CN-122019331-A - Big data mining method and equipment applied to LPDDR performance detection

CN122019331ACN 122019331 ACN122019331 ACN 122019331ACN-122019331-A

Abstract

The invention provides a big data mining method and equipment applied to LPDDR performance detection, which are used for generating a time-aligned data reconstruction set by receiving an original performance record stream and reconstructing data. And performing association rule mining on the data reconstruction set, extracting frequent co-occurrence relations among performance states in a time dimension, and screening to generate a performance influence association mode set. Mapping the set to a state transition space, constructing a directionally weighted performance state evolution map, and identifying a key state transition path through map density calculation. Based on the path, a detection adjustment instruction is generated and transmitted back to a control unit of the detection system to reconfigure the sampling frequency and the execution sequence of the detection items. The invention realizes accurate self-adaptive adjustment of the LPDDR detection flow by mining the deep association mode of the performance data.

Inventors

HUANG HUI
Xiang Yuechao

Assignees

深圳市芯片测试技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. A big data mining method applied to LPDDR performance detection, the method comprising: Receiving original performance record streams which are continuously output by an LPDDR performance detection system in a continuous detection period, wherein the original performance record streams are composed of a plurality of original performance record units which are arranged in time sequence, each original performance record unit carries an independently generated detection moment mark and LPDDR performance state original data acquired at the corresponding detection moment, and the number of the original performance record units corresponds to the detection moment one by one; carrying out data reconstruction processing on the original performance record stream, mapping the original performance record units to a unified time coordinate system according to detection moment identifiers carried by each original performance record unit, eliminating the distribution non-uniformity of the original performance record units on a time axis, and generating a data reconstruction set with a time alignment structure, wherein the data reconstruction set comprises reconstruction performance data items standardized and expressed at each detection moment; performing association rule mining operation on the data reconstruction set, traversing the reconstruction performance data items in the data reconstruction set, identifying the value fluctuation modes of the reconstruction performance data items at different detection moments, extracting the frequent co-occurrence relation of the reconstruction performance data items in the time dimension, and screening and generating a performance influence association mode set describing the performance state association strength at different detection moments according to the occurrence frequency and association stability of the frequent co-occurrence relation; Mapping the performance influence correlation mode set to a state transition space, taking each performance state in the performance influence correlation mode set as a node, taking the correlation strength in the performance influence correlation mode set as a correlation metric value of edges between the nodes, constructing a directionally weighted performance state evolution map, identifying a region with the density of a state transition path exceeding a preset map density threshold by applying a map density calculation algorithm in the performance state evolution map, and extracting a state transition sequence with the state transition frequency higher than the average transition frequency from the region as a key state transition path; And generating a detection adjustment instruction aiming at the LPDDR performance detection flow based on a state transition sequence contained in the key state transition path, returning the detection adjustment instruction to a detection control unit of the LPDDR performance detection system through a detection control interface, analyzing a detection adjustment mapping relation contained in the detection adjustment instruction by the detection control unit, and reconfiguring the sampling frequency and the detection item execution sequence of the LPDDR performance detection system in a detection period according to the detection adjustment mapping relation.
2. The big data mining method applied to LPDDR performance detection according to claim 1, wherein the performing data reconstruction processing on the original performance record stream maps the original performance record units to a unified time coordinate system according to a detection time identifier carried by each original performance record unit, eliminates distribution non-uniformity of the original performance record units on a time axis, and generates a data reconstruction set with a time alignment structure, where the data reconstruction set includes a reconstruction performance data item normalized expressed by each detection time, and the method includes: analyzing each original performance recording unit in the original performance recording stream, extracting a detection moment identifier carried by each original performance recording unit, and generating moment identifier sets corresponding to the original performance recording units in number one by one; According to the time sequence indicated by the detection time mark in the time mark set, executing integral sorting operation on the original performance recording units to generate an original performance recording sequence which is continuously arranged according to the time ascending direction; Detecting the detection time interval length between adjacent original performance recording units in the original performance recording sequence, calculating the discrete distribution degree value of all the detection time interval lengths, and positioning the continuous time interval with the detection time interval length exceeding a preset interval threshold according to the discrete distribution degree value to be used as a reconstruction target area; Performing data interpolation processing on the original performance recording units in the reconstruction target area, taking the performance state original data of the original performance recording units at the boundary of the reconstruction target area as interpolation base points, generating a supplementary performance recording unit for filling the missing time points in the reconstruction target area, and inserting the supplementary performance recording unit into a position corresponding to the detection moment mark in the original performance recording sequence; carrying out field standardization conversion on each original performance recording unit and each supplementary performance recording unit in the original performance recording sequence inserted into the supplementary performance recording unit according to a preset unified data format template to generate a reconstruction performance data item with the same data structure field name and field sequence at each detection moment; And carrying out overall aggregation packaging on the reconstruction performance data items corresponding to all the detection moments according to the time sequence identified by the detection moments, and generating a data reconstruction set which is continuous and gapless in the time dimension and has consistent data structure.
3. The big data mining method applied to LPDDR performance detection according to claim 2, wherein said parsing each original performance recording unit in the original performance recording stream, extracting a detection time identifier carried by each original performance recording unit, and generating a time identifier set corresponding to the original performance recording units in number one-to-one, includes: traversing all original performance recording units contained in the original performance recording stream, executing head information scanning operation on each original performance recording unit, and positioning the fixed offset field position used for storing the detection moment mark in each original performance recording unit; reading binary time coding data from the fixed offset field position of each original performance recording unit, and converting the binary time coding data into a standardized time format character string capable of being ordered according to a preset time decoding rule; each standardized time format character string is allocated with a temporary index number which is bound with the sequence position of an original performance recording unit from which the character string is derived in the original performance recording stream; Storing the standardized time format character strings and the corresponding temporary index numbers in an associated manner to form a time identification set containing the corresponding relation between the standardized time format character strings and the temporary index numbers; And globally sorting all the standardized time format character strings in the time identification set according to the time sequence, and updating the temporary index number associated with each standardized time format character string according to the sorting result to ensure that the updated temporary index number is consistent with the time sequence of the standardized time format character strings.
4. The big data mining method applied to LPDDR performance detection according to claim 2, wherein the performing a data interpolation process on the original performance recording unit in the reconstruction target area to reconstruct performance state original data of the original performance recording unit at a boundary of the target area as an interpolation base point, generating a supplementary performance recording unit that fills up a missing time point in the reconstruction target area, and inserting the supplementary performance recording unit into a position corresponding to a detection time mark in an original performance recording sequence, includes: identifying an initial boundary original performance recording unit and a termination boundary original performance recording unit of the reconstruction target area, and extracting first performance state original data carried by the initial boundary original performance recording unit and second performance state original data carried by the termination boundary original performance recording unit; Calculating the quantity of missing detection moments to be filled in a reconstruction target area, uniformly segmenting a numerical variation interval between the first performance state original data and the second performance state original data according to the quantity of missing detection moments, and generating intermediate state interpolation data corresponding to each missing detection moment one by one; Creating a blank recording unit template for each intermediate state interpolation data, filling a detection moment identification field of the blank recording unit template into a corresponding missing detection moment, filling a performance state original data field of the blank recording unit template into corresponding intermediate state interpolation data, and generating a complementary performance recording unit; Each generated supplementary performance recording unit is inserted into a corresponding position between adjacent original performance recording units in the original performance recording sequence according to the detection time mark, so that the detection time interval length of the original performance recording sequence in a reconstruction target area is uniform; and checking the original performance recording sequence after the supplementary performance recording units are inserted, and outputting the original performance recording sequence after interpolation processing after confirming that the detection time interval length of all the adjacent recording units in the original performance recording sequence is smaller than a preset interval threshold value.
5. The big data mining method applied to LPDDR performance detection according to claim 1, wherein the performing an association rule mining operation on the data reconstruction set, traversing the reconstruction performance data items in the data reconstruction set, identifying a value fluctuation mode of the reconstruction performance data items at different detection moments, extracting a frequent co-occurrence relationship between the reconstruction performance data items in a time dimension, and generating a performance influence association mode set describing performance state association strength at different detection moments according to occurrence frequency and association stability screening of the frequent co-occurrence relationship, includes: Dividing all reconstruction performance data items in the data reconstruction set into a plurality of continuous time window units according to the detection moment identification sequence, wherein each time window unit comprises reconstruction performance data items corresponding to continuous detection moments of fixed capacity; performing discretization interval division on performance state original data carried by the reconstructed performance data items in each time window unit, mapping the continuously valued performance state original data into discrete state identifiers, and generating performance state labels corresponding to each reconstructed performance data item; Counting the co-occurrence times of performance state labels in different detection moments in the same time window unit in all time window units, calculating the co-occurrence frequency between any two performance state labels and the ratio of the co-occurrence frequency to the occurrence frequency of a single performance state label, and generating an initial association rule candidate set; filtering the redundancy rules of the initial association rule candidate set, removing association rules with co-occurrence frequency lower than a preset frequency lower limit, and reserving association rules with the ratio of the co-occurrence frequency to the occurrence frequency of a single performance state label higher than the preset ratio lower limit to generate a simplified association rule set; And converting each association rule in the reduced association rule set into a directed association line segment pointing from a previous performance state to a subsequent performance state, and endowing each directed association line segment with an association strength measurement value, wherein the association strength measurement value is in positive monotonic relation with the co-occurrence frequency, and aggregating all the directed association line segments to form a performance influence association mode set.
6. The big data mining method applied to LPDDR performance detection according to claim 5, wherein the dividing all the reconstruction performance data items in the data reconstruction set into a plurality of consecutive time window units according to the detection time identification order, each time window unit containing reconstruction performance data items corresponding to consecutive detection times of a fixed capacity, includes: acquiring the total number of the reconstruction performance data items in the data reconstruction set and the detection moment identification of each reconstruction performance data item, and confirming that a time axis of the data reconstruction set has no breaking point according to the continuity of the detection moment identification; presetting a fixed capacity of a reconstruction performance data item contained in a single time window unit, taking the fixed capacity as a window sliding step length, and performing non-overlapping segmentation on a reconstruction performance data item sequence from a first reconstruction performance data item of a data reconstruction set; When the method is divided to the tail of the reconstruction performance data item sequence, if the number of the remaining reconstruction performance data items is smaller than a preset fixed capacity, the remaining reconstruction performance data items are independently formed into a tail time window unit; Generating a unique window identifier for each divided time window unit, and establishing a one-to-many mapping relation between the window identifier and detection moment identifiers of all reconstruction performance data items belonging to the time window unit; And sequencing all the time window units according to the detection time mark starting sequence containing the reconstruction performance data item to generate a time window unit sequence which is sequentially arranged according to the time window starting time.
7. The big data mining method applied to LPDDR performance detection according to claim 5, wherein said performing redundancy rule filtering on the initial association rule candidate set, removing association rules with co-occurrence frequency lower than a preset frequency lower limit, and retaining association rules with co-occurrence frequency and single performance state label occurrence frequency ratio higher than a preset ratio lower limit, generating a reduced association rule set includes: Analyzing each association rule in the initial association rule candidate set into a precursor performance state label, a subsequent performance state label and co-occurrence frequency of the precursor performance state label and the subsequent performance state label in the same time window unit; Counting the independent occurrence total frequency of each precursor performance state label in all time window units in the whole data reconstruction set, and storing the independent occurrence total frequency and the common occurrence frequency in a data structure of the same association rule; Performing first-round screening on the initial association rule candidate set, removing association rules with co-occurrence frequency values smaller than a preset co-occurrence frequency lower limit, and reserving the association rules with co-occurrence frequency reaching or exceeding the preset co-occurrence frequency lower limit as a first-round reservation rule set; Performing second-round screening on the first-round reserved rule set, calculating the percentage ratio of the total frequency of the co-occurrence frequency of each association rule in the first-round reserved rule set and the independent occurrence frequency of the precursor performance state label, eliminating the association rule with the percentage ratio smaller than the lower limit of the preset percentage ratio, and taking the association rule with the reserved percentage ratio reaching or exceeding the lower limit of the preset percentage ratio as a second-round reserved rule set; And classifying and aggregating each association rule in the second round of reservation rule set according to the precursor performance state labels, and sequencing the subsequent performance state labels according to the sequence from high to low of the co-occurrence frequency under the same precursor performance state label to generate a structured and stored simplified association rule set.
8. The big data mining method applied to LPDDR performance detection according to claim 1, wherein mapping the performance impact association pattern set to a state transition space, taking each performance state in the performance impact association pattern set as a node, taking an association strength in the performance impact association pattern set as an association metric value of an inter-node edge, constructing a directionally weighted performance state evolution graph, identifying a region in which a state transition path density exceeds a preset pattern density threshold by applying a pattern density calculation algorithm, and extracting a state transition sequence with a state transition frequency higher than an average transition frequency from the region as a key state transition path, wherein the method comprises: Traversing each directed associated line segment in the performance influence associated mode set, analyzing a precursor performance state from each directed associated line segment as a map node starting point, analyzing a subsequent performance state as a map node ending point, and extracting an associated intensity metric value carried by the directed associated line segment; De-coinciding all the analyzed spectrum node starting points and spectrum node ending points to generate a node set of the performance state evolution spectrum, wherein each performance state node in the node set has a unique identifier; Establishing directed connection edges between corresponding performance state nodes in the node set according to the pointing relation between the precursor performance state and the subsequent performance state of each directed correlation line segment, and endowing the correlation strength measurement value of the directed correlation line segment with the corresponding directed connection edge to generate a directed weighted performance state evolution map comprising the nodes and the directed connection edges; In the directed weighted performance state evolution map, performing breadth-first traversal by taking each performance state node as a starting point, and recording all path sequences which can reach from the starting point and the accumulated sum of associated intensity measurement values of all path sequences passing through nodes; And identifying a path sequence with the accumulated sum of the correlation intensity measurement values exceeding a preset accumulated sum threshold value in all path sequences as a high-weight path region, and extracting continuous node subsequences with the node occurrence frequency exceeding a preset frequency threshold value from the high-weight path region as key state transition paths.
9. The big data mining method applied to LPDDR performance detection according to claim 8, wherein traversing each directed associated line segment in the performance-affecting associated pattern set, parsing a precursor performance state from each directed associated line segment as a graph node start point, parsing a subsequent performance state as a graph node end point, and extracting an associated intensity metric value carried by the directed associated line segment, comprises: Sequentially reading an unprocessed directed associated line segment record from the performance influence associated mode set, carrying out field segmentation on the directed associated line segment record, and positioning a precursor performance state identification field, a subsequent performance state identification field and an associated intensity measurement value field; Converting the content of the precursor performance state identification field into candidate names of node starting points in the performance state evolution spectrum, and converting the content of the subsequent performance state identification field into candidate names of node ending points in the performance state evolution spectrum; judging whether the candidate name of the node starting point exists in the currently constructed temporary storage set of the performance state nodes, if not, creating a new performance state node for the candidate name and storing the new performance state node into the temporary storage set, and if so, directly referencing the existing performance state node; Judging whether the candidate name of the node terminal exists in the currently constructed temporary storage set of the performance state nodes, if not, creating a new performance state node for the candidate name and storing the new performance state node into the temporary storage set, and if so, directly referencing the existing performance state node; And taking the measurement value of the association strength measurement value field as an initial association measurement value of a directional connecting edge which is to be established and points to a node end point from the node start point, and storing the three association measurement values of the node start point, the node end point and the initial association measurement value in a temporary edge set.
10. A computer device, comprising: A processor; and a memory, wherein the memory has stored therein computer readable code which, when executed by the processor, causes the processor to perform the method of any of claims 1-9.

Description

Big data mining method and equipment applied to LPDDR performance detection Technical Field The invention relates to the field of semiconductor detection and data processing, in particular to a big data mining method and equipment applied to LPDDR performance detection. Background The LPDDR performance detection is a key link for guaranteeing stable operation in a high-speed data transmission scene in dynamic random access, performance fluctuation and abnormal symptoms can be found in time by continuously monitoring and collecting the working state of a memory, and the application of the large data mining method in the field aims at extracting evolution characteristics of the performance state from continuous detection records. At present, a preset fixed sampling frequency and a detection item execution sequence are generally adopted to periodically detect a memory, and collected performance state original data is compared with a preset static threshold value to judge whether the current performance is in a normal section or not. However, in the method, when the method faces dynamic factors such as workload change, environmental temperature fluctuation or application scene switching, real-time evolution trend of performance states is difficult to respond effectively, sporadic abnormality is often missed due to insufficient sampling density caused by fixed detection configuration, detection resource redundancy is caused by continuous collection in a stable state, in addition, the prior art lacks deep mining capability of transferring rules of the performance states in a time dimension, implicit association and key transfer paths between the states are difficult to identify from continuous detection records, detection adjustment depends on manual experience to intervene afterwards, and self-adaptive cooperation cannot be formed with the evolution of the performance states. Disclosure of Invention In view of the above, the present invention provides a big data mining method and apparatus for LPDDR performance detection. According to an aspect of the embodiment of the present invention, there is provided a big data mining method applied to LPDDR performance detection, including: Receiving original performance record streams which are continuously output by an LPDDR performance detection system in a continuous detection period, wherein the original performance record streams are composed of a plurality of original performance record units which are arranged in time sequence, each original performance record unit carries an independently generated detection moment mark and LPDDR performance state original data acquired at the corresponding detection moment, and the number of the original performance record units corresponds to the detection moment one by one; Carrying out data reconstruction processing on the original performance record flow, mapping the original performance record units to a unified time coordinate system according to the detection moment mark carried by each original performance record unit, eliminating the distribution non-uniformity of the original performance record units on a time axis, and generating a data reconstruction set with a time alignment structure, wherein the data reconstruction set comprises reconstruction performance data items which are standardized and expressed at each detection moment; Performing association rule mining operation on the data reconstruction set, traversing the reconstruction performance data items in the data reconstruction set, identifying the value fluctuation modes of the reconstruction performance data items at different detection moments, extracting the frequent co-occurrence relation of the reconstruction performance data items in the time dimension, and screening and generating a performance influence association mode set describing the association strength of the performance states at different detection moments according to the occurrence frequency and the association stability of the frequent co-occurrence relation; Mapping the performance influence correlation mode set to a state transition space, taking each performance state in the performance influence correlation mode set as a node, taking the correlation strength in the performance influence correlation mode set as a correlation metric value of edges between the nodes, constructing a directionally weighted performance state evolution map, identifying a region with the density of a state transition path exceeding a preset map density threshold by applying a map density calculation algorithm in the performance state evolution map, and extracting a state transition sequence with the state transition frequency higher than the average transition frequency from the region as a key state transition path; And generating a detection adjustment instruction aiming at the LPDDR performance detection flow based on a state transition sequence contained in the key state transition path, transmitting the de