075] If the current optimization iteration count reaches a preset K times, satisfying the merging condition, then the average predicted fault warning accuracy of each of the existing Q updated solution sets is first calculated. Specifically, for each updated solution set, the predicted fault warning accuracy of all perception factor clusters, including optimal solutions and inferior solutions that have been adjusted or replaced, is arithmetically averaged to obtain the overall average accuracy of the solution set. Subsequently, the Q updated solution sets are sorted in descending order of their calculated average predicted fault warning accuracy to generate a new sequence of current updated solution sets.

[0076] Next, from the sorted sequence of currently updated solutions, the top 10% of solutions are selected as the elite solution set. The elite solution set represents the set of solutions that performed best in the current optimization process. Simultaneously, the bottom 10% of solutions in the sequence are selected as the subset solution set. Then, the subset solution sets are added to the elite solution set in their original order in the sequence, completing the solution set merging operation.

[0077] Furthermore, after completing one solution set merging, the next round of iterative optimization is carried out based on the merged new solution set. This includes steps such as clustering and assigning inferior solutions, adjusting inferior solutions according to the direction of superior solutions, checking and replacing unqualified solutions, and comparing and replacing superior solutions. At the same time, the K value is reduced accordingly after each merging, and is used to determine the merging conditions for the next time.

[0078] This process is repeated iteratively, with optimization and solution set merging performed continuously, until the number of remaining solutions is reduced to or less than a preset threshold, for example, 5 solutions. At this point, the optimization process is considered to have converged, and all iterations and merging operations are stopped.

[0079] Finally, from the remaining converged solution set, the optimal solution with the highest prediction accuracy for fault warning is selected, and this optimal solution is determined and output as the final adaptive sensing factor cluster. This adaptive sensing factor cluster is the optimal combination of sensing strategies obtained through the above complex optimization process after comprehensively considering multiple dimensions such as time constraints, global indicator range, warning accuracy, and resource consumption.

[0080] S40: Based on the adaptive monitoring and early warning sliding window and the adaptive sensing factor cluster, perform intelligent monitoring and early warning of cascading avalanche faults on the power database within the preset time period.

[0081] In this embodiment of the application, after determining the adaptive sensing factor cluster, real-time intelligent monitoring and early warning of cascading avalanche faults are carried out on the power database within a preset time period based on the dynamic time limit characteristics of the adaptive monitoring and early warning sliding window.

[0082] Specifically, firstly, multi-dimensional data collection is initiated at the start of each adaptation monitoring and early warning sliding window, based on the various perception indicators included in the adaptation perception factor cluster. Then, during the duration of the adaptation monitoring and early warning sliding window, real-time data is compared with a preset threshold range for that indicator within the global perception monitoring indicator threshold to determine if abnormal fluctuations or exceeding the threshold occur. As the adaptation monitoring and early warning sliding window is about to close, all analysis results within the window are integrated to make a final early warning judgment.

[0083] Specifically, step S40 in the method includes: The preset time period is divided according to the adaptive monitoring and early warning sliding window to generate an adaptive monitoring and early warning sliding window sequence; According to the adaptive monitoring and early warning sliding window sequence, multi-dimensional data monitoring is performed based on the adaptive sensing factor cluster to generate the adaptive monitoring dataset for the Nth sliding window, and based on the adaptive monitoring dataset, fault risk prediction and monitoring and early warning are performed for the power database cascade avalanche fault in the N+1th sliding window.

[0084] In this embodiment, the preset time period is first divided into equal intervals according to the window size and sliding step of the adaptive monitoring and early warning sliding window, thereby generating a continuous and partially overlapping adaptive monitoring and early warning sliding window sequence.

[0085] For example, if the preset time period is 24 hours, the window size is 26 minutes, and the sliding step size is 4 minutes, then the first window covers the period from 0 to 26 minutes, the second window covers 4 to 30 minutes, the third window covers 8 to 34 minutes, and so on, until the entire preset time period is covered, forming a sequence containing multiple consecutive sliding windows. Each sliding window is arranged sequentially on the time axis, with an overlap of 26 minutes - 4 minutes = 22 minutes between adjacent windows. This overlap design ensures the continuity of data monitoring and avoids the omission of key data points due to window sliding.

[0086] Secondly, according to the generated sequence of adaptive monitoring and early warning sliding windows, for each sliding window in the sequence, a multi-dimensional data acquisition mechanism is triggered at the start time of the window based on the various sensing indicators determined in the adaptive sensing factor cluster. Specifically, for the Nth sliding window, where N is a positive integer representing the sequence number in the window sequence, real-time operating data corresponding to various indicators of the adaptive sensing factor cluster are synchronously collected from the power database starting from its start time point, including but not limited to key parameters such as equipment load rate, data transmission latency, node response time, cache hit rate, and number of database connections.

[0087] During data collection, the indicator weights and collection frequencies set by the adaptive perception factor cluster are followed to ensure that the sampling density of high-priority indicators is higher than that of low-priority indicators, thereby achieving targeted and efficient data collection. The collected data are standardized and timestamped, outliers and missing values ​​are removed, and finally integrated into the adaptive monitoring dataset for the Nth sliding window.

[0088] Subsequently, based on the adaptive monitoring dataset of the Nth sliding window, fault risk prediction and monitoring early warning are performed for the cascading avalanche faults in the power database during the following N+1 sliding window. Specifically, the adaptive monitoring dataset of the Nth sliding window is first input into a pre-trained cascading avalanche fault prediction model. This model is based on a deep learning-based time-series prediction model or an ensemble learning model incorporating multi-dimensional features. Its input is the data sequence of various indicators in the adaptive monitoring dataset, and its output is the probability value or risk level of a cascading avalanche fault occurring within the N+1th sliding window.

[0089] Furthermore, if the output fault probability value exceeds a preset warning threshold, such as 0.75, or the risk level reaches the "high risk" level, the warning mechanism is immediately triggered. Simultaneously, the adapted monitoring dataset of the Nth sliding window and the corresponding warning results are stored in a historical case library for subsequent iterative optimization of the model and dynamic adjustment of the perception factor cluster. Through a sliding window-based time-series prediction and warning mechanism, monitoring of cascading avalanche faults in the power database is achieved, providing maintenance personnel with more time to respond and effectively reducing the probability and scope of fault occurrence.

[0090] In summary, compared with existing technologies, this application dynamically constructs an adaptive monitoring and early warning sliding window to match the early warning response requirements of the database under different operating states, and achieves accurate screening and optimized combination of sensing indicators based on an iterative optimization strategy of multi-dimensional sensing factor clusters, effectively improving the accuracy and timeliness of cascading avalanche fault prediction.

[0091] In summary, the embodiments of this application have at least the following technical effects: This application provides a database-based intelligent monitoring and early warning method based on large models and multi-dimensional sensing data. First, it predicts fault risks within a preset time period by combining historical global sensing monitoring data sequences, thus proactively sensing potential cascading avalanche fault trends. Second, it configures an adaptive monitoring and early warning sliding window based on the predicted fault risk index, achieving dynamic monitoring window matching with risk levels. When the predicted risk is high, the window can be narrowed to increase monitoring density and response speed; when the risk is low, the window can be appropriately expanded to optimize resource usage, avoiding the problem of delayed early warnings in high-risk situations or resource waste in low-risk situations caused by fixed windows. Third, based on the adaptive monitoring and early warning sliding window, it optimizes multi-dimensional sensing strategies and obtains adaptive sensing factor clusters. It selects the sensing indicator combination that best matches the current risk status and window characteristics from the global sensing monitoring indicator thresholds, ensuring the comprehensiveness and relevance of the sensing dimensions and avoiding the insufficient ability of single or fixed indicator sets to capture complex fault characteristics. Finally, intelligent monitoring and early warning are achieved based on the adaptive monitoring and early warning sliding window and the adaptive sensing factor cluster. Through the dynamically adjusted window and optimized sensing factor cluster, early warning of cascading avalanche faults in the power database is realized, which effectively improves the timeliness of risk perception and the accuracy of early warning, thereby ensuring the stable and reliable operation of the power database.

[0092] Through the above technical solutions, this application effectively solves the technical problems of existing methods that rely on a single fixed indicator threshold for judgment, have a lag in risk perception, have unreasonable allocation of monitoring resources, and have low early warning accuracy and efficiency, and provides an intelligent and dynamic monitoring and early warning solution for the safe and stable operation of the power database.

[0093] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0094] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0095] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A database intelligent monitoring and early warning method based on large models and multi-dimensional sensing data, characterized in that, The method includes: By combining historical global perception monitoring data sequences, the fault risk of cascading avalanche in the power database within a preset time period is predicted, and the predicted fault risk index is output. Configure the adaptive monitoring and early warning sliding window within the preset time period based on the predicted fault risk index; Based on the adaptive monitoring and early warning sliding window, multi-dimensional perception strategy optimization is performed to obtain an adaptive perception factor cluster. Based on the adaptive monitoring and early warning sliding window and the adaptive sensing factor cluster, intelligent monitoring and early warning of cascaded avalanche faults are performed on the power database within the preset time period.

2. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 1, characterized in that, By combining historical global sensing and monitoring data sequences, the risk of cascading avalanches in the power database within a preset time period is predicted, and a predicted fault risk index is output, including: Configure the global perception monitoring index threshold for the power database, wherein the global perception monitoring index threshold includes an internal index set, a business index set, an environmental index set, and a log index set; According to the global perception monitoring index threshold, monitor and acquire the global perception monitoring data sequence within the historical time zone; Using a pre-built fault risk prediction plugin, the fault risk of cascading avalanche in the power database within a preset time period is predicted based on the historical global perception monitoring data sequence, and the predicted fault risk index is output.

3. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 2, characterized in that, The internal category indicator set includes performance indicators, capacity indicators, health indicators, and availability indicators; the business category indicator set includes business load indicators, critical transaction indicators, data flow indicators, and business rule indicators; the environment category indicator set includes infrastructure indicators, related system indicators, natural factor indicators, and time factor indicators; and the log category indicator set includes error log indicators, security log indicators, audit log indicators, and slow query log indicators.

4. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 2, characterized in that, The construction process of the fault risk prediction plugin includes: Based on the historical operation and maintenance monitoring records of the power database, and constrained by the historical time zone, several sample global perception monitoring data sequences are collected, and the proportion of cascading avalanche faults that occur in the power database within a preset historical time period is collected as the sample fault risk index, thus obtaining several sample fault risk indices. Using the global perception monitoring data sequences of the aforementioned samples as input data and the fault risk index of the aforementioned samples as supervision data, a long short-term memory network is trained until convergence to generate a fault risk prediction plugin.

5. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 1, characterized in that, Configure an adaptive monitoring and early warning sliding window within the preset time period based on the predicted fault risk index, including: The ratio of the preset standard fault risk index for cascading avalanches in the power database to the predicted fault risk index is used as the window time limit compensation coefficient. The product of the preset standard monitoring and early warning time window and the window time limit compensation coefficient is used as the adaptive monitoring and early warning sliding window within the preset time period.

6. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 2, characterized in that, Based on the aforementioned adaptive monitoring and early warning sliding window, a multi-dimensional perception strategy optimization is performed to obtain an adaptive perception factor cluster, including: Using the adaptive monitoring and early warning sliding window as a time limit constraint, the sensing indicators are randomly selected within the threshold of the global sensing monitoring indicators to generate several sensing factor clusters. Based on the historical operation and maintenance monitoring records of the power database, several sample perception factor clusters are collected, and the historical early warning accuracy of different sample perception factor clusters is statistically analyzed as the sample early warning accuracy, thus obtaining several sample early warning accuracies. Using the aforementioned clusters of sample perception factors as input data and the aforementioned sample early warning accuracy as supervision data, the deep learning model is trained under supervision until convergence, thereby generating a fault accuracy predictor. Using the fault accuracy predictor, fault warning accuracy is predicted for each of the several sensing factor clusters, and several predicted fault warning accuracies are output. Using the adaptive monitoring and early warning sliding window as a time limit constraint and the global perception monitoring index threshold as the optimization space, a multi-dimensional perception strategy optimization is performed based on the several perception factor clusters and several predicted fault early warning accuracies to obtain the adaptive perception factor clusters.

7. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 6, characterized in that, Using the adaptive monitoring and early warning sliding window as a time limit constraint, random selection of sensing indicators is performed within the threshold of the global perception monitoring indicators to generate several clusters of sensing factors, including: Within the threshold of the global perception monitoring index, perception indicators are randomly selected to generate a first initial perception factor cluster. Based on the historical operation and maintenance monitoring records of the power database, the average duration of the first historical monitoring and early warning required for the first initial sensing factor cluster is calculated. If the average duration of the first historical monitoring and early warning is greater than the adaptive monitoring and early warning sliding window, then the first initial perception factor cluster is discarded. If the average duration of the first historical monitoring and early warning is less than or equal to the adaptive monitoring and early warning sliding window, the first initial perception factor cluster is added to several perception factor clusters and iteratively selected randomly until the preset selection quantity is reached.

8. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 6, characterized in that, Using the adaptive monitoring and early warning sliding window as a time constraint and the global perception monitoring index threshold as the optimization space, a multi-dimensional perception strategy optimization is performed based on the several perception factor clusters and several predicted fault early warning accuracies to obtain the adaptive perception factor clusters, including: The clusters of sensing factors are arranged in descending order of the accuracy of the predicted fault warning, and the clusters of sensing factors are used as the initial solutions to generate an initial solution sequence. The first Q solutions of the initial solution sequence are selected as optimal solutions and the last K solutions are selected as inferior solutions. Based on the Q optimal solutions, the K inferior solutions are clustered to generate Q solution sets. The higher the prediction fault warning accuracy of the optimal solutions in the solution set, the more inferior solutions there are in the solution set. Within each solution set, with the optimal solution as the adjustment direction, the inferior solutions within the solution set are adjusted once according to the preset optimization step size to obtain Q one-time updated solution sets. If the updated inferior solution does not satisfy the adaptive monitoring and early warning sliding window and / or the global perception monitoring index threshold, then any solution that satisfies the adaptive monitoring and early warning sliding window is randomly selected from the global perception monitoring index threshold and replaced. Identify the Q sets of one-time updated solutions. If, within the same set, the prediction fault warning accuracy of a substandard solution is greater than that of a superior solution, then the substandard solution replaces the superior solution. Based on the Q one-time update solution sets, continue to optimize the multi-dimensional perception strategy to obtain the appropriate perception factor cluster.

9. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 8, characterized in that, Based on the Q one-time update solution sets, continue to optimize the multi-dimensional perception strategy to obtain the adaptive perception factor cluster, including: Based on the Q one-time updated solution sets, continue to perform multi-dimensional perception strategy optimization, and determine whether the current number of optimization iterations meets the preset merging condition. The preset merging condition is K consecutive optimization iterations, the initial value of K is 100, and K decreases as the number of merging increases. If satisfied, calculate the average of the Q predicted fault warning accuracies of the Q current updated solution sets, and arrange the Q current updated solution sets in descending order of the average of the Q predicted fault warning accuracies to generate a current updated solution set sequence; Select the top 10% of the updated solution sets in the current updated solution set sequence as the elite solution set, select the bottom 10% of the updated solution sets in the current updated solution set sequence as the sub-solution set, and add the sub-solution sets into the elite solution set in order to complete the solution set merging; Continue iterative optimization and merging until the number of remaining solutions is less than or equal to the preset number of solutions. Then the optimization converges, and the optimal solution corresponding to the maximum predicted fault warning accuracy in the remaining solution set is output as the adaptive sensing factor cluster.

10. The database intelligent monitoring and early warning method based on large models and multi-dimensional perception data according to claim 1, characterized in that, Based on the adaptive monitoring and early warning sliding window and the adaptive sensing factor cluster, intelligent monitoring and early warning of cascading avalanche faults are performed on the power database within the preset time period, including: The preset time period is divided according to the adaptive monitoring and early warning sliding window to generate an adaptive monitoring and early warning sliding window sequence; According to the adaptive monitoring and early warning sliding window sequence, multi-dimensional data monitoring is performed based on the adaptive sensing factor cluster to generate the adaptive monitoring dataset for the Nth sliding window, and based on the adaptive monitoring dataset, fault risk prediction and monitoring and early warning are performed for the power database cascade avalanche fault in the N+1th sliding window.

Images