Method and device for analyzing lpddr grain performance based on timing characteristics

Abstract

This application discloses a method and device for performance analysis of LPDDR particles based on time-series features. The method includes: collecting performance status monitoring data of LPDDR particles throughout their entire lifecycle; extracting time-series feature sequences based on the performance status monitoring data; performing time-series feature pattern recognition on the time-series feature sequences to obtain pattern evolution paths; determining the mutation nodes and cyclical evolution trends of the LPDDR particles during the gradual change process through the pattern evolution paths, and generating a time-series feature evolution diagram by combining the mutation nodes and the cyclical evolution trends; mining the corresponding mapping relationship between the time-series feature evolution patterns of the LPDDR particles and the causes of performance degradation based on each evolution stage in the time-series feature evolution diagram; generating a dynamic prediction strategy by combining the corresponding mapping relationship and the operating scenario characteristics of the LPDDR particles; and outputting targeted optimization guidance information for the LPDDR particles based on the dynamic prediction strategy.

Description

Technical Field

[0001] This application relates to the field of data analysis technology, and in particular to a method and device for analyzing the performance of LPDDR particles based on timing characteristics. Background Technology

[0002] LPDDR (Low Power Double Data Rate) memory chips are core storage components in mobile terminals, automotive electronics, and other devices, and their performance stability and lifespan directly affect the overall performance of these devices. Current research on LPDDR chip performance analysis primarily focuses on static timing characteristic analysis. This involves collecting performance data of the chips at specific times or under specific operating scenarios, extracting static features of single or a few performance indicators (such as read / write bandwidth and latency), and then evaluating and classifying chip performance based on these static features. Simultaneously, some studies employ machine learning algorithms to detect anomalies in chip performance data to identify whether chips have malfunctioned or experienced performance degradation. Furthermore, during the chip design and manufacturing stages, accelerated aging tests are typically conducted to collect performance data at different aging levels to analyze aging patterns and assess lifespan.

[0003] However, existing technologies mainly focus on static timing characteristic analysis, failing to fully consider the dynamic evolution logic of LPDDR chips' timing characteristic patterns throughout their entire lifecycle, as factors such as usage scenario switching and aging increase. Furthermore, existing technologies do not delve deeply enough into the causes of performance degradation, failing to establish a precise mapping relationship between timing characteristic evolution patterns and performance degradation causes. In addition, the lack of dynamic prediction strategies and targeted optimization guidance information for different operating scenarios makes it difficult to achieve forward-looking control of chip performance. Summary of the Invention

[0004] This application provides a method and apparatus for performance analysis of LPDDR chips based on timing characteristics.

[0005] This application provides, in one aspect, a method for performance analysis of LPDDR chips based on timing characteristics, applied to computer equipment. The method includes: Collect performance status monitoring data of LPDDR chips throughout their entire life cycle. The performance status monitoring data covers a variety of performance indicators of the LPDDR chips under different operating scenarios. Based on the performance status monitoring data, a time-series feature sequence characterizing the performance index changes over time is extracted, and the time-series feature sequence reflects the continuous evolution process of the LPDDR particle performance status. Temporal feature pattern recognition is performed on the time-series feature sequence to analyze the regularity and anomaly in performance evolution and obtain the pattern evolution path that characterizes the performance state change. The mutation nodes and cyclical evolution trends of the LPDDR particles in the gradual process are determined by the pattern evolution path. The mutation nodes and cyclical evolution trends are combined to generate a time-series feature evolution diagram that reflects the performance state stage changes and time-series correlation. Based on each evolution stage in the time-series feature evolution diagram, the corresponding mapping relationship between the time-series feature evolution pattern of the LPDDR particle and the performance degradation cause is explored. The corresponding mapping relationship reveals the association between different evolution stages and potential performance degradation root causes. A dynamic prediction strategy is generated by combining the corresponding mapping relationship and the operating scenario characteristics of the LPDDR particles. Based on the dynamic prediction strategy, targeted optimization guidance information for the LPDDR particles is output. The targeted optimization guidance information is used to adapt to targeted performance maintenance and adjustment for different operating scenarios.

[0006] One embodiment of this application provides a computer device, including: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement any of the aforementioned LPDDR particle performance analysis methods based on timing characteristics.

[0007] One embodiment of this application provides a readable storage medium on which a program or instruction is stored. When the program or instruction is executed by a processor, it implements the steps of the LPDDR particle performance analysis method based on timing characteristics.

[0008] This application, through a comprehensive technical solution including full lifecycle performance status monitoring data acquisition, timing feature sequence extraction, pattern evolution path analysis, timing feature evolution graph construction, evolution pattern and decay cause mapping relationship mining, and dynamic prediction strategy generation, breaks through the limitations of traditional static timing feature analysis and realizes dynamic, full-cycle, multi-dimensional cognition and forward-looking control of LPDDR particle performance status. Specifically, by collecting monitoring data of multiple performance indicators under different operating scenarios, the system comprehensively covers the performance of particles throughout their entire lifecycle, from initial use to aging and degradation. The time-series feature sequences extracted from the monitoring data accurately reflect the continuous evolution of particle performance status. The pattern evolution paths obtained through time-series feature pattern recognition accurately characterize the regularity and anomalies of performance status changes, achieving qualitative and quantitative analysis of performance evolution trends. The time-series feature evolution diagram generated by combining abrupt change nodes and cyclical evolution trends intuitively presents the stage changes and time-series correlations of performance status. The mapping relationship between the mined time-series feature evolution patterns and performance degradation triggers reveals the intrinsic connection between different evolution stages and potential performance degradation roots, providing a scientific basis for performance risk prediction. The dynamic prediction strategies and targeted optimization guidance information generated based on the mapping relationship and operating scenario characteristics can be adapted to different operating scenarios for targeted performance maintenance and adjustment, realizing a shift from passive fault handling to proactive and forward-looking regulation.

[0009] This application realizes the dynamic evolution of LPDDR chip performance status, the precise correlation of performance degradation causes, and the forward-looking control of performance risks from an overall global perspective, providing more forward-looking guidance for LPDDR chip design optimization, lifespan extension, and dynamic performance scheduling of terminal devices. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies 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.

[0011] Figure 1 This is a flowchart illustrating a time-series characteristic-based LPDDR particle performance analysis method provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of the basic structure of a computer device provided in an embodiment of this application.

[0013] Figure 3 This is a functional block diagram of an LPDDR particle performance analysis device provided in an embodiment of this application. Detailed Implementation

[0014] 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 of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0015] Please see Figure 1 , Figure 1 This is a flowchart of a method for analyzing the performance of LPDDR particles based on timing characteristics, provided in an embodiment of this application. The method can be executed by a computer device or by a computer device and a server. The method may include steps 110-160.

[0016] Step 110: Collect performance status monitoring data of LPDDR chips throughout their entire life cycle. The performance status monitoring data covers various performance indicators of the LPDDR chips under different operating scenarios.

[0017] In this embodiment, the computer device initiates a full lifecycle data acquisition process by connecting to the sensor system of the test platform or actual operating device where the LPDDR chip is located. The acquired performance status monitoring data needs to cover the entire process of the LPDDR chip from factory testing and aging tests to actual application, specifically including various performance indicators under different operating scenarios. For example, under high-load scenarios, indicators such as read / write bandwidth, latency, and error correction code count of the LPDDR chip are collected; under low-power scenarios, indicators such as standby current and sleep / wake-up time are collected. The acquisition frequency is dynamically adjusted according to the scenario, set to once per second under high-load scenarios and once every 10 minutes under low-power scenarios. All acquired data is stored according to timestamps, forming a multidimensional dataset containing time, scenario, and performance indicator dimensions.

[0018] To ensure data accuracy and integrity, the computer equipment performs real-time data verification during the acquisition process, removing outliers and missing values. For example, if the read / write bandwidth data at a certain moment deviates significantly from the normal range, it is marked as an outlier and a supplementary acquisition mechanism is initiated to interpolate and supplement the data from adjacent time points. Simultaneously, the computer equipment establishes a data version management mechanism, recording the parameter configurations and environmental conditions for each data acquisition.

[0019] Step 120: Extract time-series feature sequences representing the changes in performance indicators over time based on the performance status monitoring data. The time-series feature sequences reflect the continuous evolution of the performance status of the LPDDR particles.

[0020] In this embodiment, the computer device preprocesses the collected performance status monitoring data and then extracts time-series feature sequences using a sliding window method. First, the original data is normalized to eliminate dimensional differences between different performance indicators. For example, read / write bandwidth data is normalized to between 0 and 1, and latency data is also normalized to between 0 and 1. Then, the sliding window size is set to 100 data points, with a step size of 10 data points, and the normalized performance status monitoring data is segmented by sliding window to obtain several continuous window data segments. For each window data segment, the computer device extracts features in the time domain and frequency domain. Time domain features include mean, variance, maximum, minimum, and trend slope; frequency domain features are obtained through Fast Fourier Transform and include main frequency components and energy distribution. The time domain and frequency domain features of each window data segment are concatenated to form a multi-dimensional feature vector. Arrange the feature vectors corresponding to all window data segments in chronological order to obtain a time-series feature sequence. This time-series feature sequence reflects the continuous evolution of the performance status of LPDDR particles over time, with each feature vector corresponding to the performance status feature within a time window.

[0021] Step 130: Perform time-series feature pattern recognition on the time-series feature sequence to analyze the regularity and anomaly in the performance evolution and obtain the pattern evolution path that characterizes the performance state change.

[0022] In this embodiment of the application, the computer device performs time-series feature pattern recognition on the time-series feature sequence obtained in step 120, which specifically includes the following sub-steps: Step 131: Based on the time-series feature sequence, a sliding window of preset length is used to perform sliding segmentation on the time-series feature sequence to generate several continuous sliding window sub-sequences that have partial overlap.

[0023] In this embodiment, after obtaining the timing feature sequence, the computer device further segments it using a sliding window method. The preset length of the sliding window is 50 feature vectors, and the step size is 5 feature vectors, ensuring partial overlap between the generated sliding window sub-sequences to preserve the continuity of timing features. For example, the first sliding window sub-sequence contains feature vectors 1 to 50, the second sliding window sub-sequence contains feature vectors 6 to 55, and so on. Each sliding window sub-sequence corresponds to a set of performance state features of the LPDDR chip over a period of time, and the partial overlap design helps to capture the gradual change process of performance state. During the segmentation process, the computer device records the start and end times of each sliding window sub-sequence.

[0024] Step 132: Perform multi-dimensional feature extraction processing on each sliding window subsequence to generate a corresponding multi-resolution feature vector, which covers feature parameters in the time domain and frequency domain.

[0025] In this embodiment, the computer device performs multi-dimensional feature extraction on each sliding window subsequence. For time-domain features, statistical methods are used to calculate parameters such as the mean, variance, median, and percentiles of each sliding window subsequence. For frequency-domain features, a discrete Fourier transform is used to convert the sliding window subsequence from the time domain to the frequency domain, extracting parameters such as the amplitude, phase, and energy distribution of the main frequency components. The time-domain and frequency-domain features are then fused to generate a multi-resolution feature vector. For example, the time-domain features of a sliding window subsequence include the mean, variance, maximum, and minimum values, while the frequency-domain features include the amplitude and phase of the main frequency components. These features are concatenated to form a multi-dimensional feature vector, which comprehensively reflects the time and frequency characteristics of the LPDDR chip performance state within the sliding window subsequence.

[0026] Step 133: Perform clustering pattern classification on the multi-resolution feature vectors corresponding to all sliding window sub-sequences to obtain several different pattern categories. Each pattern category corresponds to a performance change state of the LPDDR particle.

[0027] In this embodiment, the computer device uses the K-means clustering algorithm to classify the clustering patterns of the multi-resolution feature vectors corresponding to all sliding window sub-sequences. First, the number of clusters K is determined; based on prior knowledge of LPDDR particle performance changes and data distribution characteristics, K is set to 5. Then, K multi-resolution feature vectors are randomly selected as initial cluster centers. Next, the distance between each multi-resolution feature vector and the initial cluster center is calculated, and it is assigned to the category of the nearest cluster center. Afterward, the mean of each cluster center is recalculated, and the cluster centers are updated. This process is repeated until the cluster centers no longer change or a preset number of iterations is reached. Finally, five different pattern categories are obtained, each corresponding to a performance change state of the LPDDR particle. For example, pattern category 1 corresponds to a stable performance state, pattern category 2 to a slowly declining performance state, pattern category 3 to a rapidly declining performance state, pattern category 4 to a performance recovery state, and pattern category 5 to an abnormal performance state.

[0028] Step 134: Perform anomaly detection processing on the clustered pattern categories, identify the first abnormal subsequence corresponding to the multi-resolution feature vector that deviates from the normal pattern category, and determine the second abnormal subsequence corresponding to the sudden deviation and persistent anomaly in the time-series feature sequence.

[0029] In this embodiment, the computer device performs anomaly detection processing on the clustered pattern categories. First, normal pattern categories are defined as pattern category 1 and pattern category 2, and abnormal pattern categories are defined as pattern category 3, pattern category 4, and pattern category 5. Then, the distance between each multi-resolution feature vector and the center of the normal pattern category is calculated, and multi-resolution feature vectors with a distance exceeding a preset threshold are marked as anomalies. The sliding window subsequence corresponding to these anomalies is the first anomaly subsequence. Next, the distribution of anomalies in the time-series feature sequence is analyzed to identify sudden deviations and persistent anomalies. Sudden deviations are manifested as the continuous appearance of one or a few anomalies, while persistent anomalies are manifested as the continuous appearance of multiple anomalies. The sliding window subsequences corresponding to sudden deviations and persistent anomalies are marked as the second anomaly subsequences. For example, when three consecutive sliding window subsequences belonging to pattern category 3 are detected, they are marked as persistent anomalies, and the corresponding sliding window subsequences are the second anomaly subsequences.

[0030] Step 135: Sort all pattern categories in chronological order and combine the first and second abnormal subsequences to analyze the transition relationships and frequencies between different pattern categories. Based on the transition relationships and frequencies, extract the time node information corresponding to each pattern category to generate a state transition sequence.

[0031] In this embodiment, the computer device concatenates and sorts all pattern categories in chronological order to obtain a pattern category sequence. Then, combining the first and second anomalous subsequences, the transformation relationships and frequencies between different pattern categories are analyzed. Transformation relationships include direct and indirect transformations between pattern categories, and transformation frequencies include the number of transformations and probabilities between pattern categories. For example, the transformation from pattern category 1 to pattern category 2 occurs 10 times with a probability of 0.2; the transformation from pattern category 2 to pattern category 3 occurs 5 times with a probability of 0.1. Based on the transformation relationships and frequencies, the time node information corresponding to each pattern category is extracted, including the start and end times of the pattern category's occurrence. The pattern category sequence and the corresponding time node information are combined to generate a state transition sequence.

[0032] Step 136: Generate transition trajectories between different mode categories based on the state transition sequence, and construct a mode evolution path that characterizes the performance state transition process of the LPDDR particle by combining the transition trajectories and the performance change characteristics corresponding to each sliding window subsequence.

[0033] In this embodiment, the computer device generates transition trajectories between different mode categories based on state transition sequences. The transition trajectories are represented by directed line segments, pointing from one mode category to another. The length of the line segment represents the time interval between transitions, and the width of the line segment represents the frequency of transitions. Then, combining the transition trajectories with the performance change characteristics corresponding to each sliding window subsequence, a mode evolution path is constructed. Performance change characteristics include the trend and magnitude of performance index changes within the sliding window subsequence. For example, the performance change characteristics corresponding to the transition trajectory from mode category 1 to mode category 2 are a slow decrease in read / write bandwidth and a slow increase in latency. Combining the transition trajectories and performance change characteristics forms a mode evolution path, which characterizes the transition process of the LPDDR particle performance state, including the transition relationships between mode categories, the transition frequency, and the performance change characteristics.

[0034] Step 140: Determine the mutation nodes and cyclical evolution trends of the LPDDR particles in the gradual process through the pattern evolution path, and generate a timing feature evolution diagram that reflects the performance state stage changes and timing correlation by combining the mutation nodes and the cyclical evolution trends.

[0035] In this embodiment of the application, the computer device determines the mutation nodes and cyclical evolution trends of LPDDR particles during the gradual evolution process through the pattern evolution path obtained in step 130, specifically including the following sub-steps: Step 141: Based on the evolution trajectory diagram corresponding to the mode evolution path, extract the state information and time node information corresponding to all mode categories, and construct the state transition matrix of the LPDDR particle performance state change. The state transition matrix represents the transition probability and number of transitions between different mode categories.

[0036] In this embodiment, the computer device first generates an evolution trajectory diagram based on the pattern evolution path. This diagram uses time as the horizontal axis and pattern category as the vertical axis, with curves representing the changes in pattern categories over time. Then, it extracts state information and time node information corresponding to all pattern categories from the evolution trajectory diagram. The state information includes the pattern category identifier and performance characteristics, and the time node information includes the start and end times of the pattern category's appearance. Next, a state transition matrix is ​​constructed. The rows and columns of the matrix represent different pattern categories, and the elements in the matrix represent the transition probability and the number of transitions from the row pattern category to the column pattern category. For example, the element in the first row and second column of the state transition matrix indicates that the transition probability from pattern category 1 to pattern category 2 is 0.2, and the number of transitions is 10.

[0037] Step 142: Extract the path node sequence corresponding to each mode category transformation from the state transition matrix, perform trend analysis on the path node sequence, and identify the nodes whose performance characteristic parameter change rate exceeds the preset threshold by calculating the change rate of each node in the path node sequence and determine them as inflection points in the gradual change process of the LPDDR particle performance. Each inflection point corresponds to a jump moment, and the nodes corresponding to all jump moments together constitute the mutation node.

[0038] In this embodiment, the computer device extracts the path node sequence corresponding to each mode category transition from the state transition matrix. The path node sequence consists of the start node and the end node of the mode category transition. Then, trend analysis processing is performed on the path node sequence to calculate the rate of change of the performance characteristic parameters corresponding to each node. Performance characteristic parameters include read / write bandwidth, latency, error correction code count, etc. The rate of change is calculated by dividing the difference between the performance characteristic parameter of the current node and the performance characteristic parameter of the previous node by the time interval. For example, in the transition from mode category 1 to mode category 2, the rate of change of read / write bandwidth is (read / write bandwidth of mode category 2 - read / write bandwidth of mode category 1) / time interval. Nodes with a rate of change exceeding a preset threshold are identified as inflection points, and each inflection point corresponds to a transition moment. All nodes corresponding to transition moments together constitute a mutation node. For example, when the rate of change of read / write bandwidth exceeds 0.5, the corresponding node is a mutation node.

[0039] Step 143: Perform periodic analysis on the path node sequence, mine the repeated pattern category combinations in the path node sequence, determine the periodic cyclic characteristics of the performance state change of the LPDDR particle, combine the performance change trend corresponding to each pattern category, analyze the performance change law in the periodic cyclic process, and obtain the cyclical evolution trend in the gradual performance change process of the LPDDR particle.

[0040] In this embodiment, the computer device performs periodic analysis on the path node sequence, employing the autocorrelation function method to mine recurring pattern category combinations within the sequence. The autocorrelation function method determines the period length of the periodic cycle by calculating the correlation coefficient between the path node sequence and itself. For example, when the combination of pattern category 1, pattern category 2, and pattern category 3 recurs in the path node sequence, the period length is 3 time units. Then, combining the performance change trends corresponding to each pattern category, the performance change patterns during the periodic cycle are analyzed. Performance change trends include increases, decreases, or stabilization of performance indicators. For example, during the periodic cycle, read / write bandwidth first increases and then decreases, while latency first decreases and then increases. Based on these patterns, the cyclical evolution trend of LPDDR chip performance during the gradual change process is obtained.

[0041] Step 144: Based on the jump time corresponding to the mutation node, perform gradual interval segmentation on the performance state change process of the LPDDR particle throughout its entire life cycle to obtain several continuous gradual intervals. Each gradual interval corresponds to a stable change stage of the performance state of the LPDDR particle.

[0042] In this embodiment, the computer device performs gradual interval segmentation processing on the performance state change process of the LPDDR particle throughout its entire life cycle based on the jump time corresponding to the mutation node. The time period between two adjacent jump times is divided into a gradual interval, and each gradual interval corresponds to a stable change stage of the LPDDR particle's performance state. For example, if the jump times are t1, t2, and t3, then the gradual intervals are [t0, t1), [t1, t2), [t2, t3), and [t3, t4], where t0 is the start time and t4 is the end time. The performance state change within each gradual interval is relatively stable, with fewer mode category transitions.

[0043] Step 145: Perform trend fitting processing on the path node sequence in each gradient interval to generate a corresponding trend fitting curve. The trend fitting curve reflects the changing trend of the LPDDR particle performance status in the corresponding gradient interval.

[0044] In this embodiment, the computer device performs trend fitting processing on the path node sequence within each gradient interval, and generates a trend fitting curve using a linear regression algorithm. The linear regression algorithm fits the relationship between performance characteristic parameters and time in the path node sequence using the least squares method, obtaining the slope and intercept of the trend fitting curve. The slope represents the rate of change of the performance characteristic parameter, and the intercept represents the initial value of the performance characteristic parameter. For example, for a path node sequence with read / write bandwidth, a positive slope of the trend fitting curve indicates that the read / write bandwidth increases with time, while a negative slope indicates that the read / write bandwidth decreases with time. The trend fitting curve reflects the changing trend of the LPDDR chip performance state within the corresponding gradient interval.

[0045] Step 146: Construct a directed time series graph based on the segmented gradual transition intervals, mutation nodes, cyclical evolution trends, and trend fitting curves. Mark the pattern category corresponding to each gradual transition interval, the jump time corresponding to each mutation node, and the periodic information of the cyclical evolution trend in the directed time series graph. Simultaneously construct a stage transition network between each gradual transition interval. Combine the directed time series graph and the stage transition network to generate the time series feature evolution graph.

[0046] In this embodiment, the computer device constructs a directed time series graph based on segmented gradual transition intervals, mutation nodes, cyclical evolution trends, and trend fitting curves. The directed time series graph uses time as the horizontal axis and gradual transition intervals as the vertical axis, with nodes representing gradual transition intervals and directed edges representing the transition relationships between them. The directed time series graph marks the mode category corresponding to each gradual transition interval, the jump time corresponding to the mutation node, and the periodic information of the cyclical evolution trend. For example, the mode category corresponding to gradual transition interval 1 is mode category 1, the jump time is t1, and the periodic information is 3 time units. Simultaneously, a stage transition network is constructed between each gradual transition interval. The stage transition network uses nodes to represent gradual transition intervals and edges to represent the transition probability and number of transitions between gradual transition intervals. Combining the directed time series graph and the stage transition network, a time series feature evolution graph is generated. This time series feature evolution graph reflects the stage transitions and time series correlations of the LPDDR particle performance state, including the division of gradual transition intervals, changes in mode categories, the position of mutation nodes, the period of the cyclical evolution trend, and the probability and number of stage transitions.

[0047] Step 150: Based on each evolution stage in the time-series feature evolution diagram, mine the corresponding mapping relationship between the time-series feature evolution pattern of the LPDDR particle and the performance degradation cause. The corresponding mapping relationship reveals the association between different evolution stages and potential performance degradation root causes.

[0048] In this embodiment of the application, the computer device uses the evolution stages in the timing feature evolution diagram obtained in step 140 as a basis to mine the corresponding mapping relationship between the timing feature evolution pattern of LPDDR particles and the performance degradation causes, specifically including the following sub-steps: Step 151: Based on the time-series feature evolution map, extract the stage transition network and the information corresponding to each transition interval, generate the performance state change stage division map of the LPDDR particle, and combine the mode category information corresponding to the mode evolution path to construct the time-series feature evolution topology map. The performance state change stage division map is used to mark the time range and core performance characteristics of each evolution stage, and the evolution topology map is used to present the mode category and mode transformation relationship corresponding to each evolution stage.

[0049] In this embodiment, the computer device first extracts the stage transition network and information corresponding to each transition interval from the temporal feature evolution graph, including the time range, mode category, and performance characteristic parameters of the transition interval. Then, it generates a performance state change stage division graph, with time as the horizontal axis and evolution stage as the vertical axis, using rectangles to represent each evolution stage. The time range and core performance characteristics of each evolution stage are marked within the rectangles. Core performance characteristics include the changing trends and magnitudes of read / write bandwidth, latency, and error correction code count. Next, combining the mode category information corresponding to the mode evolution path, a temporal feature evolution topology graph is constructed. The temporal feature evolution topology graph uses nodes to represent mode categories and edges to represent the transformation relationships between mode categories. The edge weights represent the transformation probability and the number of transformations. For example, the edge weight from mode category 1 to mode category 2 has a transformation probability of 0.2 and a transformation count of 10.

[0050] Step 152: Classify each evolution stage according to the performance state change stage division diagram and the evolution topology diagram. Combine the performance change trend and mode category corresponding to each evolution stage to divide the evolution stage into a stable stage and a decline stage.

[0051] In this embodiment of the application, the computer device classifies each evolution stage according to the performance state change stage division diagram and the evolution topology diagram, specifically including the following sub-steps: Step 1521: Extract the time span descriptor and key performance feature set for each evolution stage based on the performance state change stage division diagram. The time span descriptor includes the start time and end time of each evolution stage, and the key performance feature set includes the values ​​and variation range of key performance indicators marked in the time-series feature evolution diagram for each evolution stage.

[0052] In this embodiment, the computer device extracts the time span descriptor and key performance feature set for each evolution stage from the performance state change stage segmentation diagram. The time span descriptor includes the start and end time points of each evolution stage; for example, the start time point of evolution stage 1 is t0, and the end time point is t1. The key performance feature set includes the values ​​of key performance indicators marked in the time-series feature evolution diagram for each evolution stage and their range of variation; for example, the key performance indicator values ​​for evolution stage 1 are read / write bandwidth of 800MB / s to 1000MB / s and latency of 10ns to 15ns.

[0053] Step 1522: Based on the evolutionary topology graph, retrieve the pattern category identifiers associated with each evolutionary stage and the connection topology of the pattern category identifiers in the stage transition network, wherein the connection topology includes the transition probability and transition direction between pattern categories.

[0054] In this embodiment, the computer device retrieves the pattern category identifiers associated with each evolutionary stage and the connection topology of these pattern category identifiers in the stage transition network based on an evolutionary topology graph. The pattern category identifiers include pattern category 1, pattern category 2, etc., and the connection topology includes the transition probability and direction between pattern categories. For example, the pattern category identifier associated with evolutionary stage 1 is pattern category 1, and the connection topology has a transition probability of 0.2 from pattern category 1 to pattern category 2, with a positive direction.

[0055] Step 1523: Based on the start time point and end time point in the time span descriptor, segment the performance data subsequences corresponding to the time intervals from the time series feature sequence. Each performance data subsequence constitutes a representative segment of the performance state of the corresponding evolution stage.

[0056] In this embodiment, the computer device segments a performance data subsequence corresponding to a time interval from the time-series feature sequence based on the start and end time points in the time span descriptor. For example, if the start time point of evolution stage 1 is t0 and the end time point is t1, a performance data subsequence between t0 and t1 is segmented from the time-series feature sequence, and this performance data subsequence constitutes a representative segment of the performance state of evolution stage 1.

[0057] Step 1524: Perform trend modeling on representative performance state segments for each evolutionary stage, fit a trend line using a linear regression algorithm, calculate the slope of the trend line as a trend slope metric, and calculate the average deviation between the data points of the representative performance state segments and the trend line as a volatility metric. Combine the variation range of the key performance indicators in the key performance feature set within the evolutionary stage to calculate a comprehensive stability evaluation index for each evolutionary stage. The comprehensive stability evaluation index is obtained by multiplying the weighted product of the absolute value of the trend slope metric and the reciprocal of the volatility metric, and then multiplying it by the normalization coefficient of the variation range.

[0058] In this embodiment, the computer device performs trend modeling on representative performance state segments of each evolutionary stage and fits a trend line using a linear regression algorithm. The slope of the trend line is calculated as a trend slope metric; the smaller the absolute value of the slope, the more stable the performance change. The average deviation between the data points of the representative performance state segments and the trend line is calculated as a volatility metric; the smaller the average deviation, the smaller the performance fluctuation. Combining the variation amplitude of key performance indicators in the key performance feature set within the evolutionary stage, a comprehensive stability evaluation index for each evolutionary stage is calculated. The comprehensive stability evaluation index is obtained by weighting the absolute value of the trend slope metric and the reciprocal of the volatility metric, and then multiplying it by a normalization coefficient of the variation amplitude. For example, if the absolute value of the trend slope metric is 0.1, the volatility metric is 0.2, and the normalization coefficient of the variation amplitude is 0.5, then the comprehensive stability evaluation index is 0.1 × (1 / 0.2) × 0.5.

[0059] Step 1525: Set stability threshold intervals, mark the evolution stage where the comprehensive stability evaluation index falls into the first stability threshold sub-interval as a stable stage candidate, and mark the evolution stage where the comprehensive stability evaluation index falls into the second stability threshold sub-interval as a decay stage candidate.

[0060] In this embodiment, the computer device is configured with stability threshold ranges. A first stability threshold sub-range is from 0 to 0.3, and a second stability threshold sub-range is from 0.3 to 1. Evolutionary stages where the comprehensive stability evaluation index falls within the first stability threshold sub-range are marked as candidates for stable stages. For example, the comprehensive stability evaluation index for evolutionary stage 1 is 0.25, falling within the first stability threshold sub-range, and is thus marked as a candidate for stable stages. Evolutionary stages where the comprehensive stability evaluation index falls within the second stability threshold sub-range are marked as candidates for decay stages. For example, the comprehensive stability evaluation index for evolutionary stage 2 is 0.4, falling within the second stability threshold sub-range, and is thus marked as a candidate for decay stages.

[0061] Step 1526: Refer to the pattern category identifier to query the preset pattern category classification mapping table. If the pattern category identifier corresponds to the set of stable pattern categories in the preset pattern category classification mapping table, then add a stability enhancement factor to the preliminary identification result of the corresponding evolution stage. If the pattern category identifier corresponds to the set of declining pattern categories in the preset pattern category classification mapping table, then add a decline enhancement factor to the preliminary identification result of the corresponding evolution stage.

[0062] In this embodiment, the computer device queries a preset mode category classification mapping table using a reference mode category identifier. The preset mode category classification mapping table includes mode category 1 and mode category 2 as stable mode categories, and mode category 3, mode category 4, and mode category 5 as decaying mode categories. If the mode category identifier corresponds to the stable mode category set, a stability enhancement factor is added to the initial identification result of the corresponding evolution stage. For example, if the mode category identifier for evolution stage 1 is mode category 1, corresponding to the stable mode category set, a stability enhancement factor of 1.2 is added. If the mode category identifier corresponds to the decaying mode category set, a decay enhancement factor is added to the initial identification result of the corresponding evolution stage. For example, if the mode category identifier for evolution stage 2 is mode category 3, corresponding to the decaying mode category set, a decay enhancement factor of 1.5 is added.

[0063] Step 1527: Integrate the preliminary identification results with the stability enhancement factor or the decline enhancement factor, and calculate the final classification score for each evolution stage based on the deviation of the stability comprehensive evaluation index from the stability threshold range and the weight value of the stability enhancement factor or the decline enhancement factor; set a classification boundary value according to the final classification score, classify the evolution stage with a score higher than the classification boundary value as the stable stage, classify the evolution stage with a score lower than or equal to the classification boundary value as the decline stage, and generate a state stage classification output table containing classification labels for each evolution stage.

[0064] In this embodiment, the computer device integrates the preliminary identification results with the stability enhancement factor or the decay enhancement factor. Based on the deviation of the comprehensive stability evaluation index from the stability threshold range and the weight values ​​of the stability enhancement factor or the decay enhancement factor, the final classification score for each evolution stage is calculated. For example, if the comprehensive stability evaluation index for evolution stage 1 is 0.25, the deviation is 0.05, and the weight value of the stability enhancement factor is 0.8, then the final classification score is 0.25 × (1 - 0.05) × 0.8; if the comprehensive stability evaluation index for evolution stage 2 is 0.4, the deviation is 0.1, and the weight value of the decay enhancement factor is 0.9, then the final classification score is 0.4 × (1 + 0.1) × 0.9. A classification boundary value of 0.3 is set based on the final classification score. Evolution stages with scores higher than the classification boundary value are classified as stable stages. For example, if the final classification score for evolution stage 1 is lower than the classification boundary value, it is classified as a decay stage; if the final classification score for evolution stage 2 is higher than the classification boundary value, it is classified as a stable stage. Finally, a state stage classification output table containing classification labels for each evolution stage is generated.

[0065] Step 153: Based on the category feature parameters of the pattern category corresponding to each evolution stage, determine the conversion law between the pattern categories corresponding to different evolution stages, and obtain the temporal feature evolution mode corresponding to each evolution stage. The temporal feature evolution mode covers the steady evolution mode of the stable stage and the gradual decay mode and rapid decay mode of the decay stage.

[0066] In this embodiment, the computer device determines the transition rules between pattern categories corresponding to different evolutionary stages based on the category feature parameters of the pattern categories corresponding to each evolutionary stage. Category feature parameters include the mean, variance, maximum value, and minimum value of the pattern categories. The transition rules include the transition order, transition probability, and transition time interval between pattern categories. For example, in the stable stage, the transition order from pattern category 1 to pattern category 2 is sequential, with a transition probability of 0.2 and a transition time interval of 10 time units; in the decay stage, the transition order from pattern category 3 to pattern category 4 is random, with a transition probability of 0.5 and a transition time interval of 5 time units. Based on the transition rules, the temporal feature evolution patterns corresponding to each evolutionary stage are obtained. The temporal feature evolution pattern in the stable stage is a steady evolution pattern, with performance indicators changing slowly and stably; the temporal feature evolution pattern in the decay stage is a gradual decay pattern or a rapid decay pattern, with performance indicators changing rapidly and unstable.

[0067] Step 154: Perform correlation analysis on the temporal characteristic evolution patterns corresponding to each evolution stage and various influencing factors in the preset influencing factor set, and construct a correlation matrix. The correlation matrix represents the correlation strength between different temporal characteristic evolution patterns and various performance degradation causes.

[0068] In this embodiment, the preset set of influencing factors for computer equipment includes performance degradation factors such as temperature, voltage, load, and aging time. Then, a correlation analysis is performed between the temporal characteristic evolution patterns corresponding to each evolution stage and various influencing factors in the preset set of influencing factors, and the correlation matrix is ​​calculated using the Pearson correlation coefficient method. The Pearson correlation coefficient method calculates the correlation coefficient by calculating the ratio of the covariance to the standard deviation between two variables. The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation between the two variables. For example, the correlation coefficient between the stationary evolution pattern and temperature is 0.1, and the correlation coefficient with voltage is 0.2; the correlation coefficient between the gradual degradation pattern and load is 0.5, and the correlation coefficient with aging time is 0.6. The elements in the correlation matrix represent the correlation strength between different temporal characteristic evolution patterns and various performance degradation factors.

[0069] Step 155: Based on the correlation matrix, select the target performance degradation causes that have a set correlation strength with each time series feature evolution mode, and establish a corresponding mapping relationship by combining the target performance degradation causes and their corresponding time series feature evolution modes. The corresponding mapping relationship is used to reflect the root cause of performance degradation behind different performance change modes.

[0070] In this embodiment, the computer device sets the correlation strength to 0.5, filtering out target performance degradation causes whose correlation strength with each time-series feature evolution mode reaches the set strength. For example, the correlation strength between the steady evolution mode and temperature is 0.1, which does not reach the set strength, so it is not filtered; the correlation strength between the gradual decay mode and load is 0.5, which reaches the set strength, so it is filtered as a target performance degradation cause. Then, a corresponding mapping relationship is established by combining the target performance degradation causes and their corresponding time-series feature evolution modes. For example, the gradual decay mode corresponds to the load, and the rapid decay mode corresponds to the aging time. The corresponding mapping relationship is used to reflect the root causes of performance degradation behind different performance change modes.

[0071] Step 160: Generate a dynamic prediction strategy by combining the corresponding mapping relationship and the operating scenario characteristics of the LPDDR particle, and output targeted optimization guidance information for the LPDDR particle according to the dynamic prediction strategy. The targeted optimization guidance information is used to adapt to targeted performance maintenance and adjustment for different operating scenarios.

[0072] In this embodiment of the application, the computer device generates a dynamic prediction strategy by combining the corresponding mapping relationship obtained in step 150 and the operating scenario characteristics of the LPDDR particles, specifically including the following sub-steps: Step 161: Based on the corresponding mapping relationship, extract the performance degradation causes and correlation strengths corresponding to different time-series feature evolution patterns, and construct the corresponding mapping rule set and correlation dictionary. The mapping rule set is used to define the correspondence rules between time-series feature evolution patterns and performance degradation causes, and the correlation dictionary records the degree of influence and response priority of each performance degradation cause.

[0073] In this embodiment, the computer device first extracts the performance degradation causes and their correlation strengths corresponding to different temporal feature evolution modes from the corresponding mapping relationship. For example, there are no performance degradation causes for the steady evolution mode; the performance degradation cause for the gradual degradation mode is load, with a correlation strength of 0.5; and the performance degradation cause for the rapid degradation mode is aging time, with a correlation strength of 0.6. Then, a mapping rule set is constructed, containing the correspondence rules between temporal feature evolution modes and performance degradation causes. For example, rule 1: if the temporal feature evolution mode is a gradual degradation mode, then the performance degradation cause is load; rule 2: if the temporal feature evolution mode is a rapid degradation mode, then the performance degradation cause is aging time. Simultaneously, an association dictionary is constructed, recording the impact degree and response priority corresponding to each performance degradation cause. The impact degree is determined based on the correlation strength; the higher the correlation strength, the greater the impact degree. The response priority is determined based on the impact degree; the greater the impact degree, the higher the response priority. For example, the impact degree of load is medium, and the response priority is 2; the impact degree of aging time is high, and the response priority is 1.

[0074] Step 162: Collect the operating scenario features of the LPDDR chip during actual operation, extract the load mode and environmental parameters from the operating scenario features, and generate corresponding scenario classification labels according to different combinations of the load mode and the environmental parameters.

[0075] In this embodiment, the computer device collects the operating scenario characteristics of the LPDDR chip during actual operation, including load mode and environmental parameters. Load modes include high load, medium load, and low load; environmental parameters include temperature, voltage, and humidity. Then, the load mode and environmental parameters are extracted from the operating scenario characteristics, and corresponding scenario classification labels are generated based on different combinations of load mode and environmental parameters. For example, if the load mode is high load and the environmental parameters are temperature 30℃ and voltage 1.2V, the generated scenario classification label is "High Load_30℃_1.2V"; if the load mode is low load and the environmental parameters are temperature 25℃ and voltage 1.1V, the generated scenario classification label is "Low Load_25℃_1.1V".

[0076] Step 163: Match the scene classification label with the mapping rule set and the associated dictionary, combine the current temporal feature evolution mode of the LPDDR particle, predict the potential performance degradation causes and potential performance change trends of the LPDDR particle, and construct an adaptive early warning strategy based on the prediction results.

[0077] In this embodiment of the application, the computer device matches scene classification labels with mapping rule sets and associated dictionaries, specifically including the following sub-steps: Step 1631: Parse the load pattern code and environment parameter vector in the scene classification label, and traverse the set of rule entries in the mapping rule set that match the combination of the load pattern code and environment parameter vector. The set of rule entries includes a list of associated time-series feature evolution pattern types and a performance degradation cause index sequence.

[0078] In this embodiment, the computer device parses the load pattern code and environmental parameter vector from the scene classification label. The load pattern code includes high load code, medium load code, low load code, etc., and the environmental parameter vector includes numerical values ​​of parameters such as temperature, voltage, and humidity. Then, the set of rule entries corresponding to the combination of load pattern code and environmental parameter vector is traversed in the mapping rule set. The set of rule entries contains a list of associated temporal feature evolution mode types and a performance degradation cause index sequence. For example, if the scene classification label is "high load_30℃_1.2V", the corresponding set of rule entries contains a temporal feature evolution mode type list of [gradual degradation mode, rapid degradation mode] and a performance degradation cause index sequence of [load, aging time].

[0079] Step 1632: Retrieve the descriptive information corresponding to the performance degradation cause index sequence from the associated dictionary. The descriptive information includes the impact intensity value and response priority ranking of each performance degradation cause.

[0080] In this embodiment, the computer device retrieves descriptive information corresponding to the performance degradation cause index sequence from an associated dictionary. The descriptive information includes the impact intensity value and response priority ranking of each performance degradation cause. For example, if the performance degradation cause index sequence is [load, aging time], the corresponding descriptive information is that the impact intensity value of load is 0.5 and the response priority is 2; the impact intensity value of aging time is 0.6 and the response priority is 1.

[0081] Step 1633: Obtain the current timing feature evolution mode of the LPDDR particle, calculate the similarity matrix between the timing feature evolution mode and the timing feature evolution mode type list in the rule entry set, filter out timing feature evolution mode types with similarity exceeding the threshold as matching mode types based on a preset similarity threshold, and extract the corresponding performance degradation cause index as a candidate performance degradation cause set.

[0082] In this embodiment, the computer device acquires the current timing feature evolution mode of the LPDDR chip. Then, it calculates a similarity matrix between the timing feature evolution mode and the timing feature evolution mode type list in the rule entry set. The similarity is calculated using cosine similarity, which represents the degree of similarity between two timing feature evolution modes. Based on a preset similarity threshold (e.g., 0.7), timing feature evolution mode types with similarity exceeding the threshold are selected as matching mode types, and the corresponding performance degradation cause indexes are extracted as a candidate performance degradation cause set. For example, if the current timing feature evolution mode has a similarity of 0.8 with the gradual decay mode, exceeding the threshold, the matching mode type is the gradual decay mode, and the candidate performance degradation cause set is [load].

[0083] Step 1634: Based on the influence intensity value of the candidate performance degradation cause set, and combined with the historical evolution trajectory data of the time-series feature evolution mode, infer the performance change trend curve of the LPDDR particle within a future preset time window through a time series prediction model. The performance change trend curve includes the predicted sequence of performance index values ​​and its confidence band.

[0084] In this embodiment, the computer device, based on the influence intensity values ​​of a set of candidate performance degradation causes and combined with historical evolution trajectory data of time-series feature evolution patterns, infers the performance change trend curve of LPDDR particles within a future preset time window using a time series prediction model. The time series prediction model employs an LSTM model, which can capture long-term dependencies in time-series data. The performance change trend curve includes a predicted sequence of performance index values ​​and its confidence band. For example, if the future preset time window is 1 hour, the predicted sequence of the performance change trend curve is a decrease in read / write bandwidth from 800MB / s to 700MB / s, with a confidence band of ±50MB / s.

[0085] Step 1635: Based on the performance change trend curve, analyze the time point when the performance index value drops beyond the preset baseline, and in conjunction with the response priority ranking, set multi-level early warning trigger conditions. The multi-level early warning trigger conditions are associated with threshold combinations corresponding to different drop amplitudes and drop rates.

[0086] In this embodiment, the computer device analyzes the point in time when the performance index value drops beyond a preset baseline based on the performance change trend curve. The preset baseline is 90% of the initial value of the performance index. For example, if the initial value of the read / write bandwidth is 800MB / s, the preset baseline is 720MB / s, and the point in time when the performance index value drops beyond the preset baseline is 30 minutes later. Multi-level warning trigger conditions are set based on response priority ranking. These multi-level warning trigger conditions are associated with threshold combinations corresponding to different decrease magnitudes and decrease rates. For example, a level one warning trigger condition is a decrease magnitude exceeding 10% and a decrease rate exceeding 5% / minute, while a level two warning trigger condition is a decrease magnitude exceeding 20% ​​and a decrease rate exceeding 10% / minute.

[0087] Step 1636: Construct an early warning response logic topology, which maps the multi-level early warning triggering conditions to corresponding early warning action sequences. The early warning action sequences include an arrangement of data recording instructions, notification generation instructions, or parameter adjustment instructions.

[0088] In this embodiment, the computer device constructs an early warning response logical topology. This topology maps multi-level early warning triggering conditions to corresponding early warning action sequences. The early warning action sequence includes an arrangement of data recording instructions, notification generation instructions, or parameter adjustment instructions. For example, the early warning action sequence corresponding to a first-level early warning triggering condition is a data recording instruction followed by a notification generation instruction; the early warning action sequence corresponding to a second-level early warning triggering condition is a data recording instruction followed by a notification generation instruction followed by a parameter adjustment instruction.

[0089] Step 1637: Based on the multi-level early warning triggering conditions and early warning response logic topology, create an initial adaptive early warning strategy set; use historical fragments of the performance status monitoring data throughout the entire life cycle of the LPDDR particles to perform simulation evaluation on the initial adaptive early warning strategy set, adjust the sensitivity of the preset baseline and the execution timing of the early warning action sequence, and optimize the response efficiency of the initial adaptive early warning strategy set; configure the optimized initial adaptive early warning strategy set as the adaptive early warning strategy.

[0090] In this embodiment, the computer device creates an initial adaptive early warning strategy set based on multi-level early warning triggering conditions and early warning response logic topology. Then, historical segments of performance status monitoring data throughout the entire lifecycle of the LPDDR chips are used to simulate and evaluate the initial adaptive early warning strategy set. The simulation evaluation includes calculating indicators such as early warning accuracy, false alarm rate, and response latency. Based on the simulation evaluation results, the sensitivity of the preset baseline and the execution timing of the early warning action sequence are adjusted to optimize the response efficiency of the initial adaptive early warning strategy set. For example, the sensitivity of the preset baseline is adjusted to 1.2 times the original value, and the execution timing of the early warning action sequence is advanced by 10 seconds. Furthermore, the optimized initial adaptive early warning strategy set is configured as an adaptive early warning strategy.

[0091] Step 164: Invoke the preset strategy adjustment algorithm, combine the performance status monitoring data and historical evolution patterns of the LPDDR particles throughout their entire life cycle, optimize the warning threshold and response logic in the adaptive warning strategy, and use the optimized adaptive warning strategy as the dynamic prediction strategy.

[0092] In this embodiment, the computer device invokes a preset strategy adjustment algorithm, combining performance status monitoring data and historical evolution patterns throughout the entire lifecycle of the LPDDR chip, to optimize the warning threshold and response logic in the adaptive warning strategy. Specifically, this includes the following sub-steps: Step 1641: Load the adaptive early warning strategy, parse the early warning threshold parameter set and response logic rule network of the adaptive early warning strategy. The early warning threshold parameter set includes the threshold values ​​corresponding to the early warning trigger conditions of each level, and the response logic rule network includes the execution conditions and sequence relationship of the early warning action sequence.

[0093] In this embodiment, the computer device loads an adaptive early warning strategy and parses the early warning threshold parameter set and response logic rule network of the adaptive early warning strategy. The early warning threshold parameter set includes threshold values ​​corresponding to the early warning trigger conditions at each level. For example, the threshold value for a level one early warning trigger condition is a decrease of 10% at a decrease rate of 5% / minute; the threshold value for a level two early warning trigger condition is a decrease of 20% at a decrease rate of 10% / minute. The response logic rule network includes the execution conditions and sequence relationships of the early warning action sequence. For example, the execution condition is that the performance indicator value decreases beyond a preset baseline, and the sequence relationship is data recording instruction -- notification generation instruction -- parameter adjustment instruction.

[0094] Step 1642: Call the preset genetic algorithm as the strategy adjustment algorithm to encode the set of early warning threshold parameters and the response logic rule network into a chromosome individual representation, wherein the chromosome individual representation contains parameter gene segments and logical gene segments.

[0095] In this embodiment, the computer device invokes a preset genetic algorithm as the strategy adjustment algorithm. A genetic algorithm is an optimization algorithm that simulates natural selection and genetic mechanisms, capable of finding the optimal solution in a complex search space. The set of warning threshold parameters and the network of response logic rules are encoded as chromosome individual representations, each containing parameter gene segments and logical gene segments. The parameter gene segments encode the numerical values ​​of the warning threshold parameter set, and the logical gene segments encode the execution conditions and order relationships of the response logic rule network.

[0096] Step 1643: Extract performance data subsets from multiple historical time periods from the performance status monitoring data throughout the entire life cycle of the LPDDR particles, and use them as the algorithm training dataset; construct a performance degradation simulation environment based on the historical evolution law, evaluate the performance of the early warning strategy corresponding to the chromosome individual in the performance degradation simulation environment, and calculate the fitness score, which is a weighted index of early warning accuracy, false alarm rate and response latency.

[0097] In this embodiment, the computer device extracts performance data subsets from multiple historical time periods from the performance status monitoring data throughout the entire lifecycle of LPDDR particles, using these subsets as the algorithm training dataset. For example, a subset of performance data from the past year is extracted as the training dataset. A performance degradation simulation environment is constructed based on historical evolution patterns, capable of simulating the performance degradation process under different scenarios. The performance of the early warning strategy corresponding to the chromosome individual is evaluated in the performance degradation simulation environment, and a fitness score is calculated. The fitness score is a weighted index combining early warning accuracy, false alarm rate, and response latency. For example, the weight of early warning accuracy is 0.5, the weight of false alarm rate is 0.3, and the weight of response latency is 0.2.

[0098] Step 1644: Iteratively optimize the chromosome individual population through selection, crossover, and mutation operations of a genetic algorithm to generate a new generation of early warning strategy encoding sequences; decode the optimized chromosome individuals to obtain an updated set of early warning threshold parameters and a response logic rule network; integrate the updated set of early warning threshold parameters and the response logic rule network into the adaptive early warning strategy, replacing the original set of early warning threshold parameters and the response logic rule network, to obtain an optimized and adjusted adaptive early warning strategy.

[0099] In this embodiment, a computer device iteratively optimizes a population of chromosome individuals through selection, crossover, and mutation operations using a genetic algorithm to generate a new generation of early warning strategy encoding sequences. The selection operation chooses chromosome individuals with high fitness scores, the crossover operation exchanges gene segments between chromosome individuals, and the mutation operation randomly alters gene segments of chromosome individuals. Decoding the optimized chromosome individuals yields an updated set of early warning threshold parameters and a response logic rule network. The updated set of early warning threshold parameters and the response logic rule network are then integrated into an adaptive early warning strategy, replacing the original set of early warning threshold parameters and the response logic rule network, resulting in an optimized and adjusted adaptive early warning strategy.

[0100] Step 1645: Use an independent verification subset of the performance status monitoring data throughout the entire lifecycle of the LPDDR chip to perform performance testing on the optimized adaptive early warning strategy, and output the adaptive early warning strategy that passes the performance test as the dynamic prediction strategy.

[0101] In this embodiment, the computer device uses an independent validation subset of performance status monitoring data throughout the entire lifecycle of the LPDDR chips to perform performance testing on the optimized adaptive early warning strategy. The independent validation subset does not overlap with the training dataset, ensuring the objectivity of the test results. Performance testing includes calculating indicators such as early warning accuracy, false alarm rate, and response latency. If the performance test results meet preset performance requirements (e.g., early warning accuracy exceeding 90%, false alarm rate below 5%, and response latency below 1 second), the optimized adaptive early warning strategy is output as a dynamic prediction strategy.

[0102] Step 165: Based on the dynamic prediction strategy, combined with the load mode and environmental parameters corresponding to the scenario classification label, generate parameter adjustment suggestions for the LPDDR particles, and generate a corresponding maintenance operation task set based on the priority of the performance degradation causes; generate an optimization action sequence based on the parameter adjustment suggestions and the maintenance operation task set as targeted optimization guidance information for the LPDDR particles.

[0103] In this embodiment, the computer device generates parameter adjustment suggestions based on a dynamic prediction strategy, combined with the load mode and environmental parameters corresponding to the scene classification label. For example, if the scene classification label is "high load_30℃_1.2V", the warning threshold in the dynamic prediction strategy is 80% load, and the parameter adjustment suggestion is to reduce the load to below 70%. Simultaneously, a maintenance operation task set is generated based on the priority of responses to performance degradation causes. Maintenance operation tasks corresponding to performance degradation causes with higher priority are executed first. For example, the priority for responding to aging time is 1, and the priority for responding to load is 2; therefore, the maintenance operation task set first includes reducing the load, and then includes replacing the aged LPDDR chips. Based on the parameter adjustment suggestions and the maintenance operation task set, an optimization action sequence is generated. The optimization action sequence is arranged in chronological order, including the time of parameter adjustment, the time of maintenance operation, and the specific operation performed. For example, the optimization action sequence is: reduce the load to below 70% at time t1, and replace the aged LPDDR chips at time t2. The optimization action sequence serves as targeted optimization guidance information for LPDDR chips, used to adapt to targeted performance maintenance and adjustment for different operating scenarios.

[0104] As an optional embodiment, the method further includes: Step 210: Obtain the set of performance constraint rules and the set of key performance event annotations associated with the LPDDR chip model and operating scenario from the preset domain knowledge base; perform rule matching processing on the time-series feature sequence and the set of performance constraint rules, identify abnormal segment sequences that violate performance constraints in the time-series feature sequence and mark the corresponding constraint violation tags.

[0105] In this embodiment, the computer device first obtains a set of performance constraint rules and a set of key performance event annotations associated with the LPDDR chip model and operating scenario from a preset domain knowledge base. The set of performance constraint rules includes constraints such as the minimum read / write bandwidth, the maximum latency, and the maximum error correction code count; the set of key performance event annotations includes features and timestamps of key performance events such as initialization completion, performance degradation, and fault occurrence. Then, the time-series feature sequence obtained in step 120 is matched with the set of performance constraint rules. The rule matching process is implemented using a rule engine, which matches each feature vector in the time-series feature sequence with a rule in the set of performance constraint rules. If a feature vector violates a rule, it is marked as an abnormal segment sequence and labeled with the corresponding constraint violation tag. For example, if the latency of a feature vector in the time-series feature sequence exceeds the maximum value in the set of performance constraint rules, it is marked as an abnormal segment sequence, and the constraint violation tag is "excessive latency".

[0106] Step 220: Perform event matching processing on the time-series feature sequence and the key performance event annotation set to identify significant event fragment sequences in the time-series feature sequence that meet the key performance event characteristics and label them with the corresponding event type labels.

[0107] In this embodiment, the computer device performs event matching processing between the time-series feature sequence and the key performance event annotation set. The event matching processing is implemented using a pattern recognition algorithm, which matches each feature vector in the time-series feature sequence with the key performance event features in the key performance event annotation set. If a feature vector matches the characteristics of a certain key performance event, it is marked as a significant event segment sequence, and the corresponding event type label is assigned. For example, if the read / write bandwidth of a feature vector in the time-series feature sequence suddenly increases, matching the characteristic of initialization completion, it is marked as a significant event segment sequence, and the event type label is "initialization completed".

[0108] Step 230: Based on the abnormal fragment sequence and its constraint violation label, the significant event fragment sequence and its event type label, perform knowledge annotation processing on the original temporal feature sequence to generate a composite knowledge sequence containing rule knowledge and event knowledge; perform knowledge graph modeling processing on the composite knowledge sequence to construct a performance knowledge graph composed of temporal nodes and knowledge relationship edges. The temporal nodes are generated by fusing feature points in the temporal feature sequence with the constraint violation label and event type label, and the knowledge relationship edges represent the knowledge logical association between different temporal nodes.

[0109] In this embodiment, the computer device performs knowledge annotation processing on the original temporal feature sequence based on the abnormal fragment sequence and its constraint violation label, and the significant event fragment sequence and its event type label. The knowledge annotation process adds the constraint violation label and event type label to the corresponding feature vector, generating a composite knowledge sequence containing rule knowledge and event knowledge. For example, the feature vector in the abnormal fragment sequence is labeled with the constraint violation label "excessive delay," and the feature vector in the significant event fragment sequence is labeled with the event type label "initialization complete." Then, the composite knowledge sequence is subjected to knowledge graph modeling processing to construct a performance knowledge graph. The performance knowledge graph consists of temporal nodes and knowledge relationship edges. Temporal nodes are generated by fusing feature points from the temporal feature sequence with constraint violation labels and event type labels, and knowledge relationship edges represent the logical knowledge associations between different temporal nodes. For example, temporal node 1 contains feature point 1 and the constraint violation label "excessive delay," temporal node 2 contains feature point 2 and the event type label "initialization complete," and the knowledge relationship edges represent the causal relationship between temporal node 1 and temporal node 2.

[0110] Step 240: Extract time-series node subgraphs with knowledge logical associations from the performance knowledge graph, map the time-series node subgraphs back to the time-series dimension, reconstruct the feature expression sequence guided by the fusion domain knowledge, and use the feature expression sequence guided by the fusion domain knowledge as a new time-series feature sequence to update the input of pattern recognition and evolutionary analysis.

[0111] In this embodiment, the computer device extracts a temporal node subgraph with logical knowledge relationships from a performance knowledge graph. These logical knowledge relationships include causal relationships, temporal relationships, and associative relationships. For example, a temporal node subgraph containing the constraint violation label "excessive delay" and the event type label "initialization complete" is extracted. Then, the temporal node subgraph is mapped back to the temporal dimension to reconstruct a feature expression sequence guided by fused domain knowledge. This feature expression sequence contains feature points from the original temporal feature sequence and information about logical knowledge relationships. Using this fused domain knowledge-guided feature expression sequence as a new temporal feature sequence updates the input to pattern recognition and evolutionary analysis, improving the accuracy of pattern recognition and evolutionary analysis.

[0112] As an optional embodiment, the method further includes: Step 310: Obtain the newly added performance status monitoring data of the LPDDR particles collected in the current batch; extract the newly added time-series feature sequence fragments based on the newly added performance status monitoring data; input the newly added time-series feature sequence fragments into the constructed mode evolution path, and calculate the membership distribution of the newly added time-series feature sequence fragments with each existing mode category in the mode evolution path.

[0113] In this embodiment, the computer device acquires newly added performance status monitoring data of the LPDDR chips collected in the current batch. The newly added performance status monitoring data includes indicators such as read / write bandwidth, latency, and error correction code count within the current time period. Then, based on the newly added performance status monitoring data, newly added time-series feature sequence fragments are extracted. The extraction method is the same as in step 120, using a sliding window method to extract features in the time and frequency domains, generating multi-dimensional feature vectors, and arranging them in chronological order to obtain the newly added time-series feature sequence fragments. Next, the newly added time-series feature sequence fragments are input into the constructed pattern evolution path, and the membership degree distribution between the newly added time-series feature sequence fragments and each existing pattern category in the pattern evolution path is calculated. The membership degree distribution is calculated using the fuzzy C-means clustering algorithm. The fuzzy C-means clustering algorithm obtains the membership value by calculating the distance between each feature vector and the center of the pattern category. The closer the membership degree value is to 1, the more the feature vector belongs to that pattern category.

[0114] Step 320: Based on the membership distribution, determine whether the newly added time-series feature sequence fragment belongs to any existing pattern category. If it belongs entirely to an existing pattern category, extend the time span of that existing pattern category to cover the newly added time-series feature sequence fragment. If the membership distribution shows a dispersed characteristic and does not clearly belong to any single existing pattern category, mark the newly added time-series feature sequence fragment as a candidate new pattern fragment.

[0115] In this embodiment, the computer device determines whether a newly added temporal feature sequence fragment belongs to any existing pattern category based on the membership degree distribution. If the membership degree values ​​of all feature vectors in the newly added temporal feature sequence fragment exceed a preset threshold (e.g., 0.8) and all belong to the same existing pattern category, it is determined that the fragment completely belongs to a certain existing pattern category, and the time span of that existing pattern category is extended to cover the newly added temporal feature sequence fragment. For example, if the time span of existing pattern category 1 is [t0, t1), the time span of the newly added temporal feature sequence fragment is [t1, t2), and all feature vectors belong to pattern category 1, then the time span of pattern category 1 is extended to [t0, t2). If the membership degree values ​​of the feature vectors in the newly added temporal feature sequence fragment are scattered across multiple existing pattern categories, and the membership degree value of no single pattern category exceeds the preset threshold, it is determined that the membership degree distribution exhibits a dispersed characteristic and does not clearly belong to any single existing pattern category, and the newly added temporal feature sequence fragment is marked as a candidate new pattern fragment.

[0116] Step 330: Perform feature re-extraction and pattern abstraction processing on the candidate new pattern fragments to generate pattern feature descriptions of candidate new pattern categories; calculate the pattern distance matrix between the pattern feature descriptions of the candidate new pattern categories and the pattern feature descriptions of all existing pattern categories; based on the pattern distance matrix, if the pattern distance between the candidate new pattern category and all existing pattern categories is greater than a preset new pattern determination threshold, then the candidate new pattern category is determined as a valid new pattern category and inserted into the corresponding time position of the pattern evolution path.

[0117] In this embodiment, the computer device performs feature re-extraction and pattern abstraction on candidate new pattern fragments. Feature re-extraction uses the same method as in step 120 to extract features in the time domain and frequency domain, generating multi-dimensional feature vectors. Pattern abstraction uses principal component analysis (PCA) to reduce the multi-dimensional feature vectors to a lower dimension, generating pattern feature descriptions of candidate new pattern categories. The pattern feature descriptions include information such as the mean, variance, and principal components of the candidate new pattern categories. Then, the pattern distance matrix between the pattern feature descriptions of the candidate new pattern categories and the pattern feature descriptions of all existing pattern categories is calculated. The pattern distance is calculated using Euclidean distance, which represents the degree of difference between two pattern feature descriptions. Based on the pattern distance matrix, if the pattern distance between the candidate new pattern category and all existing pattern categories is greater than a preset new pattern determination threshold (e.g., 0.5), the candidate new pattern category is determined as a valid new pattern category and inserted into the corresponding time position in the pattern evolution path. For example, if the pattern distance between the candidate new pattern category and all existing pattern categories is greater than 0.5, it is inserted into the pattern evolution path at the time span of the newly added temporal feature sequence fragment.

[0118] Step 340: Based on the temporal transition characteristics between the candidate new mode fragment and the adjacent mode categories, update the relevant state transition probabilities and transition trajectory descriptions in the mode evolution path; based on the updated mode category set and state transition relationships, iteratively optimize the structure and parameters of the mode evolution path to generate a dynamically evolving and self-improving mode evolution path map.

[0119] In this embodiment, the computer device updates the state transition probabilities and transition trajectory descriptions related to the pattern evolution path based on the temporal transition characteristics between candidate new pattern fragments and adjacent pattern categories. The temporal transition characteristics include the trend and magnitude of feature vector changes. For example, if the feature vector change trend between a candidate new pattern fragment and the previous pattern category is upward with a magnitude of 0.3, and the feature vector change trend between it and the next pattern category is downward with a magnitude of 0.2, then the state transition probabilities and transition trajectory descriptions are updated. Then, based on the updated set of pattern categories and state transition relationships, the structure and parameters of the pattern evolution path are iteratively optimized. Iterative optimization employs a genetic algorithm, using operations such as selection, crossover, and mutation to optimize the structure (e.g., the number of pattern categories, the number of transition relationships) and parameters (e.g., state transition probabilities, transition time intervals) of the pattern evolution path. This generates a dynamically evolving and self-improving pattern evolution path map, which is continuously updated and improved with newly added performance status monitoring data, enhancing the accuracy and adaptability of the pattern evolution path.

[0120] As an optional embodiment, the method further includes: Step 410: Collect full lifecycle performance status monitoring data of multiple other LPDDR particles of the same model or type as the LPDDR particle, and construct a group performance time-series feature database; extract the time-series feature sequences corresponding to each particle from the group performance time-series feature database, and perform pattern evolution path construction processing on the time-series feature sequence of each particle to obtain a group pattern evolution path set.

[0121] In this embodiment, a computer device collects full-lifecycle performance status monitoring data of multiple LPDDR chips of the same model or type as the LPDDR chip, including indicators such as read / write bandwidth, latency, and error correction code count. Then, a group performance timing feature database is constructed, containing performance status monitoring data and corresponding timing feature sequences of multiple LPDDR chips. The timing feature sequences corresponding to each chip are extracted from the group performance timing feature database, and a pattern evolution path construction process is performed on the timing feature sequence of each chip, using the same method as step 130, resulting in a group pattern evolution path set. The group pattern evolution path set contains the pattern evolution paths of multiple LPDDR chips, and each pattern evolution path characterizes the performance status transition process of the corresponding LPDDR chip.

[0122] Step 420: Encode the dynamic prediction strategy currently generated by the LPDDR particle into a strategy feature vector, and retrieve several reference mode evolution paths with the highest similarity to the current mode evolution path of the LPDDR particle from the group mode evolution path set; extract the historical optimization strategy record corresponding to each reference mode evolution path, the historical optimization strategy record records the optimization measures taken at each stage of the reference mode evolution path and their effect feedback.

[0123] In this embodiment, the computer device encodes the dynamically predicted strategy currently generated by the LPDDR particle into a strategy feature vector. The strategy feature vector includes information such as the warning threshold, response logic, and parameter adjustment suggestions in the dynamic prediction strategy. Then, from the set of group pattern evolution paths, several reference pattern evolution paths with the highest similarity to the current pattern evolution path of the LPDDR particle are retrieved. The similarity is calculated using cosine similarity, which represents the degree of similarity between two pattern evolution paths. Historical optimization strategy records corresponding to each reference pattern evolution path are extracted. These records document the optimization measures taken at each stage of the reference pattern evolution path and their effect feedback. Optimization measures include parameter adjustments and maintenance operations, while effect feedback includes the magnitude of performance improvement and the number of failures reduced.

[0124] Step 430: Based on the strategy feature vector and the strategy features of each historical optimization strategy record, perform strategy feature matching and effect evaluation, and select a subset of target historical optimization strategies with significant effect gains and matching applicable conditions; perform strategy fusion and adaptive adjustment processing on the subset of target historical optimization strategies, and generate an enhanced hybrid prediction strategy set by combining the current specific operating scenario characteristics and performance degradation causes of the LPDDR particle.

[0125] In this embodiment, the computer device performs strategy feature matching and effect evaluation based on the strategy feature vector and the strategy features of each historical optimization strategy record. Strategy feature matching uses vector similarity calculation, and effect evaluation uses a weighted average calculation of effect feedback. A subset of target historical optimization strategies with significant effect gains (e.g., performance improvement exceeding 10%) and matching applicable conditions (e.g., identical operating scenario characteristics and performance degradation causes) is selected. The target subset of historical optimization strategies undergoes strategy fusion and adaptive adjustment. Strategy fusion uses a weighted average method to merge parameter adjustment suggestions and maintenance operation task sets from multiple historical optimization strategies. Adaptive adjustment combines the current specific operating scenario characteristics and performance degradation causes of the LPDDR granules to adjust the fused strategy parameters and rules. An enhanced hybrid prediction strategy set is generated, which contains multiple optimized prediction strategies, improving the accuracy and adaptability of dynamic prediction.

[0126] Step 440: Utilize recent performance status monitoring data of the LPDDR particles to perform strategy deduction and effect simulation on the enhanced hybrid prediction strategy set. Adjust the decision weights and action triggering logic in the strategy set based on the simulation results. Integrate the enhanced hybrid prediction strategy set optimized by simulation with the dynamic prediction strategy to generate a collaborative prediction strategy that integrates individual evolutionary patterns and group experience knowledge. Use the collaborative prediction strategy as the basis for generating the final output directional optimization guidance information.

[0127] In this embodiment, the computer device uses recent performance status monitoring data of LPDDR particles to perform strategy deduction and effect simulation on an enhanced hybrid prediction strategy set. Strategy deduction employs Monte Carlo simulation to simulate performance changes under different scenarios; effect simulation uses historical data verification to verify the accuracy and effectiveness of the prediction strategy. Based on the simulation results, the decision weights and action triggering logic in the strategy set are adjusted. The decision weights are adjusted according to the feedback, and the action triggering logic is adjusted according to scenario changes. The enhanced hybrid prediction strategy set optimized by simulation is integrated with the dynamic prediction strategy. The strategy integration uses a weighted average method to merge the parameter adjustment suggestions and maintenance operation task sets of the enhanced hybrid prediction strategy set and the dynamic prediction strategy. A collaborative prediction strategy that integrates individual evolutionary patterns and group experience knowledge is generated. This strategy can comprehensively integrate individual evolutionary patterns and group experience knowledge to improve the accuracy and reliability of targeted optimization guidance information. Using the collaborative prediction strategy as the basis for generating the final output targeted optimization guidance information, targeted optimization guidance information for LPDDR particles is generated for targeted performance maintenance and adjustment adapted to different operating scenarios.

[0128] When implementing the above technical solution, those skilled in the art will recognize that the solution involves processing, calculating, and fusing performance indicators and characteristic parameters with various physical meanings and dimensions. For example, characteristic parameters extracted from the time domain and frequency domain may have different dimensions and numerical scales. In subsequent steps such as clustering, trend analysis, calculation of comprehensive evaluation indicators (e.g., weighting the absolute value of the slope, the reciprocal of the volatility measure, and the magnitude of change), and strategy performance evaluation (e.g., comprehensive accuracy, false alarm rate, etc.), if the numerical values ​​of the original dimensions are directly manipulated, there may be problems such as dimension mismatch or the lack of clear physical meaning when adding parameters of different scales.

[0129] By employing adaptive normalization techniques, raw parameters from different sources and with different dimensions can be mapped to a unified dimensionless scale or a specific common scale. For example, through methods such as min-max normalization and Z-score standardization, differences in the dimensions and numerical ranges of different features can be eliminated, making parameters with different physical meanings comparable and operable in subsequent mathematical operations (such as weighted summation, distance calculation, and similarity comparison). This ensures the mathematical consistency and physical feasibility of a series of algorithmic steps, from multi-resolution feature vector extraction to stability comprehensive evaluation index calculation and policy fitness assessment.

[0130] Specifically, adaptive normalization of parameters with different physical meanings and dimensions to overcome dimensional errors is a common technique implemented by those skilled in the art based on existing technologies. In signal processing, data mining, and machine learning, normalization or standardization before processing multidimensional heterogeneous data is a standard preprocessing procedure. When a scheme involves comprehensive calculations (e.g., weighted product, similarity matrix construction, fitness score calculation) of multiple parameters such as time span, performance index values, rate of change, slope, volatility, correlation strength, and influence strength, technicians will naturally choose an appropriate normalization function based on the parameter characteristics. For example, for bounded parameters, min-max scaling to the normalization interval can be used; for parameters conforming to or approximately Gaussian distributions, Z-score standardization can be used to convert them to zero mean and unit variance; for proportional data, logarithmic transformation can be used, etc. These methods can transform parameters with different dimensions into pure numerical values, thus enabling subsequent steps such as cluster analysis, matrix operations, trend fitting, and fitness calculation in genetic algorithms to proceed smoothly without worrying about computational errors caused by inconsistent units. Furthermore, in the processes of constructing early warning thresholds and setting similarity thresholds, the normalized data makes the threshold settings more universal and interpretable.

[0131] Furthermore, this normalization process not only solves the problem of dimension matching but also often helps improve the convergence speed and generalization performance of the model. In the complex processes described in this solution, such as temporal feature pattern recognition, state transition probability calculation, correlation matrix construction, policy encoding and optimization, the internal mechanisms of the core algorithms (such as clustering algorithms, regression algorithms, genetic algorithms, etc.) usually assume that the input data are at a relatively consistent scale. Therefore, before executing the specific steps of the technical solution, systematically normalizing the original performance status monitoring data, extracted temporal features, and calculated various metrics is a recognized and essential basic operation in this field. Technical personnel can flexibly select and combine different normalization techniques according to the specific distribution of the data, the requirements of the algorithm, and the needs of the business logic. They can even design phased normalization strategies, such as first normalizing the original monitoring data according to the indicator type, and then performing a secondary scale adjustment on the derived comprehensive indicators. This ensures that even if each normalization step is not explicitly described in the solution document, in actual engineering implementation, technical personnel will rely on their professional knowledge and conventional practices to introduce an adaptive normalization module to guarantee the consistency of all mathematical operations in the entire analysis process, from data input to result output, in terms of dimensions and physical rationality. This overcomes potential problems such as "summing up different scale units," ensuring the feasibility and effectiveness of the overall technical solution. In short, by using the above mature normalization techniques, all the calculation, comparison, fusion, and optimization processes described in the solution can be implemented on a correct mathematical basis, without affecting the accuracy of the final analysis conclusions and guiding information due to confusion in dimensions or scales.

[0132] This application, through a comprehensive technical solution including full lifecycle performance status monitoring data acquisition, timing feature sequence extraction, pattern evolution path analysis, timing feature evolution graph construction, evolution pattern and decay cause mapping relationship mining, and dynamic prediction strategy generation, breaks through the limitations of traditional static timing feature analysis and realizes dynamic, full-cycle, multi-dimensional cognition and forward-looking control of LPDDR particle performance status. Specifically, by collecting monitoring data of multiple performance indicators under different operating scenarios, the system comprehensively covers the performance of particles throughout their entire lifecycle, from initial use to aging and degradation. The time-series feature sequences extracted from the monitoring data accurately reflect the continuous evolution of particle performance status. The pattern evolution paths obtained through time-series feature pattern recognition accurately characterize the regularity and anomalies of performance status changes, achieving qualitative and quantitative analysis of performance evolution trends. The time-series feature evolution diagram generated by combining abrupt change nodes and cyclical evolution trends intuitively presents the stage changes and time-series correlations of performance status. The mapping relationship between the mined time-series feature evolution patterns and performance degradation triggers reveals the intrinsic connection between different evolution stages and potential performance degradation roots, providing a scientific basis for performance risk prediction. The dynamic prediction strategies and targeted optimization guidance information generated based on the mapping relationship and operating scenario characteristics can be adapted to different operating scenarios for targeted performance maintenance and adjustment, realizing a shift from passive fault handling to proactive and forward-looking regulation.

[0133] This application realizes the dynamic evolution of LPDDR chip performance status, the precise correlation of performance degradation causes, and the forward-looking control of performance risks from an overall global perspective, providing more forward-looking guidance for LPDDR chip design optimization, lifespan extension, and dynamic performance scheduling of terminal devices.

[0134] Please see Figure 2 The figure is a schematic diagram of the basic structure of a computer device 200 provided in an embodiment of this application. The computer device 200 includes: a processor 201; a storage device 202 on which a computer program 2020 is stored; and a network interface 203 for providing network communication functions. When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the time-series-based LPDDR particle performance analysis methods described above.

[0135] Please see Figure 3 This application provides a functional block diagram of an LPDDR particle performance analysis device, which includes: The monitoring data acquisition module is used to collect performance status monitoring data of LPDDR chips throughout their entire life cycle. The performance status monitoring data covers a variety of performance indicators of the LPDDR chips under different operating scenarios. The time-series feature extraction module is used to extract time-series feature sequences that characterize the changes in performance indicators over time based on the performance status monitoring data. The time-series feature sequences reflect the continuous evolution process of the performance status of the LPDDR particles. An evolution path generation module is used to perform temporal feature pattern recognition on the temporal feature sequence to analyze the regularity and anomaly in the performance evolution and obtain the pattern evolution path that characterizes the performance state change. The feature evolution analysis module is used to determine the mutation nodes and cyclic evolution trends of the LPDDR particles in the gradual process through the pattern evolution path, and to generate a time-series feature evolution diagram that reflects the performance state stage changes and time-series correlation by combining the mutation nodes and the cyclic evolution trends. The mapping relationship mining module is used to mine the corresponding mapping relationship between the time-series feature evolution pattern of the LPDDR particle and the performance degradation cause based on each evolution stage in the time-series feature evolution diagram. The corresponding mapping relationship reveals the association between different evolution stages and potential performance degradation root causes. The guidance information output module is used to generate a dynamic prediction strategy by combining the corresponding mapping relationship and the operating scenario characteristics of the LPDDR particle, and output targeted optimization guidance information for the LPDDR particle according to the dynamic prediction strategy. The targeted optimization guidance information is used to adapt to targeted performance maintenance and adjustment for different operating scenarios.

[0136] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0137] Furthermore, it should be noted that this application also provides a computer program product, which may include a computer program that can be stored in a computer-readable storage medium. The processor of a computer device reads the computer program from the computer-readable storage medium, and the processor can execute the computer program, causing the computer device to perform the aforementioned... Figure 1 The methods described in the corresponding embodiments are already known, and therefore will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program product embodiments related to this application, please refer to the description of the method embodiments of this application.

[0138] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A method for LPDDR die performance analysis based on timing characteristics, characterized in that, The method includes: Collect performance status monitoring data of LPDDR chips throughout their entire life cycle. The performance status monitoring data covers a variety of performance indicators of the LPDDR chips under different operating scenarios. Based on the performance status monitoring data, a time-series feature sequence characterizing the performance index changes over time is extracted, and the time-series feature sequence reflects the continuous evolution process of the LPDDR particle performance status. Temporal feature pattern recognition is performed on the time-series feature sequence to analyze the regularity and anomaly in the performance evolution and obtain the pattern evolution path that characterizes the performance state change. The mutation nodes and cyclical evolution trends of the LPDDR particles in the gradual process are determined by the pattern evolution path. The mutation nodes and cyclical evolution trends are combined to generate a time-series feature evolution diagram that reflects the performance state stage changes and time-series correlation. Based on each evolution stage in the time-series feature evolution diagram, the corresponding mapping relationship between the time-series feature evolution pattern of the LPDDR particle and the performance degradation cause is explored. The corresponding mapping relationship reveals the association between different evolution stages and potential performance degradation root causes. A dynamic prediction strategy is generated by combining the corresponding mapping relationship and the operating scenario characteristics of the LPDDR particles. Based on the dynamic prediction strategy, targeted optimization guidance information for the LPDDR particles is output. The targeted optimization guidance information is used to adapt to targeted performance maintenance and adjustment for different operating scenarios.

2. The method of claim 1, wherein, The step of performing time-series feature pattern recognition on the time-series feature sequence to analyze the regularity and anomalies in performance evolution and obtain the pattern evolution path characterizing the performance state transition includes: Based on the time-series feature sequence, a sliding window of preset length is used to perform sliding segmentation on the time-series feature sequence to generate several continuous sliding window sub-sequences that have some overlap. Multi-dimensional feature extraction is performed on each sliding window subsequence to generate a corresponding multi-resolution feature vector, which covers feature parameters in the time domain and frequency domain. Clustering pattern classification is performed on the multi-resolution feature vectors corresponding to all sliding window sub-sequences to obtain several different pattern categories. Each pattern category corresponds to a performance change state of the LPDDR particle. Anomaly detection processing is performed on the clustered pattern categories to identify the first abnormal subsequence corresponding to the multi-resolution feature vector that deviates from the normal pattern category, and to determine the second abnormal subsequence corresponding to the sudden deviation and persistent anomaly in the time-series feature sequence. All pattern categories are sequentially sorted in chronological order, and the transition relationships and frequencies between different pattern categories are analyzed by combining the first and second abnormal subsequences. Based on the transition relationships and frequencies, the time node information corresponding to each pattern category is extracted to generate a state transition sequence. Based on the state transition sequence, a transition trajectory between different mode categories is generated. Combining the transition trajectory and the performance change characteristics corresponding to each sliding window subsequence, a mode evolution path that characterizes the performance state transition process of the LPDDR particle is constructed.

3. The method as described in claim 1, characterized in that, The process involves determining the abrupt change nodes and cyclical evolution trends of the LPDDR particles during the gradual evolution process through the pattern evolution path, and combining the abrupt change nodes and the cyclical evolution trends to generate a time-series feature evolution diagram reflecting the performance state stage transitions and timing correlations, including: Based on the evolution trajectory diagram corresponding to the mode evolution path, extract the state information and time node information corresponding to all mode categories, and construct the state transition matrix of the LPDDR particle performance state change. The state transition matrix represents the transition probability and number of transitions between different mode categories. Extract the path node sequence corresponding to each mode category transformation from the state transition matrix, perform trend analysis on the path node sequence, and identify nodes whose performance characteristic parameter change rate exceeds a preset threshold by calculating the change rate of each node in the path node sequence and determine them as inflection points in the gradual performance change process of the LPDDR particle. Each inflection point corresponds to a jump moment, and the nodes corresponding to all jump moments together constitute the mutation node. The path node sequence is subjected to periodic analysis to mine the repeated pattern category combinations in the path node sequence, determine the periodic cyclic characteristics of the performance state change of the LPDDR particle, and combine the performance change trend corresponding to each pattern category to analyze the performance change law in the periodic cycle process, thereby obtaining the cyclical evolution trend of the LPDDR particle performance gradual change process. Based on the jump time corresponding to the mutation node, the performance state change process of the LPDDR particle throughout its entire life cycle is divided into several continuous gradual intervals, each of which corresponds to a stable change stage of the LPDDR particle's performance state. A trend fitting process is performed on the path node sequence within each gradient interval to generate a corresponding trend fitting curve. The trend fitting curve reflects the changing trend of the performance state of the LPDDR particle within the corresponding gradient interval. A directed time series graph is constructed based on the segmented gradual intervals, mutation nodes, cyclical evolution trends, and trend fitting curves. The pattern category corresponding to each gradual interval, the jump time corresponding to each mutation node, and the periodic information of the cyclical evolution trend are marked in the directed time series graph. A stage transition network between each gradual interval is constructed simultaneously. The time series feature evolution graph is generated by combining the directed time series graph and the stage transition network.

4. The method as described in claim 1, characterized in that, The step of mining the mapping relationship between the time-series feature evolution pattern of the LPDDR particles and the performance degradation causes, based on each evolution stage in the time-series feature evolution diagram, includes: Based on the time-series feature evolution map, extract the stage transition network and the information corresponding to each gradual interval to generate the performance state change stage division map of the LPDDR particle. Combined with the mode category information corresponding to the mode evolution path, construct the time-series feature evolution topology map. The performance state change stage division map is used to mark the time range and core performance characteristics of each evolution stage. The evolution topology map is used to present the mode category and mode transformation relationship corresponding to each evolution stage. Based on the performance state change stage division diagram and the evolution topology diagram, each evolution stage is classified and processed. Combining the performance change trend and mode category corresponding to each evolution stage, the evolution stage is divided into a stable stage and a decline stage. Based on the category feature parameters of the pattern category corresponding to each evolution stage, the conversion law between the pattern categories corresponding to different evolution stages is determined, and the temporal feature evolution mode corresponding to each evolution stage is obtained. The temporal feature evolution mode covers the steady evolution mode of the stable stage and the gradual decay mode and rapid decay mode of the decay stage. A correlation analysis is performed on the temporal characteristic evolution patterns corresponding to each evolution stage and various influencing factors in the preset influencing factor set to construct a correlation matrix. The correlation matrix represents the correlation strength between different temporal characteristic evolution patterns and various performance degradation causes. Based on the correlation matrix, target performance degradation causes with a correlation strength reaching a set intensity with each time series feature evolution mode are selected. A corresponding mapping relationship is established by combining the target performance degradation causes and their corresponding time series feature evolution modes. The corresponding mapping relationship is used to reflect the root cause of performance degradation behind different performance change modes.

5. The method as described in claim 4, characterized in that, The evolutionary stages are classified according to the performance state change stage division diagram and the evolutionary topology diagram. Combining the performance change trends and pattern categories corresponding to each evolutionary stage, the evolutionary stages are divided into stable stages and decay stages, including: Based on the performance state change stage division diagram, extract the time span descriptor and key performance feature set for each evolution stage. The time span descriptor includes the start time point and end time point of each evolution stage, and the key performance feature set includes the values ​​and variation range of key performance indicators marked in the time-series feature evolution diagram for each evolution stage. Based on the evolutionary topology graph, retrieve the pattern category identifiers associated with each evolutionary stage and the connection topology of the pattern category identifiers in the stage transition network. The connection topology includes the transition probability and transition direction between pattern categories. Based on the start and end time points in the time span descriptor, the performance data subsequences corresponding to the time intervals are segmented from the time-series feature sequence, and each performance data subsequence constitutes a representative segment of the performance state of the corresponding evolution stage. For representative performance segments of each evolutionary stage, trend modeling is performed. A linear regression algorithm is used to fit a trend line, and the slope of the trend line is calculated as a trend slope metric. The average deviation between the data points of the representative performance segments and the trend line is calculated as a volatility metric. Combining the variation range of the key performance indicators in the key performance feature set within the evolutionary stage, a comprehensive stability evaluation index for each evolutionary stage is calculated. The comprehensive stability evaluation index is obtained by multiplying the absolute value of the trend slope metric by the reciprocal of the volatility metric, and then multiplying it by the normalization coefficient of the variation range. Set a stability threshold range, mark the evolution stage where the comprehensive stability evaluation index falls into the first stability threshold sub-range as a candidate for stable stage, and mark the evolution stage where the comprehensive stability evaluation index falls into the second stability threshold sub-range as a candidate for decay stage. Referring to the pattern category identifier, a preset pattern category classification mapping table is queried. If the pattern category identifier corresponds to the set of stable pattern categories in the preset pattern category classification mapping table, a stability enhancement factor is added to the preliminary identification result of the corresponding evolution stage. If the pattern category identifier corresponds to the set of decay pattern categories in the preset pattern category classification mapping table, a decay enhancement factor is added to the preliminary identification result of the corresponding evolution stage. Integrating the preliminary identification results with the stability enhancement factor or decline enhancement factor, and based on the deviation of the comprehensive stability evaluation index from the stability threshold range and the weight values ​​of the stability enhancement factor or decline enhancement factor, the final classification score for each evolution stage is calculated; a classification boundary value is set according to the final classification score, and evolution stages with scores higher than the classification boundary value are classified as stable stages, while evolution stages with scores lower than or equal to the classification boundary value are classified as decline stages, generating a state stage classification output table containing classification labels for each evolution stage.

6. The method as described in claim 1, characterized in that, The process involves generating a dynamic prediction strategy by combining the corresponding mapping relationship and the operating scenario characteristics of the LPDDR particles, and outputting targeted optimization guidance information for the LPDDR particles based on the dynamic prediction strategy, including: Based on the corresponding mapping relationship, the performance degradation causes and correlation strengths corresponding to different time-series feature evolution modes are extracted, and corresponding mapping rule sets and correlation dictionaries are constructed. The mapping rule set is used to define the correspondence rules between time-series feature evolution modes and performance degradation causes, and the correlation dictionary records the degree of influence and response priority of each performance degradation cause. The operating scenario features of the LPDDR chip during actual operation are collected, the load mode and environmental parameters in the operating scenario features are extracted, and corresponding scenario classification labels are generated according to different combinations of the load mode and the environmental parameters. The scene classification labels are matched with the mapping rule set and the associated dictionary. Combined with the current time-series feature evolution mode of the LPDDR particles, the potential performance degradation causes and potential performance change trends of the LPDDR particles are predicted. An adaptive early warning strategy is constructed based on the prediction results. The preset strategy adjustment algorithm is invoked, and the performance status monitoring data and historical evolution patterns of the LPDDR particles throughout their entire life cycle are combined to optimize the warning threshold and response logic in the adaptive warning strategy. The optimized and adjusted adaptive warning strategy is then used as the dynamic prediction strategy. Based on the dynamic prediction strategy, combined with the load mode and environmental parameters corresponding to the scenario classification tags, parameter adjustment suggestions are generated for the LPDDR particles, and a corresponding maintenance operation task set is generated based on the priority of addressing the performance degradation causes. An optimization action sequence is generated based on the parameter adjustment suggestions and the maintenance operation task set, and serves as targeted optimization guidance information for the LPDDR particles.

7. The method as described in claim 6, characterized in that, The process involves matching the scene classification labels with the mapping rule set and the associated dictionary, combining this with the current temporal feature evolution mode of the LPDDR particles, predicting the potential performance degradation causes and potential performance change trends of the LPDDR particles, and constructing an adaptive early warning strategy based on the prediction results, including: The load pattern code and environmental parameter vector in the scene classification label are parsed, and the set of rule entries corresponding to the combination of the load pattern code and environmental parameter vector are traversed in the mapping rule set. The set of rule entries includes a list of associated time-series feature evolution pattern types and a performance degradation cause index sequence. Retrieve descriptive information corresponding to the performance degradation cause index sequence from the associated dictionary. The descriptive information includes the impact intensity value and response priority ranking of each performance degradation cause. Obtain the current time-series feature evolution mode of the LPDDR particle, calculate the similarity matrix between the time-series feature evolution mode and the time-series feature evolution mode type list in the rule entry set, filter out the time-series feature evolution mode types with similarity exceeding the threshold as matching mode types based on a preset similarity threshold, and extract the corresponding performance degradation cause index as a candidate performance degradation cause set. Based on the influence intensity value of the candidate performance degradation cause set, combined with the historical evolution trajectory data of the time series feature evolution mode, the performance change trend curve of the LPDDR particle in the future preset time window is inferred by the time series prediction model. The performance change trend curve includes the predicted sequence of performance index values ​​and its confidence band. Based on the performance change trend curve, analyze the time points when the performance index value drops beyond the preset baseline, and combine the response priority ranking to set multi-level early warning trigger conditions. The multi-level early warning trigger conditions are associated with threshold combinations corresponding to different drop amplitudes and drop rates. Construct an early warning response logic topology, which maps the multi-level early warning triggering conditions to corresponding early warning action sequences, wherein the early warning action sequences include an arrangement of data recording instructions, notification generation instructions, or parameter adjustment instructions; Based on the multi-level early warning triggering conditions and early warning response logic topology, an initial adaptive early warning strategy set is created; using historical fragments of performance status monitoring data throughout the entire lifecycle of the LPDDR particles, the initial adaptive early warning strategy set is simulated and evaluated, the sensitivity of the preset baseline and the execution timing of the early warning action sequence are adjusted, and the response efficiency of the initial adaptive early warning strategy set is optimized; the optimized initial adaptive early warning strategy set is configured as the adaptive early warning strategy.

8. The method as described in claim 6, characterized in that, The method of invoking a preset strategy adjustment algorithm, combined with the performance status monitoring data and historical evolution patterns throughout the entire lifecycle of the LPDDR particles, optimizes the warning threshold and response logic in the adaptive warning strategy, and uses the optimized adaptive warning strategy as the dynamic prediction strategy, including: The adaptive early warning strategy is loaded, and the early warning threshold parameter set and response logic rule network of the adaptive early warning strategy are parsed. The early warning threshold parameter set includes the threshold values ​​corresponding to the early warning trigger conditions at each level, and the response logic rule network includes the execution conditions and sequence relationship of the early warning action sequence. A preset genetic algorithm is invoked as a strategy adjustment algorithm to encode the set of early warning threshold parameters and the response logic rule network into a chromosome individual representation, wherein the chromosome individual representation contains parameter gene segments and logical gene segments; Multiple historical time period performance data subsets are extracted from the performance status monitoring data of the LPDDR particles throughout their entire life cycle as the algorithm training dataset; a performance degradation simulation environment is constructed based on the historical evolution law, and the performance of the early warning strategy corresponding to the chromosome individual is evaluated in the performance degradation simulation environment, and the fitness score is calculated. The fitness score is a weighted index that comprehensively considers the early warning accuracy, false alarm rate and response latency. The chromosome population is iteratively optimized through selection, crossover, and mutation operations using a genetic algorithm to generate a new generation of early warning strategy encoding sequences. The optimized chromosomes are then decoded to obtain an updated set of early warning threshold parameters and a response logic rule network. The updated set of early warning threshold parameters and the response logic rule network are integrated into the adaptive early warning strategy, replacing the original set of early warning threshold parameters and the response logic rule network, to obtain an optimized and adjusted adaptive early warning strategy. The optimized adaptive warning strategy is tested using an independent validation subset of the performance status monitoring data throughout the entire lifecycle of the LPDDR particles. The adaptive warning strategy that passes the performance test is then output as the dynamic prediction strategy.

9. A computer device, characterized in that, include: A processor; a storage device having a computer program stored thereon; a network interface for providing network communication functions; when the computer program is executed by the processor, the processor enables the processor to implement the LPDDR particle performance analysis method based on timing characteristics as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the LPDDR particle performance analysis method based on timing characteristics as described in any one of claims 1-8.

Images