A time series database prefetch method for feature vectors

Abstract

The present application relates to the technical field of data prefetching, and particularly to a time series database prefetching method for feature vectors, comprising the following steps: extracting static and dynamic characteristics of a measuring point to construct an initial feature vector, mapping and weighting to generate a preloading priority, comparing a predicted deviation with an actual request to update a regression coefficient, adaptively adjusting to generate a dynamic preloading parameter, and performing prefetching and dynamically correcting a model according to the parameter. In the present application, a multi-dimensional feature vector is constructed by comprehensively extracting static characteristics and dynamic operating characteristics of a target measuring point, and a regression coefficient is dynamically updated by comparing a deviation between a predicted request distribution and an actual request distribution using a multiple linear regression algorithm, while parameters such as preloading duration and loading quantity are adaptively adjusted, so as to realize accurate evaluation of time series data preloading priority and dynamic correction of prefetching parameters. In the premise of not affecting normal operation of a unit, the present application ensures that unit operating data is more real-time and stable under a high load state.

Description

Technical Field

[0001] This invention relates to the field of data prefetching technology, and in particular to a prefetch method for a time-series database of feature vectors. Background Technology

[0002] Data prefetching technology primarily involves a series of technical operations that preload necessary data from conventional storage media into a faster cache before the computer system processes the data. This encompasses cache management strategies, storage hierarchy architecture optimization, and access pattern analysis mechanisms. Its main purpose is to bridge the gap between the processing unit's computing speed and memory read speed, thereby reducing system latency. It is widely used in large computing centers, cloud computing platforms, and large-scale database management. In distributed control systems, querying time-series data is an effective means of analyzing unit operating status and identifying the causes of unit malfunctions. When AI is combined with distributed control systems, large models require access to significantly more time-series data for prediction, analysis, and control. This results in a much higher load on the time-series database for units with AI-enabled models compared to ordinary units. Ensuring more real-time and stable acquisition of unit operating data without affecting operation is a key technology for time-series databases.

[0003] Traditional time-series database prefetching technology can compensate for the difference between the processing unit's computing speed and the memory read speed and reduce system latency by loading the required data into the cache in advance. However, under high load conditions where the demand for time-series data calls surges due to the deep integration of AI large models and distributed control systems, there are problems such as the inability to adapt to the characteristics of scene fluctuations due to the use of static prefetching strategies based on historical access frequencies, and the excessive memory consumption caused by blind prefetching, which seriously affects the normal operation of the system. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a prefetch method for time-series databases based on feature vectors.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a prefetch method for a time-series database of feature vectors, comprising the following steps: S1: Obtain the business type, writing cycle, access request volume, average request value, hit rate, access count, and alarm status of the target measurement point. Construct static features of the measurement point based on the business type and writing cycle. Construct dynamic features of the measurement point based on the access request volume, average request value, hit rate, access count, and alarm status. Concatenate the static features and dynamic features of the measurement point to generate an initial feature vector. S2: The initial feature vector is mapped to generate a standard feature vector set. The preceding feature vectors associated with the requested data are selected from the standard feature vector set and the initial regression coefficients are configured. The weighted evaluation is performed based on the preceding feature vectors and the initial regression coefficients to generate a preloading priority order. S3: Perform data prefetching to construct the predicted request distribution state according to the preloading priority order, synchronously collect the actual request distribution state, extract the deviation through multiple linear regression based on the predicted request distribution state and the actual request distribution state, update the initial regression coefficient, and generate the target regression coefficient; S4: Obtain the continuous request duration, default preload duration, maximum request duration, actual request points, and default load count associated with the target regression coefficient, perform adaptive parameter adjustment, and generate dynamic preload parameters; S5: Perform preloading processing according to the dynamic preloading parameters and preloading priority order, dynamically correct the parameters of the multiple linear regression model, and generate the feature vector time series database prefetch results.

[0006] As a further aspect of the present invention, during the process of correcting the parameters of the multiple linear regression model, the predicted request distribution state and the actual request distribution state are analyzed and the prefetch hit rate is statistically calculated. When the prefetch hit rate is lower than the preset hit rate threshold, correction is performed.

[0007] As a further aspect of the present invention, the preloading priority order specifically refers to the order obtained by arranging the weighted evaluation results in descending order of numerical values.

[0008] As a further aspect of the present invention, the specific steps for obtaining the initial feature vector are as follows: S111: Obtain the target measurement point business type identifier and write cycle parameter, perform encoding conversion and mapping processing operations to extract the corresponding discrete scalar data, integrate multiple discrete scalar data spatial arrangement sequences to construct attribute sequence parameters, extract the set of distribution coordinate values ​​within the attribute sequence parameters, and establish static feature quantities of the measurement point. S112: Monitor the number of access requests, average number of requests, historical hit rate, number of accesses, and alarm status indicators of the target measurement point. Perform extreme value normalization processing on the parameters of access request volume and average number of requests, extract the set of percentage mapping sequences corresponding to each scalar, and construct the dynamic feature quantity of the measurement point. S113: Based on the static and dynamic features of the measurement point, perform vector length verification and multi-dimensional spatial parameter alignment, match the corresponding position nodes of the static and dynamic features of the measurement point, and simultaneously collect timestamp sequences to extract alignment node parameters. Then, splice the feature dimension value distribution sequence of each alignment node parameter to generate an initial feature vector.

[0009] As a further aspect of the present invention, the specific steps for obtaining the preloading priority order are as follows: S211: Based on the initial feature vector, perform data hierarchical probability mapping and numerical standardization distribution processing to obtain basic vector parameters. Perform structural combination and hierarchical sequence aggregation processing on each feature component inside the basic vector parameters, extract set indexes, summarize all node component sequences to which the set indexes belong, and establish a standard feature vector set. S212: For the standard feature vector set, filter the request preceding feature parameters associated with the requested data, obtain the basic configuration weight values ​​corresponding to the benchmark features, configure the basic configuration weight values ​​for the request preceding feature parameters as prior experience coefficient configuration items, and combine the prior experience coefficient configuration items with the request preceding feature parameter set to establish initial regression coefficients. S213: Based on the standard feature vector set, extract the request preceding feature parameters and initial regression coefficients, perform a weighted summation operation, output the score quantification value, sort and reorganize all data in descending order, assign sequence level priority according to the sorting and reorganization position, and generate a preloaded priority order.

[0010] As a further aspect of the present invention, the specific steps for obtaining the target regression coefficient are as follows: S311: Execute advance data prefetching operation instructions on data nodes based on the preloading priority order, statistically analyze the frequency distribution information of each item corresponding to the advance data prefetching operation instructions, and construct the corresponding predicted request distribution state quantity by combining the numerical range distribution interval of the frequency distribution information of the request sequence. S312: Collect the actual request distribution data sequence running inside the time series database, perform multiple linear regression fitting operation on the predicted request distribution state quantity and the actual request distribution data sequence, extract the corresponding predicted residual value distribution difference during the fitting operation, and establish a deviation gradient index. S313: Based on the initial regression coefficient and the deviation gradient index, perform iterative update operations on the coefficients, combine the downward change direction of the deviation gradient index to backpropagate and adjust the value of the internal values ​​of the initial regression coefficient, extract the convergence coefficient variable data after the loop stops, and obtain the target regression coefficient.

[0011] As a further aspect of the present invention, the specific steps for obtaining the dynamic preloading parameters are as follows: S411: Obtain the continuous request duration and default preload duration for the node data associated with the target regression coefficient, compare the numerical difference between the continuous request duration and the default preload duration, extract the node duration that fully meets the numerical replacement comparison conditions, and merge them to establish a reference request duration value. S412: Collect the maximum request duration of the historical data access query, perform a corresponding replacement operation with the reference request duration value, extract the corresponding time dimension boundary data indicators after the replacement operation, and integrate to obtain the amount of adjustment preload duration. S413: Monitor the current actual request point data and the default loading number, and perform an adaptive multiple ratio adjustment operation on the default loading number based on the current actual request point data to extract the node number allocation threshold. Combine the adjusted preloading time and the node number allocation threshold to construct dynamic preloading parameters.

[0012] As a further aspect of the present invention, the specific steps for obtaining the feature vector time series database prefetch result are as follows: S511: Based on the dynamic preloading parameters and the preloading priority order, execute resource preloading processing instructions for the corresponding hierarchical nodes, extract memory access log data sequence information during the execution process, merge and integrate it, and construct a preloading processing state quantity; S512: For the predicted request distribution state quantity and the actual request distribution data, perform a node item overlap comparison operation, calculate the percentage of total access volume of the corresponding request node data that has been hit, and obtain the prefetch hit rate. S513: Obtain the preset hit rate threshold, determine whether the difference between the prefetched hit rate and the preset hit rate threshold is in the negative range, if it is in the negative range, correct the configuration weight parameter values ​​of the multiple linear regression operation, combine the time series execution parameters of the corrected configuration parameters, and generate the feature vector time series database prefetch results.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, a multi-dimensional feature vector is constructed by comprehensively extracting static features (such as business type and write cycle) and dynamic operating features (such as access request volume, hit rate, and alarm status) of the target measurement points. The regression coefficients are dynamically updated by comparing the deviation between the predicted request distribution and the actual request distribution using a multiple linear regression algorithm. At the same time, the preloading duration, loading quantity, and other parameters are adaptively adjusted to achieve accurate evaluation of the preloading priority of time-series data and dynamic correction of preloading parameters. This effectively avoids blind data preloading, significantly reduces invalid memory occupation, and adapts to fluctuations in complex scenarios. Thus, without affecting the normal operation of the unit, it ensures more real-time and stable acquisition of unit operation data under high load conditions. Attached Figure Description

[0014] Figure 1 This is a flowchart of the main steps of the present invention; Figure 2This is a flowchart of the initial feature vector acquisition process of the present invention; Figure 3 This is a flowchart illustrating the preloading priority order acquisition process of this invention. Figure 4 This is a flowchart of the process for obtaining the target regression coefficients in this invention; Figure 5 This is a flowchart of the dynamic preloading parameter acquisition process of the present invention; Figure 6 This is a flowchart illustrating the process of obtaining the prefetch results from the feature vector time series database of this invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 A prefetch method for a time-series database of feature vectors includes the following steps: S1: Obtain the target measurement point's business type, write cycle, access request volume, average request value, hit rate, access count, and alarm status. Construct static features of the measurement point based on the business type and write cycle. Construct dynamic features of the measurement point based on the access request volume, average request value, hit rate, access count, and alarm status. Concatenate the static and dynamic features of the measurement point to generate an initial feature vector. S2: Map the initial feature vector to generate a standard feature vector set, select the preceding feature vectors associated with the requested data from the standard feature vector set and configure the initial regression coefficients, perform a weighted evaluation based on the preceding feature vectors and the initial regression coefficients, sort the weighted evaluation results in descending order of values, and generate a preload priority order. S3: Execute data prefetching to construct the predicted request distribution state according to the preloading priority order, synchronously collect the actual request distribution state, and extract the deviation gradient index through multiple linear regression based on the predicted request distribution state and the actual request distribution state. Update the initial regression coefficients based on the deviation gradient index and generate the target regression coefficients. S4: Obtain the continuous request duration, default preload duration, maximum request duration, actual request points, and default load count associated with the target regression coefficient, perform adaptive parameter adjustment, and generate dynamic preload parameters; S5: Perform preloading processing according to dynamic preloading parameters and preloading priority order, analyze and predict the request distribution status and the actual request distribution status, and calculate the prefetch hit rate. When the prefetch hit rate is lower than the preset hit rate threshold, correct the parameters of the multiple linear regression model and generate the feature vector time series database prefetch results.

[0017] Please see Figure 2 The specific steps of S1 are as follows: S111: Obtain the target measurement point business type identifier and write cycle parameter, perform encoding conversion and mapping processing operations to extract the corresponding discrete scalar data, integrate multiple discrete scalar data spatial arrangement sequences to construct attribute sequence parameters, extract the set of distribution coordinate values ​​within the attribute sequence parameters, and establish static feature quantities of the measurement point. The system reads the target measurement point's service type identifier and write cycle parameter. The target measurement point's service type identifier is determined by extracting the character-type registration tag from the device management configuration list. The write cycle parameter is obtained by collecting the underlying clock pulse transmission interval. Encoding conversion and mapping operations are performed on the target measurement point's service type identifier and write cycle parameter. The character-type registration tag is converted to an integer format number according to a specific character mapping dictionary, and the corresponding discrete scalar data is extracted. The discrete scalar data is then arranged in a sequential order according to the acquisition time, combining multiple discrete scalar data spatial arrangement sequences and merging them to construct an attribute sequence parameter. The numerical values ​​within the attribute sequence parameter are extracted and arranged into index numbers within the sequence set. These index numbers are then aggregated to establish a set of corresponding distribution coordinate values. Finally, a one-to-one correspondence binding operation is performed between the actual numerical values ​​contained in the attribute sequence parameter and the index numbers contained in the distribution coordinate value set to establish the static characteristic quantity of the measurement point.

[0018] S112: Monitor the number of access requests, average number of requests, historical hit rate, number of accesses, and alarm status indicators of the target measurement point. Perform extreme value normalization processing on the parameters of access request volume and average number of requests, extract the set of percentage mapping sequences corresponding to each scalar, and construct the dynamic feature quantity of the measurement point. The system monitors the access request volume, average request value, historical hit rate, access count, and alarm status indicators of the target measurement points. Extreme value normalization is performed on the access request volume and average request value parameters. The process involves collecting the maximum peak and minimum floor values ​​of the access request volume and average request value recorded by all measurement point nodes within the past 24-hour monitoring period. The minimum floor value is subtracted from the monitored access request volume and average request value to obtain the difference, and the minimum floor value is subtracted from the maximum peak value to obtain the range. The difference is divided by the range value to perform a ratio calculation. The ratio calculation results are used to extract the percentage mapping sequence set corresponding to each scalar. This percentage mapping sequence set is then sequentially combined with the historical hit rate, access count, and alarm status indicators to construct the dynamic feature quantities of the measurement points.

[0019] S113: Based on the static and dynamic features of the measurement points, perform vector length verification and multi-dimensional spatial parameter alignment, match the corresponding position nodes of the static and dynamic features of the measurement points, collect timestamp sequences to extract alignment node parameters, splice the feature dimension value distribution sequence of each alignment node parameter, and generate the initial feature vector. Retrieve the static and dynamic features of the measurement point, and verify that the total number of parameter elements in the static and dynamic features is completely consistent. If the total number of elements is confirmed to be consistent, perform a multi-dimensional spatial parameter alignment operation to match the corresponding position node data of the static and dynamic features at the same physical acquisition time. Call the underlying clock interface to collect the number of seconds elapsed since the reference time as a timestamp sequence, and extract the synchronization alignment time represented by the timestamp sequence as the alignment node parameter. Perform a sequential concatenation operation on the static features containing actual values ​​and permutation index numbers, the dynamic features containing normalized ratios and running identifiers, and the alignment node parameters, fusing the numerical distribution sequences of the feature dimensions of each alignment node parameter to generate an initial feature vector.

[0020] Please see Figure 3 The specific steps of S2 are as follows: S211: Based on the initial feature vector, perform data hierarchical probability mapping and numerical standardization distribution processing to obtain the basic vector parameters. Then, perform structural combination and hierarchical sequence aggregation processing on each feature component within the basic vector parameters, extract the set index, summarize all node component sequences to which the set index belongs, and establish a standard feature vector set. The initial feature vector is obtained and input into the fully connected feedforward computation process, which performs hierarchical probability mapping and numerical standardization distribution processing. The fully connected feedforward computation process consists of an input layer, a hidden layer containing 64 computation nodes, and an output layer. The input layer receives the initial feature vector and passes it to the hidden layer. The hidden layer uses a modified linear activation function to truncate input feature values ​​less than 0 to 0 and retain positive feature values. The output layer summarizes the feature values ​​passed from the hidden layer and outputs the basic vector parameter. The feature components within the basic vector parameter are structurally combined and arranged, sorted in ascending order according to their numerical values. Hierarchical sequence aggregation processing is performed on feature components belonging to the same numerical range. A set index representing the density of each range is extracted. The sequence of all node components belonging to the set index is summarized, deduplicated, and fused to establish a standard feature vector set.

[0021] S212: For the standard feature vector set, filter the request-related request-preceding feature parameters associated with the requested data, obtain the basic configuration weight values ​​corresponding to the benchmark features, configure the basic configuration weight values ​​for the request-preceding feature parameters as prior experience coefficient configuration items, and combine the prior experience coefficient configuration items with the request-preceding feature parameter set to establish initial regression coefficients. For the standard feature vector set, retrieve and filter the request-preceding feature parameters associated with the requested data. Obtain the basic configuration weight values ​​corresponding to the baseline features. The basic configuration weight values ​​are set based on the frequency data of valid data access requests triggered by all feature parameters in the historical operating environment. Calculate the mathematical average of the valid data access request frequency data set, and divide the mathematical average by the highest frequency value to obtain the basic configuration weight value. Configure the obtained basic configuration weight values ​​for the request-preceding feature parameters as prior experience coefficient configuration items. Bind the prior experience coefficient configuration items to the values ​​contained in the request-preceding feature parameter set and store them in the registry record node to establish initial regression coefficients.

[0022] S213: Based on the standard feature vector set, extract the request preceding feature parameters and initial regression coefficients, perform a weighted summation operation, output the score quantification value, sort and reorganize all data in descending order, assign sequence level priority according to the position of the descending sort and reorganization, and generate a preloaded priority order. Extract the requested preceding feature parameters and initial regression coefficients from the standard feature vector set. Multiply the requested preceding feature parameters (including their numerical values) with the initial regression coefficients to obtain the product. Sum all the product results and output the resulting score quantification value. Based on the score quantification value contained in each data node, perform a descending sorting and recombination operation on all data nodes. Obtain the descending sorting and recombination position index and assign sequence priority values ​​from high to low according to the index. Read the assigned sequence priority values ​​to generate a preloaded priority order.

[0023] Please see Figure 4 The specific steps of S3 are as follows: S311: Execute advance data prefetching operation instructions on data nodes based on the preloading priority order, statistically analyze the frequency distribution information of each item corresponding to the advance data prefetching operation instructions based on the request sequence, and construct the corresponding predicted request distribution state quantity by combining the numerical range distribution interval of the frequency distribution information based on the request sequence. The system reads the preloading priority order and executes advance data prefetching operations on the data nodes at the top of the order. The advance data prefetching operation involves loading data files from the physical disk storage area into the cache processing space in advance. It then statistically analyzes the frequency distribution information based on the request sequence for each advance data prefetching operation. This frequency distribution information represents the total number of request calls within a specific monitoring period. The system divides the numerical range starting from 0 into multiple consecutive numerical ranges based on a fixed span of 20. By combining the frequency distribution information based on the request sequence, it finds the numerical range within which the total number of request calls falls, extracts the pre-defined interval codes for each numerical range, and constructs the corresponding predicted request distribution state.

[0024] S312: Collect the actual request distribution data sequence running inside the time series database, perform multiple linear regression fitting operation on the predicted request distribution state quantity and the actual request distribution data sequence, extract the corresponding predicted residual value distribution difference during the fitting operation, and establish the deviation gradient index. The system collects actual request distribution data sequences from the time-series database. These actual request distribution data sequences are extracted directly from the time-series database access logs, representing the true read frequencies. A multiple linear regression fitting operation is performed on the predicted request distribution state quantity and the actual request distribution data sequences. This operation uses the mean squared error (MSE) calculation principle. The difference between the predicted request distribution state quantity (including the predicted value) and the actual request distribution data sequence (including the true read frequencies) at the same time point is calculated. This difference is then multiplied by itself to obtain the square. The squared values ​​from all time points are summed and divided by the total number of samples collected. The final result obtained using the MSE calculation principle is extracted as the distribution difference of the corresponding predicted residual values, establishing a bias gradient index.

[0025] S313: Based on the initial regression coefficient and the deviation gradient index, perform mutual iterative update operation on the coefficient, combine the downward change direction of the deviation gradient index to backpropagate and adjust the value of the internal value of the initial regression coefficient, extract the convergence coefficient variable data after the loop stops, and obtain the target regression coefficient. The initial regression coefficients and the deviation gradient index are extracted and iteratively updated. The deviation gradient index value is determined to be either in the positive (greater than 0) or negative (less than 0) range. Based on the direction of the deviation gradient index's descent, the values ​​within the initial regression coefficients are backpropagated and adjusted. If the deviation gradient index is greater than 0, the initial regression coefficients are subtracted by a fixed step size quantization value of 0.01; if the deviation gradient index is less than 0, the initial regression coefficients are added by the fixed step size quantization value of 0.01. The magnitude of the deviation gradient index change is monitored. When the difference in the deviation gradient index change is less than the lower limit boundary value of 0.001 for five consecutive iterations, a convergence state is determined, triggering a loop stop. The target regression coefficients are obtained by extracting the coefficient variable data corresponding to the convergence state after the loop stops.

[0026] Please see Figure 5 The specific steps of S4 are as follows: S411: Obtain the continuous request duration and default preload duration for the node data associated with the target regression coefficient, compare the numerical differences between the continuous request duration and the default preload duration, extract the node duration that fully meets the numerical replacement comparison conditions, and merge them to establish a reference request duration value. Read the data of the nodes associated with the target regression coefficient to obtain the continuous request duration during which these nodes maintain their invoked state. Retrieve the default preload duration from the basic runtime configuration file. Compare the continuous request duration with the default preload duration; specifically, subtract the default preload duration from the continuous request duration to obtain the time difference. Determine if the time difference is greater than a time retention benchmark, which is set based on the latency records of all access blocking events in the historical runtime logs, calculated by determining the standard deviation of the values ​​contained in the latency record dataset. Extract the data nodes whose time difference is greater than the time retention benchmark as nodes that fully meet the numerical replacement comparison criteria. Merge and summarize the durations of nodes that fully meet the numerical replacement comparison criteria to establish a reference request duration value.

[0027] S412: Collect data access history query maximum request duration, perform corresponding replacement operation with reference request duration value, extract the corresponding time dimension boundary limit data indicators after replacement operation, and integrate to obtain the amount of adjustment preload duration. Extract the record number of the maximum request duration from the data access history query. Perform a corresponding replacement operation between this record number and the reference request duration value. Determine if the reference request duration value is greater than the product of the record number of the maximum request duration from the data access history query multiplied by a fixed coefficient of 0.8. If the reference request duration value is greater, directly replace the original value of the scheduling parameter position with the reference request duration value. Extract the time dimension boundary data indicator that occupies the highest position after the replacement operation. Integrate all extreme value records obtained after the replacement operation on multiple types of concurrent nodes, and obtain the maximum value among all extreme value records as the adjustment amount for the preloading duration.

[0028] S413: Monitor the current actual request points data and the default load count, and perform adaptive multiplier adjustment calculations on the default load count based on the current actual request points data to extract the node number allocation threshold. Combine and adjust the overall distribution of the preloading duration and the node number allocation threshold to construct dynamic preloading parameters. Monitor the current actual request count data of active connections in the underlying network transmission channel. Extract the default load count from the global configuration library. Perform an adaptive multiplier adjustment operation on the default load count based on the current actual request count data. The adaptive multiplier adjustment operation involves dividing the current actual request count data by the default load count to obtain the expansion multiplier value, and multiplying the expansion multiplier value by the basic configuration constant 10. Extract the final result of the multiplication operation to determine the node number allocation threshold. Obtain the adjustment preloading duration and the node number allocation threshold, package the adjustment preloading duration and the node number allocation threshold into a single transmission packet as the overall distribution state, and construct the dynamic preloading parameters.

[0029] Please see Figure 6 The specific steps of S5 are as follows: S511: Based on dynamic preloading parameters and preloading priority order, execute resource preloading processing instructions for corresponding hierarchical nodes, extract memory access log data sequence information during execution, merge and integrate it, and construct preloading processing state variables. Extract dynamic preloading parameters and preloading priority order. Based on the preloading priority order (arranged from front to back), execute resource preloading processing instructions for the corresponding level nodes. The resource preloading processing instruction process involves loading disk files into the random access storage area according to the read time requirements and loading limit specified by the dynamic preloading parameters. Extract the memory access log data sequence information generated during the execution of the resource preloading processing instructions. This memory access log data sequence information includes address success feature identifiers, transmission start feature identifiers, and disk write completion feature identifiers. Merge and integrate the address success feature identifiers, transmission start feature identifiers, and disk write completion feature identifiers according to their generation time to construct the preloading processing state variable.

[0030] S512: Perform a node item overlap comparison calculation on the predicted request distribution state quantity and the actual request distribution data, calculate the percentage of total access volume of the corresponding request node data that has been hit, and obtain the prefetch hit rate. Extract the predicted request distribution state and the actual request distribution data. Perform a node item overlap comparison operation, matching the set of data object identifiers pre-loaded into the high-speed cache storage area recorded in the predicted request distribution state with the set of actual access request data object identifiers contained in the actual request distribution data. Count the total number of identical identifiers in the two sets, and use this total number of identical identifiers as the total number of data accesses to the corresponding request node that have been hit. Divide the total number of data accesses to the corresponding request node that have been hit by the total number of actual access request data object identifiers to calculate the percentage assessment value, and establish this percentage assessment value as the prefetch hit rate.

[0031] S513: Obtain the preset hit rate threshold, determine whether the difference between the prefetched hit rate and the preset hit rate threshold is within the negative range, if it is within the negative range, correct the configuration weight parameter value of the multiple linear regression operation, combine the time series execution parameters of the corrected configuration parameters, and generate the feature vector time series database prefetch result. Extract the preset hit rate threshold. The preset hit rate threshold is set based on the proportion of access hits by all request nodes within the historical operating cycle. The standard deviation of the data set of access hits by all request nodes is calculated as the fault tolerance baseline, establishing a preset hit rate threshold of 0.85. Determine if the difference between the prefetched hit rate and the preset hit rate threshold is within the negative range. The execution process is to subtract the preset hit rate threshold from the prefetched hit rate. If the subtraction result is less than 0, it is determined to be within the negative range. If it is within the negative range, the configuration weight parameter values ​​of the multiple linear regression operation are corrected by subtracting a fixed gradient value of 0.05 from the configuration weight parameter values ​​of the multiple linear regression operation. The corrected configuration parameters are then synchronized with the clock to send the timing execution parameters, generating the prefetch results of the feature vector time series database.

[0032] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for prefetching a time-series database of feature vectors, characterized in that, Includes the following steps: S1: Obtain the business type, writing cycle, access request volume, average request value, hit rate, access count, and alarm status of the target measurement point. Construct static features of the measurement point based on the business type and writing cycle. Construct dynamic features of the measurement point based on the access request volume, average request value, hit rate, access count, and alarm status. Concatenate the static features and dynamic features of the measurement point to generate an initial feature vector. S2: The initial feature vector is mapped to generate a standard feature vector set. The preceding feature vectors associated with the requested data are selected from the standard feature vector set and the initial regression coefficients are configured. The weighted evaluation is performed based on the preceding feature vectors and the initial regression coefficients to generate a preloading priority order. S3: Perform data prefetching to construct the predicted request distribution state according to the preloading priority order, synchronously collect the actual request distribution state, extract the deviation through multiple linear regression based on the predicted request distribution state and the actual request distribution state, update the initial regression coefficient, and generate the target regression coefficient; S4: Obtain the continuous request duration, default preload duration, maximum request duration, actual request points, and default load count associated with the target regression coefficient, perform adaptive parameter adjustment, and generate dynamic preload parameters; S5: Perform preloading processing according to the dynamic preloading parameters and preloading priority order, dynamically correct the parameters of the multiple linear regression model, and generate the feature vector time series database prefetch results.

2. The time-series database prefetch method for feature vectors according to claim 1, characterized in that: During the process of correcting the parameters of the multiple linear regression model, the predicted request distribution and the actual request distribution are analyzed and the prefetch hit rate is calculated. Correction is performed when the prefetch hit rate is lower than the preset hit rate threshold.

3. The time-series database prefetch method for feature vectors according to claim 1, characterized in that: The preloading priority order specifically refers to the order obtained by arranging the weighted evaluation results in descending order of numerical values.

4. The time-series database prefetch method for feature vectors according to claim 1, characterized in that, The specific steps for obtaining the initial feature vector are as follows: S111: Obtain the target measurement point service type identifier and write cycle parameter, perform encoding conversion and mapping processing operations to extract the corresponding discrete scalar data, integrate multiple discrete scalar data spatial arrangement sequences to construct attribute sequence parameters, and establish static feature quantities of the measurement point; S112: Monitor the number of access requests, average number of requests, historical hit rate, number of accesses, and alarm status indicators of the target measurement point. Perform extreme value normalization processing on the parameters of access request volume and average number of requests, extract the set of percentage mapping sequences corresponding to each scalar, and construct the dynamic feature quantity of the measurement point. S113: Based on the static and dynamic features of the measurement point, perform vector length verification and multi-dimensional spatial parameter alignment, match the corresponding position nodes of the static and dynamic features of the measurement point, and simultaneously collect timestamp sequences to extract alignment node parameters. Then, splice the feature dimension value distribution sequence of each alignment node parameter to generate an initial feature vector.

5. The time-series database prefetch method for feature vectors according to claim 4, characterized in that: In the process of establishing the static feature quantity of the measuring point, the set of distribution coordinate values ​​within the attribute sequence parameter is extracted.

6. The time-series database prefetch method for feature vectors according to claim 1, characterized in that, The specific steps for obtaining the preloading priority order are as follows: S211: Based on the initial feature vector, perform data hierarchical probability mapping and numerical standardization distribution processing to obtain basic vector parameters. Perform structural combination and hierarchical sequence aggregation processing on each feature component inside the basic vector parameters, extract set indexes, summarize all node component sequences to which the set indexes belong, and establish a standard feature vector set. S212: For the standard feature vector set, filter the request preceding feature parameters associated with the requested data, obtain the basic configuration weight values ​​corresponding to the benchmark features, configure the basic configuration weight values ​​for the request preceding feature parameters as prior experience coefficient configuration items, and combine the prior experience coefficient configuration items with the request preceding feature parameter set to establish initial regression coefficients. S213: Based on the standard feature vector set, extract the request preceding feature parameters and initial regression coefficients, perform a weighted summation operation, output the score quantification value, sort and reorganize all data in descending order, assign sequence level priority according to the sorting and reorganization position, and generate a preloaded priority order.

7. The time-series database prefetch method for feature vectors according to claim 1, characterized in that, The specific steps for obtaining the target regression coefficients are as follows: S311: Execute advance data prefetching operation instructions on data nodes based on the preloading priority order, statistically analyze the frequency distribution information of each item corresponding to the advance data prefetching operation instructions, and construct the corresponding predicted request distribution state quantity by combining the numerical range distribution interval of the frequency distribution information of the request sequence. S312: Collect the actual request distribution data sequence running inside the time series database, perform multiple linear regression fitting operation on the predicted request distribution state quantity and the actual request distribution data sequence, extract the corresponding predicted residual value distribution difference during the fitting operation, and establish a deviation gradient index. S313: Based on the initial regression coefficient and the deviation gradient index, perform iterative update operations on the coefficients, combine the downward change direction of the deviation gradient index to backpropagate and adjust the value of the internal values ​​of the initial regression coefficient, extract the convergence coefficient variable data after the loop stops, and obtain the target regression coefficient.

8. The time-series database prefetch method for feature vectors according to claim 1, characterized in that, The specific steps for obtaining the dynamic preloading parameters are as follows: S411: Obtain the continuous request duration and default preload duration for the node data associated with the target regression coefficient, compare the numerical difference between the continuous request duration and the default preload duration, extract the node duration that fully meets the numerical replacement comparison conditions, and merge them to establish a reference request duration value. S412: Collect the maximum request duration of the historical data access query, perform a corresponding replacement operation with the reference request duration value, extract the corresponding time dimension boundary data indicators after the replacement operation, and integrate to obtain the amount of adjustment preload duration. S413: Monitor the current actual request point data and the default loading number, extract the node number allocation threshold, combine the adjusted preloading time and the overall distribution of the node number allocation threshold, and construct dynamic preloading parameters.

9. The time-series database prefetch method for feature vectors according to claim 8, characterized in that, During the process of allocating the threshold for the number of extracted nodes, the default loading number is adjusted by an adaptive multiple ratio based on the current actual number of requested nodes.

10. The time-series database prefetch method for feature vectors according to claim 1, characterized in that, The specific steps for obtaining the feature vector time series database prefetch results are as follows: S511: Based on the dynamic preloading parameters and the preloading priority order, execute resource preloading processing instructions for the corresponding hierarchical nodes, extract memory access log data sequence information during the execution process, merge and integrate it, and construct a preloading processing state quantity; S512: For the predicted request distribution state quantity and the actual request distribution data, perform a node item overlap comparison operation, calculate the percentage of total access volume of the corresponding request node data that has been hit, and obtain the prefetch hit rate. S513: Obtain the preset hit rate threshold, determine whether the difference between the prefetched hit rate and the preset hit rate threshold is in the negative range, if it is in the negative range, correct the configuration weight parameter values ​​of the multiple linear regression operation, combine the time series execution parameters of the corrected configuration parameters, and generate the feature vector time series database prefetch results.

Images