Chip fault intelligent detection method and system based on operation data

Abstract

The application provides a chip fault intelligent detection method and system based on operation data, and relates to the technical field of chip operation state analysis. The method comprises the following steps: obtaining historical operation state record data of a chip and determining a plurality of load states; constructing a local steady-state feature set corresponding to each load state and a local transition feature set corresponding to any two adjacent load states; performing sliding window analysis on each local steady-state feature set based on state coordination characteristics and quantile difference characteristics, and constructing a first fault detection strategy for steady-state anomaly detection of the chip; performing fluctuation feature analysis and fluctuation coexistence analysis based on load changes on each local transition feature set, and constructing a second fault detection strategy for state transition anomaly detection of the chip; and realizing real-time fault detection of the chip through the first fault detection strategy and the second fault detection strategy. The application improves the fault detection accuracy of the chip.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chip running state analysis, and particularly relates to a chip fault intelligent detection method and system based on running data. BACKGROUND

[0002] With the rapid development of integrated circuit technology, the complexity and integration of modern chips are increasing, and the running state health management has become the key to guarantee the reliability of electronic systems. Some chip fault detection technologies based on static threshold alarm or single parameter trend analysis have problems such as insufficient dynamic load adaptability and low sensitivity to small faults in actual application. When the chip is in some relatively stable use scenarios, the degree of change of the state parameters of the chip is relatively small, which leads to a high difficulty in identifying some potential small abnormalities. In some complex application scenarios, it may be necessary to frequently switch between different load states, and if the actual differences in the behaviors of the state parameters under different load states cannot be distinguished, false positives may be generated in complex scenarios. And some early faults such as capacitance attenuation and solder crack will first show subtle abnormalities in the complex conversion process of the load, such as voltage recovery delay, current overshoot anomaly and the like. In some high-reliability scenarios such as industrial Internet of Things and vehicle-mounted electronics, chip faults often need to be discovered early before they cause local system failure. There is an urgent need for a chip fault detection scheme for multi-state dynamic scenarios. SUMMARY

[0003] In view of the deficiencies in the above background art, the present application provides a chip fault intelligent detection method and system based on running data, which generates a fault detection strategy for analyzing small faults existing in the running process of the chip by analyzing the potential relationship between a plurality of state parameters under different running states of the chip, and can effectively face complex use scenarios of the chip and improve the identification accuracy of potential small abnormalities.

[0004] The first aspect of the present application provides a chip fault intelligent detection method based on running data, comprising:

[0005] Obtaining historical running state record data of a chip, determining a plurality of load states based on the running load of the chip, pre-dividing the historical running state record data, and constructing a state feature data set corresponding to each load state, and determining a state transition phase between two adjacent load states from the historical running state record data;

[0006] According to the plurality of state transition phases, the plurality of state feature data sets are state conversion segmented, a local steady state feature set corresponding to each load state and a local conversion feature set corresponding to any two adjacent load states are generated, and the local conversion feature set includes local conversion parameter record data corresponding to a plurality of state conversion events;

[0007] respectively, and a quantile difference matrix of each local steady-state feature set is extracted, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed;

[0008] respectively, and a quantile difference matrix of each local steady-state feature set is extracted, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed;

[0009] respectively, and a quantile difference matrix of each local steady-state feature set is extracted, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed;

[0010] respectively, and a quantile difference matrix of each local steady-state feature set is extracted, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed;

[0011] The local steady-state feature set is traversed through a preset state analysis window, and a state feature sequence corresponding to each feature window in the preset state analysis window is extracted, a covariance between state feature sequences corresponding to any two state parameters is calculated, a state coordination vector in each feature window is generated, and state coordination vectors corresponding to a plurality of feature windows are spliced in time sequence to generate a state coordination matrix of the local steady-state feature set;

[0012] According to the load state to which the local steady-state feature set belongs, a plurality of load quantile parameters of the local steady-state feature set are determined, and a state quantile interval of each local steady-state feature set is constructed according to the plurality of load quantile parameters;

[0013] During the process of traversing the local steady-state feature set through the preset state analysis window, a mean value feature vector corresponding to each state quantile interval is generated, a local correlation coefficient between mean value feature vectors corresponding to any two state parameters in each state quantile interval is calculated, a quantile correlation vector of each state quantile interval is constructed, and state coordination vectors corresponding to a plurality of state quantile intervals are spliced to generate a quantile difference matrix of the local steady-state feature set.

[0014] respectively, and a quantile difference matrix of each local steady-state feature set is extracted, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed;

[0015] The fluctuation amplitude sequences of the multiple state parameters in the local transition feature set are extracted, a data fluctuation splitting strategy is determined according to a load transition rate involved in the local transition feature set, multiple transition rate intervals of the local transition feature set are generated according to the data fluctuation splitting strategy, the fluctuation amplitude sequences of the state parameters with respect to the load change are reconstructed based on time sequences according to the transition rate intervals, a local discrete characteristic vector of each transition rate interval is generated under each fluctuation amplitude sequence, and a fluctuation feature matrix corresponding to each transition rate interval in the local transition feature set is constructed.

[0016] The multiple sets of local transition parameter record data in each local transition feature set are subjected to fluctuation co-occurrence analysis according to the data fluctuation splitting strategy, including assigning each set of local transition parameter record data to multiple transition rate intervals, extracting multiple fluctuation co-occurrence sequences of each set of local transition parameter record data in each belonging transition rate interval, and generating a fluctuation co-occurrence matrix of each set of local transition parameter record data.

[0017] Preferably, for the first fault detection strategy and the second fault detection strategy, further comprising:

[0018] The state parameter importance analysis is performed on each state quantile interval according to the quantile difference matrix of the local steady-state feature set, including calculating a global correlation parameter of each state parameter under the state quantile interval according to the quantile correlation vector of the state quantile interval, generating a collaborative reference list of each state quantile interval, calculating multiple global collaborative scores of each state parameter according to the state collaborative matrix of the local steady-state feature set, generating a collaborative evaluation range of each state parameter according to the multiple global collaborative scores, and generating a first fault detection strategy for steady-state anomaly detection of the chip through the collaborative evaluation range of the multiple state parameters and the collaborative reference list of the multiple state quantile intervals.

[0019] The state parameter importance analysis is performed on each transition rate interval according to the local discrete characteristic vector, including calculating a global fluctuation characteristic parameter of each state parameter under the transition rate interval according to the local discrete characteristic vector, generating a fluctuation reference list of each transition rate interval, calculating multiple global co-occurrence scores of each state parameter according to the multiple fluctuation co-occurrence matrices, generating a fluctuation evaluation range of each state parameter according to the multiple global co-occurrence scores, and generating a second fault detection strategy for state transition anomaly detection of the chip through the fluctuation evaluation range of the multiple state parameters and the fluctuation reference list of the multiple transition rate intervals.

[0020] Preferably, for the calculation of the global collaborative scores and the global co-occurrence scores, including:

[0021] The state coordination matrix is divided based on multiple reference periods, the local coordination average between any two state parameters under each reference period is calculated, and the global coordination score of the state parameters under the reference period is calculated according to the multiple local coordination averages of each state parameter under the reference period.

[0022] The multiple fluctuation symbiotic matrices are divided based on multiple reference periods, the local symbiotic parameters between any two state parameters in each fluctuation symbiotic matrix under each reference period are calculated, and the global symbiotic score of the state parameters under the reference period is calculated according to the multiple local symbiotic parameters of each state parameter under the reference period.

[0023] Preferably, the real-time fault detection of the chip is realized through the first fault detection strategy and the second fault detection strategy, including:

[0024] After obtaining the target running monitoring data of the chip, the target load state and the target state transition stage involved in the target running monitoring data are determined;

[0025] The target steady-state data of the target load state is extracted, the state quantile interval corresponding to the target steady-state data is determined, the steady-state evaluation weight of the multiple state parameters under the target steady-state data is determined according to the coordination reference list of the state quantile interval, the state coordination analysis of the multiple state parameters under the target steady-state data is performed, the target coordination parameter between any two state parameters under the target steady-state data is calculated, the first coordination score of each state parameter is calculated through the multiple target coordination parameters, the second coordination score is generated by modifying the first coordination score based on the steady-state evaluation weight of the state parameter, the state coordination evaluation of the target steady-state data is performed based on the coordination evaluation range of the multiple state parameters under the target load state, and the steady-state abnormal detection result of the chip is generated.

[0026] The target conversion data of the target load state is extracted, the conversion rate interval corresponding to the target conversion data is determined, the fluctuation evaluation weight of the multiple state parameters under the target conversion data is generated according to the fluctuation reference list of the conversion rate interval, the fluctuation characteristic analysis of the multiple state parameters under the target conversion data is performed, the target symbiotic parameter between any two state parameters is calculated, the first symbiotic score of each state parameter is calculated through the multiple target symbiotic parameters, the second symbiotic score is generated by modifying the first symbiotic score based on the fluctuation evaluation weight of the state parameter, and the state abnormal conversion detection result of the chip is generated by performing the fluctuation characteristic evaluation of the target conversion data based on the fluctuation evaluation range of the multiple state parameters under the target conversion data.

[0027] Preferably, the multiple fluctuation symbiotic sequences of each group of local conversion parameter record data under each belonging conversion rate interval are extracted, including:

[0028] Statistical analysis is performed on the local transition parameter record data to generate a plurality of statistical characteristics of each state parameter, a fluctuation threshold of each state parameter is determined according to the plurality of statistical characteristics, fluctuation amplitude-based sliding window detection is performed on the local transition parameter record data, a plurality of time nodes at which each state parameter is greater than the fluctuation threshold are determined, and a fluctuation co-occurrence sequence of each state parameter is generated.

[0029] The second aspect provides a chip fault intelligent detection system based on operation data, which is used to implement the chip fault intelligent detection method based on operation data.

[0030] The historical data preprocessing module is configured to obtain historical operation state record data of the chip, determine a plurality of load states based on operation load of the chip, pre-divide the historical operation state record data, and construct a state feature data set corresponding to each load state, and determine a state transition phase between adjacent two load states from the historical operation state record data.

[0031] The load state analysis module is configured to perform state transition segmentation on the plurality of state feature data sets according to the plurality of state transition phases, generate a local steady-state feature set corresponding to each load state and a local transition feature set corresponding to any adjacent two load states, and the local transition feature set includes local transition parameter record data corresponding to a plurality of state transition events.

[0032] The steady-state anomaly analysis module is configured to perform sliding window analysis on each local steady-state feature set based on state collaborative characteristics and quantile difference characteristics respectively, extract a state collaborative matrix and a quantile difference matrix of each local steady-state feature set, and construct a first fault detection strategy for steady-state anomaly detection of the chip.

[0033] The state transition analysis module is configured to perform fluctuation feature analysis based on load change on each local transition feature set, generate a plurality of fluctuation feature matrices of each local transition feature set, perform fluctuation co-occurrence analysis on a plurality of groups of local transition parameter record data in the local transition feature set respectively, generate a fluctuation co-occurrence matrix of each group of local transition parameter record data, and construct a second fault detection strategy for state transition anomaly detection of the chip according to the plurality of fluctuation feature matrices and the fluctuation co-occurrence matrices.

[0034] The fault detection module is configured to perform real-time fault detection on the chip by using the first fault detection strategy and the second fault detection strategy.

[0035] The present application has the following advantages:

[0036] The application can capture the gradual degradation of the synergistic relationship between different state parameters in the chip under stable load, and identify the burst characteristics under transient abnormality in the state switching process, by combining steady-state synergistic analysis with fluctuation detection, and can construct independent detection baselines for different load states, has strong dynamic load adaptability, and the fluctuation coexistence analysis of the state conversion process can reveal the abnormal coupling between parameters, explicit the implicit fault propagation path, and realize high-precision detection of small faults in dynamic scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A flowchart of a chip fault intelligent detection method based on running data provided by an embodiment of the application.

[0038] Figure 2 A structural diagram of a chip fault intelligent detection system based on running data provided by an embodiment of the application. DETAILED DESCRIPTION

[0039] In order for those skilled in the art to better understand the technical solutions in the application, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0040] Please refer to Figure 1 As one of the implementation processes of the application, a chip fault intelligent detection method based on running data is provided, comprising the following steps:

[0041] Step S10, obtain historical running state record data of the chip, determine a plurality of load states based on the running load of the chip, pre-divide the historical running state record data, and construct state feature data sets corresponding to each load state, and determine the state transition stage between adjacent two load states from the historical running state record data.

[0042] Specifically, the historical running state record data of the chip includes various state parameters recorded and collected in the actual running process of the chip in a historical period, such as load, voltage, temperature, power consumption, current, etc., for reflecting the historical running state of the chip. According to the difference of the use scene, for example, in the application of industrial Internet of Things, automotive electronic system, sensor processing unit, the running data of the chip can be periodically collected and analyzed according to the actual detection precision requirement. By analyzing the actual power consumption of the chip in different time periods and the basic load performance parameters set after the production of the product, the load range in which the chip can run can be divided to determine multiple load states. For example, the load range in which the chip can run is divided into multiple load states with a pre-set load interval, for example, 5% or 10%, and the person skilled in the art can also determine some typical load modes that may exist in the running process of the chip according to the specific use scene of the chip, reasonably set the load state of the chip, and ensure that it is suitable for the application scene of the chip.

[0043] After determining the multiple load states of the chip, the historical data corresponding to each load state is extracted from the historical running state record data to construct a state feature data set corresponding to each load state. At the same time, for the historical running state record data, the state transition stage between adjacent two load states can be determined according to these load states, for example, each 10% load level is regarded as a load state, and the process in which the load level of the chip changes from 19% to 21% can be regarded as a state transition stage between adjacent two load states, which is used for subsequent analysis of the conversion analysis in the state transition stage.

[0044] Step S20: According to the multiple state transition stages, the state conversion segmentation is performed on the multiple state feature data sets to generate a local steady state feature set corresponding to each load state and a local conversion feature set corresponding to any adjacent two load states, and the local conversion feature set includes local conversion parameter record data corresponding to multiple state conversion events.

[0045] Specifically, after determining the multiple state transition stages of the chip, the state transition segmentation is performed on the state feature dataset based on the multiple state transition stages, that is, the data in each state feature dataset at a relatively stable level of the load state is separated from the data at a relatively dramatic change level of the transition process of the load level in different use scenarios. For example, in the Internet of Things scenario, a certain sensor device may remain in several typical working modes, and the data collection frequency will be adjusted according to the actual scenario. Without a sudden event, it may work stably in some load level modes that are relatively low, and then the working mode may be changed after detecting the environment or device exception, resulting in an increase in the load level of the control chip. After considering the transition of the load state, the state transition segmentation is performed on each state feature dataset to generate a local steady-state feature set corresponding to each load state and a local transition feature set corresponding to any two adjacent load states. The local steady-state feature set includes the running data of the chip in some relatively stable load states. The local transition feature set includes the state features of the chip in the load switching process, which includes the local transition parameter record data of multiple state transition events. Each state transition event specifically refers to a transition process of the load state of the chip.

[0046] In step S30, a sliding window analysis based on state coordination characteristics and quantile difference characteristics is performed on each local steady-state feature set respectively to extract a state coordination matrix and a quantile difference matrix of each local steady-state feature set, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed.

[0047] Specifically, for each constructed local steady-state feature set, a sliding window analysis is performed from the perspective of the coordination change characteristics of different state parameters to extract the coordination change characteristics between the state parameters to identify different performances of the state parameters of the chip in the stable state. The change characteristics of different state parameters caused by the local changes of the chip in the relatively stable working state of the load are also analyzed.

[0048] In an optional implementation process, the sliding window analysis based on state coordination characteristics and quantile difference characteristics is performed on each local steady-state feature set respectively, including the following steps:

[0049] The local steady-state feature set is traversed through a preset state analysis window, the state feature sequence corresponding to each feature window under the preset state analysis window is extracted, the covariance between the state feature sequences corresponding to any two state parameters is calculated, the state coordination vector in each feature window is generated, and the state coordination vectors corresponding to the multiple feature windows are spliced in time sequence to generate the state coordination matrix of the local steady-state feature set.

[0050] Specifically, a state analysis window is preset, such as a time window of 1 minute, 5 minutes, etc. selected according to experience, to ensure that the running period of the chip is fully covered, and then the local steady-state feature set is traversed based on the preset state analysis window. For the data in each feature window under the preset state analysis window, the state feature sequence corresponding to each state parameter in the data is extracted, such as the sequence of the temperature, current, etc. of the chip changing with time in this period of time. For the multiple state feature sequences extracted in the window, the covariance between any two state parameters is calculated, which is used to measure whether the changes of the two state parameters are synchronized, and the greater the value is, the stronger the synchronous change relationship is. The state coordination vector corresponding to this window is constructed according to the covariance between any two state parameters in each feature window. The state coordination vectors corresponding to multiple feature windows sorted by time can be spliced in the time order of the traversal process to obtain the state coordination matrix of the local steady-state feature set, which is used to represent how multiple parameters change coordinately when the chip is in a stable state.

[0051] The multiple load quantile parameters of the local steady-state feature set are determined according to the load state to which the local steady-state feature set belongs, and the state quantile interval of each local steady-state feature set is constructed according to the multiple load quantile parameters.

[0052] Specifically, the load quantile parameter distinguishes the state performance of the chip under the local change of the load level in the relatively stable load state, to further subdivide the parameter characteristics in the load state. For example, for the load state interval corresponding to 10% to 20%, multiple quantile points are determined according to a local load level change of 2%, so as to determine the specific load quantile parameters corresponding to each quantile point in the load state interval, such as 12%, 14%, etc., so that the local steady-state feature set can be divided into multiple state quantile intervals.

[0053] During the process of traversing the local steady-state feature set by the preset state analysis window, the multiple mean feature vectors corresponding to multiple state parameters in each state quantile interval are generated, the local correlation coefficient between the mean feature vectors corresponding to any two state parameters in each state quantile interval is calculated, and the quantile correlation vector of each state quantile interval is constructed.

[0054] Specifically, for each state quantile interval, the mean values of different state parameters under different feature windows in the interval are calculated, and then the mean feature vectors of each state parameter under the state quantile interval are constructed according to the mean values of each state parameter under different feature windows, representing the average level of these parameters under the load state. The local correlation coefficient, such as the Pearson correlation coefficient, between any two mean feature vectors in each state quantile interval is further calculated, which is used to represent the correlation between two parameters under a specific load interval. The quantile correlation vector of each state quantile interval is generated to represent the correlation level between each state parameter and other state parameters under a specific load interval. By splicing the quantile correlation vectors of all quantile intervals, the quantile difference matrix of the local steady-state feature set is finally generated, which is used to reveal the influence of the local slight change of the load level on different state parameters under a specific working mode of the chip and describe the change rule of the correlation between different state parameters. For example, when the load of the chip in a low-power working mode changes slightly from low to high, the correlation between the current and the temperature may change, and this implicit characteristic can be captured by the quantile difference matrix.

[0055] By performing a sliding window analysis on the local steady-state feature set based on the state collaborative characteristics and the quantile difference characteristics, a state collaborative matrix and a quantile difference matrix of the local steady-state feature set can be constructed. The state collaborative matrix reveals the change direction and intensity between two variables from the time dimension, and the quantile difference matrix reflects the correlation change characteristics between parameters under the local slight change of the load state, for example, to understand whether the dependency between different parameters changes with the change of the load.

[0056] In an optional implementation process, for the construction of the first fault detection strategy for steady-state anomaly detection of the chip, the following contents are included:

[0057] According to the quantile difference matrix of the local steady-state feature set, the state parameter importance analysis of each state quantile interval is performed, including calculating the global correlation parameters of each state parameter under the state quantile interval according to the quantile correlation vector of the state quantile interval, and generating a collaborative reference list for each state quantile interval.

[0058] Specifically, the quantile difference matrix analyzes the characteristic changes between different state parameters from the load dimension. For each state quantile interval, the multiple local correlation coefficients involved in the quantile correlation vector of the state quantile interval can be used to calculate the multiple local correlation coefficients associated with each state parameter as the global correlation parameter of the state parameter, which is used to measure the correlation degree between the state parameter and the overall correlation of the remaining state parameters within the local load level range. By sorting the multiple state parameters according to the global correlation parameter, a collaborative reference list of the state quantile interval is generated. The state parameter with a larger global correlation parameter indicates that the overall correlation between the state parameter and other state parameters is higher at a specific load level. In the process of steady-state anomaly detection of the chip, the collaborative state analysis result involving the state parameter is more referential.

[0059] According to the state synergy matrix of the local steady-state feature set, multiple global synergy scores of each state parameter are calculated. According to the multiple global synergy scores, a synergy evaluation range of each state parameter is generated. Through the synergy evaluation range of the multiple state parameters and the synergy reference list of the multiple state quantile intervals, a first fault detection strategy for steady-state anomaly detection of the chip is generated.

[0060] Specifically, the state synergy matrix of the local steady-state feature set analyzes the characteristic changes between different state parameters from the time dimension. According to the time range involved in the local steady-state feature set, multiple reference periods for data analysis can be determined, such as every 10 preset state analysis windows as a reference period. For the content involved in the state synergy matrix in each reference period, the local synergy mean between any two state parameters can be calculated, that is, the mean of multiple covariances of the two state parameters in the state synergy matrix. The multiple local synergy means involved in each state parameter in the reference period are calculated as the global synergy score of the state parameter in the reference period. Then, the range of the multiple global synergy scores in different reference periods is analyzed. The synergy evaluation range of each state parameter can be obtained, which is used as the fluctuation range of the synergy change reference of each state parameter under different load conditions, to help determine how a certain parameter should change in coordination with other parameters under normal circumstances. Combined with the relative importance of different state parameters involved in the synergy reference list, a first fault detection strategy for steady-state anomaly detection of the chip is obtained, which realizes early mining of possible minor abnormalities in the running process when the chip is in a relatively stable load state, and realizes early warning of faults.

[0061] Step S40, respectively performing load change based fluctuation feature analysis on each local conversion feature set to generate a plurality of fluctuation feature matrices of each local conversion feature set, respectively performing fluctuation symbiotic analysis on a plurality of groups of local conversion parameter record data in the local conversion feature set to generate a fluctuation symbiotic matrix of each group of local conversion parameter record data.

[0062] Specifically, the purpose of the load change based fluctuation feature analysis process is to analyze the fluctuation features of each state parameter under different load conversion rates, that is, the cooperative change law shown by the change amplitude of each state parameter under different conversion rates when the chip needs to convert the working mode. At the same time, for each specific state conversion event, the overlapping of the fluctuations of different state parameters is analyzed, for example, for the rapid load state conversion caused by the rapid switching of the chip, the current may jump abnormally at multiple time points in this process. By analyzing the time synchronization of different state parameters about abnormal jumps, the symbiotic characteristics between states in different load state conversion processes are revealed. For example, load change usually causes current to rise, but if it is found that the symbiotic relationship between current fluctuation and temperature change is abnormally changed in some cases, it may be a signal of the decline of the chip heat dissipation capability.

[0063] In an optional implementation process, the generation of the plurality of fluctuation feature matrices and the fluctuation symbiotic matrix of the local conversion feature set specifically includes:

[0064] The fluctuation amplitude sequence of each state parameter in the local conversion feature set about the load change is extracted, a data fluctuation splitting strategy is determined according to the load conversion rate involved in the local conversion feature set, and a plurality of conversion rate intervals of the local conversion feature set are generated according to the data fluctuation splitting strategy.

[0065] Specifically, the fluctuation amplitude sequence of the state parameter about the load change is specifically the overall fluctuation amplitude of the electrical parameter under the load conversion rate shown for different state conversion events in the historical data of the chip. Considering that the conversion rate level of the load corresponding to different state conversion events is different, a global range interval is determined according to the plurality of load conversion rates involved in the local conversion feature set, and then a data fluctuation splitting strategy for local analysis of the range interval is determined, for example, the range interval is divided into a plurality of equally spaced small intervals, so as to determine the data fluctuation splitting strategy for splitting different state conversion events based on the small intervals, that is, to determine a plurality of conversion rate intervals of the local conversion feature set, each conversion rate interval represents a certain rate range, for example, rapid change, smooth change, slow change, etc., realizing the classification of different state conversion events according to the fast and slow level of conversion.

[0066] The fluctuation amplitude sequence of the state parameters with respect to the load change is reconstructed based on timing according to the conversion rate interval, a local discrete characteristic vector of each conversion rate interval is generated under each fluctuation amplitude sequence, and a fluctuation characteristic matrix corresponding to each conversion rate interval in the local conversion feature set is constructed.

[0067] Specifically, the timing-based feature reconstruction specifically distributes different state conversion events according to the conversion rate interval to which the load conversion rate belongs, to obtain a local discrete characteristic vector corresponding to each conversion rate interval. After determining the local discrete characteristic vector of each state parameter in the conversion rate interval in this way, the local discrete characteristic vectors of multiple state parameters are combined to obtain a fluctuation characteristic matrix corresponding to each conversion rate interval in the local conversion feature set, i.e., the difference in fluctuation level of different state parameters under a specific conversion rate level.

[0068] And the fluctuation co-occurrence matrix of each group of local conversion parameter record data in each local conversion feature set is generated by combining the data fluctuation splitting strategy and performing fluctuation co-occurrence analysis on each group of local conversion parameter record data.

[0069] Specifically, according to the aforementioned data fluctuation splitting strategy, the multiple groups of local conversion parameter record data in the local conversion feature set are distributed to the corresponding conversion rate interval, and then for each group of local conversion parameter record data, the fluctuation co-occurrence sequence of different state parameters is extracted. In this process, for each state parameter, the local conversion parameter record data of the state parameter is statistically analyzed to calculate the mean and standard deviation, etc. The fluctuation threshold of the state parameter is determined according to the mean and standard deviation, for example, the node higher than the mean plus 2 times the standard deviation is taken as the fluctuation threshold. Then the local conversion parameter record data of the state parameter is analyzed by sliding window to determine multiple time nodes of the state parameter greater than the fluctuation threshold, i.e., a specific sliding window, and the fluctuation co-occurrence sequence of the state parameter is generated. For example, if the mean of the state parameter in a window is greater than the fluctuation threshold, the element value corresponding to the window in the fluctuation co-occurrence sequence is recorded as 1, otherwise as 0. In this way, multiple fluctuation co-occurrence sequences of each group of local conversion parameter record data under each conversion rate interval to which it belongs are extracted, and a fluctuation co-occurrence matrix of the local conversion parameter record data containing the fluctuation co-occurrence sequences corresponding to multiple state parameters is constructed, which reflects the coordinated change trend of each parameter under a specific load conversion rate. By comparing the fluctuation co-occurrence matrices of different intervals, it can be observed how each electrical parameter of the chip changes coordinately during the rapid, smooth or slow change of the load, thereby revealing possible failure modes or abnormal behaviors.

[0070] Step S50, constructing a second fault detection strategy for state transition anomaly detection of the chip according to the plurality of fluctuation feature matrices and the fluctuation co-occurrence matrix, and realizing real-time fault detection of the chip through the first fault detection strategy and the second fault detection strategy.

[0071] Specifically, the second fault strategy is used to detect possible micro-abnormalities in the chip mode transition process. The fluctuation feature matrix quantitatively analyzes the fluctuation characteristics of different parameters following the load change under normal circumstances. The fluctuation co-occurrence matrix reveals the influence of different state parameters on other parameters after the fluctuation. Based on this information, a second fault detection strategy for state transition anomaly detection of the chip is generated.

[0072] In an alternative implementation process, the construction of the second fault detection strategy for state transition anomaly detection of the chip includes the following contents:

[0073] According to the local discrete characteristic vector, the state parameter importance of each transition rate interval is analyzed, including calculating the global fluctuation characteristic parameters of each state parameter in the transition rate interval according to the local discrete characteristic vector, and generating a fluctuation reference list for each transition rate interval.

[0074] Specifically, the local discrete characteristic vector reflects the change of the fluctuation degree of the state parameter under a specific load transition rate as time evolves. As the chip gradually wears out, the performance of the chip will decline, which is manifested as in a dynamic scenario, different state parameters may not respond to the rapidly changing load level in time, or in other words, there will be a short-term abnormal fluctuation and a short-term recovery. To some extent, the local discrete characteristic vector can also be understood as the resistance level of the chip to mode transition in a dynamic scenario, that is, whether it can recover to the normal level in time or try not to appear high-amplitude state deviation under the rapidly changing load level.

[0075] By calculating the local discrete characteristic vector, the global fluctuation characteristic parameters representing the related fluctuation characteristic level of each state parameter under a specific load transition rate level can be obtained. For example, the mean value of the plurality of fluctuation amplitude parameters of the state parameter in the local discrete characteristic vector is calculated as the global fluctuation characteristic parameter of the state parameter. Based on the global fluctuation characteristic parameter, the plurality of state parameters are sorted to generate a fluctuation reference list for each transition rate interval, which is used to represent the resistance of different state parameters to abnormal states under a specific scenario. The smaller the global fluctuation characteristic parameter is, the stronger the adaptability to load state transition is, and there will be no large-amplitude level deviation. In the process of analyzing the coordination characteristics between different state parameters, the data corresponding to the more stable state parameters is more referential.

[0076] The global co-occurrence scores of each state parameter are calculated according to the plurality of fluctuation co-occurrence matrices, and the fluctuation evaluation range of each state parameter is generated according to the plurality of global co-occurrence scores. The second fault detection strategy for state transition anomaly detection of the chip is generated by the fluctuation evaluation range of the plurality of state parameters and the fluctuation reference list of the plurality of conversion rate intervals.

[0077] Specifically, the fluctuation co-occurrence matrix includes the fluctuation coordination characteristics of the plurality of state parameters under different scenarios caused by the load state conversion of the chip. The plurality of fluctuation co-occurrence matrices are divided based on a plurality of reference periods. Each reference period can include one or more fluctuation co-occurrence matrices. For any two state parameters, the local co-occurrence parameters between the two fluctuation co-occurrence sequences corresponding to the two state parameters in the fluctuation co-occurrence matrix are calculated according to the two fluctuation co-occurrence sequences. Specifically, it is analyzed whether the fluctuation characteristics of the two parameters in each window are the same, for example, both are 1, which indicates that the fluctuation characteristics of the two state parameters in the window are the same. The coverage of the window between any two state parameters in the fluctuation co-occurrence matrix with respect to the fluctuation characteristics compared to all windows is calculated as the local co-occurrence parameter between any two state parameters in the fluctuation co-occurrence matrix. The mean value of the plurality of local co-occurrence parameters of the state parameter in each fluctuation co-occurrence matrix under the reference period is obtained as the global co-occurrence score of the state parameter under the reference period. Thus, according to the global co-occurrence scores corresponding to the plurality of reference periods, the range involved is determined to generate the fluctuation evaluation range of each state parameter, which is used as a reference for the evaluation of the state parameter under different load state conversion scenarios. The state parameters appear significantly higher than the normal level of fluctuation phenomenon. In combination with the relative importance of the state parameter recorded in the fluctuation reference list during the load state conversion process, the second fault detection strategy for state transition anomaly detection of the chip is obtained. The stability of the chip in complex scenarios is realized by mining the abnormal fluctuation characteristics that may exist in the data. If an abnormal fluctuation of a certain state parameter occurs, i.e., the co-occurrence state is abnormal, it indicates that the state parameter may cause an abnormal fluctuation of the state due to a small fault of the chip. Further, the early warning of the fault is realized.

[0078] In an optional implementation process, real-time fault detection of the chip is realized by the first fault detection strategy and the second fault detection strategy, specifically including:

[0079] After obtaining the target running monitoring data of the chip, the target load state and the target state transition phase involved in the target running monitoring data are determined. Specifically, the load change data of the chip in the target running monitoring data can be analyzed to determine the load state or state transition phase that may be involved.

[0080] The target steady-state data of the target load state is extracted, a state quantile interval corresponding to the target steady-state data is determined, steady-state evaluation weights of the plurality of state parameters under the target steady-state data are determined according to a collaborative reference list of the state quantile interval, wherein the steady-state evaluation weights can be normalized to generate according to global correlation parameters of the plurality of state parameters in the collaborative reference list. Then, state collaborative analysis is performed on the plurality of state parameters under the target steady-state data, target collaborative parameters between any two state parameters under the target steady-state data are calculated, first collaborative scores of each state parameter are calculated through the plurality of target collaborative parameters, which are used to quantitatively represent the collaborative characteristics between the current state parameter and the remaining state parameters in the current period. At the same time, considering that some parameters are more important, the steady-state evaluation weights are further used to modify them. For example, a state parameter has high collaboration with the remaining state parameters in the historical state, but the collaboration under the target steady-state data decreases. In the same degree of decrease, the situation represented by the state parameter with high collaboration is more serious and needs more attention. After the first collaborative scores are modified to generate second collaborative scores based on the steady-state evaluation weights of the plurality of state parameters, state collaborative evaluation is performed on the target steady-state data based on a collaborative evaluation range of the plurality of state parameters under the target load state, and a steady-state anomaly detection result of the chip is generated, which indicates whether there is an anomaly in different state parameters of the chip under the current state, and realizes detection and early warning of possible potential micro anomalies of the chip under a relatively stable running state.

[0081] Similarly, after the target transition data of the target load state is extracted, a transition rate interval corresponding to the target transition data is determined, and fluctuation evaluation weights of the plurality of state parameters under the target transition data are generated according to a fluctuation reference list of the transition rate interval. The significance is to preferentially pay attention to the parameters with strong symbiotic relationship with the remaining state parameters. In this case, even if the remaining state parameters have an anomaly, the symbiotic relationship between them will also be abnormal, realizing multi-angle anomaly detection of each state parameter. Fluctuation characteristic analysis is performed on the plurality of state parameters under the target transition data, target symbiotic parameters between any two state parameters are calculated, first symbiotic scores of each state parameter are calculated through the plurality of target symbiotic parameters, second symbiotic scores are generated by modifying the first symbiotic scores based on the fluctuation evaluation weights of the state parameters, and state anomaly transition detection results of the chip are generated by performing fluctuation characteristic evaluation on the target transition data based on a fluctuation evaluation range of the plurality of state parameters under the target transition data.

[0082] The application can capture the progressive degradation of the synergistic relationship between different state parameters in the chip under stable load, and identify the sudden characteristics under transient abnormality in the state switching process, and construct independent detection baselines for different load states, which has strong dynamic load adaptability, and the fluctuation coexistence analysis of the state conversion process can reveal the abnormal coupling between parameters, which can make the implicit fault propagation path explicit, and can be used for high-precision fault detection of the chip in dynamic use scenarios. For different application scenarios, such as sensor networks in the production field and automotive electronics, the data analysis period can be reasonably set according to actual needs, so as to timely grasp the actual state of the chip, and the detection standard can be reasonably adjusted according to the recent performance of the chip in the fault detection process, to realize high-precision detection of small faults in dynamic scenarios.

[0083] Please refer to Figure 2 As one of the implementation processes of the application, based on the above-mentioned application concept, a chip fault intelligent detection system based on operation data is also provided, comprising:

[0084] A historical data preprocessing module is configured to obtain historical operation state record data of the chip, determine a plurality of load states based on the operation load of the chip, pre-divide the historical operation state record data, and construct state feature data sets corresponding to each load state, and determine state transition stages between adjacent two load states from the historical operation state record data.

[0085] A load state analysis module is configured to divide the plurality of state feature data sets according to the plurality of state transition stages, generate local steady-state feature sets corresponding to each load state, and local conversion feature sets corresponding to any adjacent two load states, and the local conversion feature sets include local conversion parameter record data corresponding to a plurality of state conversion events.

[0086] A steady-state abnormality analysis module is configured to perform sliding window analysis on each local steady-state feature set based on state synergistic characteristics and quantile difference characteristics, extract state synergy matrices and quantile difference matrices of each local steady-state feature set, and construct a first fault detection strategy for steady-state abnormality detection of the chip.

[0087] A state conversion analysis module is configured to perform fluctuation feature analysis on each local conversion feature set based on load changes, generate a plurality of fluctuation feature matrices of each local conversion feature set, and perform fluctuation coexistence analysis on a plurality of groups of local conversion parameter record data in the local conversion feature set, generate a fluctuation coexistence matrix of each group of local conversion parameter record data, and construct a second fault detection strategy for state conversion abnormality detection of the chip according to the plurality of fluctuation feature matrices and the fluctuation coexistence matrices.

[0088] The fault detection module is configured to implement real-time fault detection on the chip by using the first fault detection strategy and the second fault detection strategy.

[0089] The above merely describes specific embodiments of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art. The parts not described in detail in the specification belong to the prior art known to those skilled in the art.

Claims

1. A chip fault intelligent detection method based on operational data, characterized in that, include: The historical operating status records of the chip are obtained, multiple load states are determined based on the chip's operating load, the historical operating status records are pre-divided, and a state feature dataset corresponding to each load state is constructed. The state transition stage between two adjacent load states is determined from the historical operating status records. Based on multiple state transition stages, state transition segmentation is performed on multiple state feature datasets to generate a local steady-state feature set corresponding to each load state, as well as a local transition feature set corresponding to any two adjacent load states. The local transition feature set includes local transition parameter record data corresponding to multiple state transition events. For each local steady-state feature set, a sliding window analysis based on state coordination characteristics and quantile difference characteristics is performed to extract the state coordination matrix and quantile difference matrix of each local steady-state feature set, and a first fault detection strategy for steady-state anomaly detection of the chip is constructed. For each local transformation feature set, a fluctuation feature analysis based on load change is performed to generate multiple fluctuation feature matrices for each local transformation feature set. For multiple sets of local transformation parameter record data in the local transformation feature set, a fluctuation co-occurrence analysis is performed to generate a fluctuation co-occurrence matrix for each set of local transformation parameter record data. A second fault detection strategy for detecting state transition anomalies in chips is constructed based on multiple fluctuation feature matrices and fluctuation co-occurrence matrices. Real-time fault detection of chips is achieved through the first and second fault detection strategies. Among them, the first fault detection strategy and the second fault detection strategy include: Based on the quantile difference matrix of the local steady-state feature set, the importance of state parameters is analyzed for each state quantile interval. This includes calculating the global correlation parameter of each state parameter in the state quantile interval based on the quantile correlation vector of the state quantile interval, generating a cooperative reference list for each state quantile interval, calculating multiple global cooperative scores for each state parameter based on the state cooperative matrix of the local steady-state feature set, generating a cooperative evaluation range for each state parameter based on the multiple global cooperative scores, and generating a first fault detection strategy for steady-state anomaly detection of the chip through the cooperative evaluation range of multiple state parameters and the cooperative reference list of multiple state quantile intervals. The importance analysis of state parameters is performed on each switching rate interval based on the local discrete characteristic vector. This includes calculating the global fluctuation characteristic parameter of each state parameter in the switching rate interval based on the local discrete characteristic vector, generating a fluctuation reference list for each switching rate interval, calculating multiple global co-occurrence scores for each state parameter based on multiple fluctuation co-occurrence matrices, generating a fluctuation evaluation range for each state parameter based on multiple global co-occurrence scores, and generating a second fault detection strategy for detecting state transition anomalies in the chip using the fluctuation evaluation ranges of multiple state parameters and the fluctuation reference lists of multiple switching rate intervals.

2. The intelligent chip fault detection method based on operational data according to claim 1, characterized in that, For each local steady-state feature set, a sliding window analysis based on state coordination characteristics and quantile difference characteristics is performed to extract the state coordination matrix and quantile difference matrix for each local steady-state feature set, including: By traversing the local steady-state feature set through a preset state analysis window, the state feature sequences corresponding to multiple state parameters in each feature window under the preset state analysis window are extracted. The covariance between the state feature sequences corresponding to any two state parameters is calculated, and the state coordination vector in each feature window is generated. The state coordination vectors corresponding to multiple feature windows are concatenated according to the time sequence to generate the state coordination matrix of the local steady-state feature set. Based on the load state to which the local steady-state feature set belongs, determine multiple load quantile parameters of the local steady-state feature set, and construct the state quantile interval of each local steady-state feature set based on the multiple load quantile parameters; During the process of traversing the local steady-state feature set through the preset state analysis window, the mean feature vectors corresponding to multiple state parameters under each state quantile interval are generated. The local correlation coefficient between the mean feature vectors corresponding to any two state parameters under each state quantile interval is calculated, and the quantile correlation vector of each state quantile interval is constructed. The state coordination vectors corresponding to multiple state quantile intervals are concatenated to generate the quantile difference matrix of the local steady-state feature set.

3. The intelligent chip fault detection method based on operational data according to claim 2, characterized in that, The generation of multiple fluctuation feature matrices and fluctuation co-occurrence matrices for the local transformation feature set includes: Extract the fluctuation amplitude sequences of multiple state parameters with respect to load changes from the local transformation feature set. Determine the data fluctuation splitting strategy based on the load transformation rate involved in the local transformation feature set. Generate multiple transformation rate intervals of the local transformation feature set based on the data fluctuation splitting strategy. Perform time-series-based feature reconstruction on the fluctuation amplitude sequences of state parameters with respect to load changes according to the transformation rate intervals. Generate local discrete characteristic vectors for each fluctuation amplitude sequence with respect to each transformation rate interval. Construct the fluctuation feature matrix corresponding to each transformation rate interval in the local transformation feature set. By combining the data fluctuation decomposition strategy, fluctuation co-occurrence analysis is performed on multiple sets of local transformation parameter records in each local transformation feature set. This includes assigning each set of local transformation parameter records to multiple transformation rate intervals, extracting multiple fluctuation co-occurrence sequences of each set of local transformation parameter records in each corresponding transformation rate interval, and generating a fluctuation co-occurrence matrix for each set of local transformation parameter records.

4. The intelligent chip fault detection method based on operational data according to claim 3, characterized in that, The calculation of global collaboration score and global symbiosis score includes: The state coordination matrix is ​​divided based on multiple reference time periods. The local coordination mean between any two state parameters in each reference time period is calculated. The global coordination score of the state parameters in the reference time period is calculated based on the multiple local coordination means of each state parameter in the reference time period. Multiple fluctuation co-occurrence matrices are divided based on multiple reference time periods. The local co-occurrence parameters of any two state parameters in each fluctuation co-occurrence matrix under each reference time period are calculated. The global co-occurrence score of the state parameter under the reference time period is calculated based on the multiple local co-occurrence parameters of each state parameter under the reference time period.

5. The intelligent chip fault detection method based on operational data according to claim 4, characterized in that, Real-time fault detection of the chip is achieved through a first fault detection strategy and a second fault detection strategy, including: After acquiring the target operation monitoring data of the chip, determine the target load state and target state transition stage involved in the target operation monitoring data; Extract the target steady-state data of the target load state, determine the state quantile interval corresponding to the target steady-state data, determine the steady-state evaluation weights of multiple state parameters under the target steady-state data based on the collaborative reference list of the state quantile intervals, perform state collaborative analysis on multiple state parameters under the target steady-state data, calculate the target collaborative parameters between any two state parameters under the target steady-state data, calculate the first collaborative score of each state parameter through multiple target collaborative parameters, correct the first collaborative score based on the steady-state evaluation weights of the state parameters to generate the second collaborative score, perform state collaborative evaluation on the target steady-state data based on the collaborative evaluation range of multiple state parameters under the target load state, and generate the steady-state anomaly detection results of the chip; Extract the target transition data of the target load state, determine the transition rate range corresponding to the target transition data, generate the fluctuation evaluation weights of multiple state parameters under the target transition data based on the fluctuation reference list of the transition rate range, perform fluctuation characteristic analysis on multiple state parameters under the target transition data, calculate the target co-occurrence parameters between any two state parameters, calculate the first co-occurrence score of each state parameter through multiple target co-occurrence parameters, correct the first co-occurrence score based on the fluctuation evaluation weight of the state parameter to generate the second co-occurrence score, evaluate the fluctuation characteristics of the target transition data based on the fluctuation evaluation range of multiple state parameters under the target transition data, and generate the chip's state abnormal transition detection results.

6. The intelligent chip fault detection method based on operational data according to claim 5, characterized in that, Extract multiple fluctuation co-occurrence sequences from each set of local transformation parameter records within each corresponding transformation rate interval, including: Statistical analysis is performed on the local transformation parameter recording data to generate multiple statistical features for each state parameter. Based on these statistical features, the fluctuation threshold for each state parameter is determined. Sliding window detection based on fluctuation amplitude is then performed on the local transformation parameter recording data to determine multiple time points where each state parameter exceeds the fluctuation threshold, thus generating a fluctuation co-occurrence sequence for each state parameter.

7. A chip fault intelligent detection system based on operational data, characterized in that, The system is used to implement the intelligent chip fault detection method based on operational data as described in any one of claims 1-6, comprising: The historical data preprocessing module is used to acquire historical operating status record data of the chip, determine multiple load states based on the chip's operating load, pre-divide the historical operating status record data, construct a state feature dataset corresponding to each load state, and determine the state transition stage between two adjacent load states from the historical operating status record data. The load state analysis module is used to perform state transition segmentation on multiple state feature datasets based on multiple state transition stages, generate a local steady-state feature set corresponding to each load state, and a local transition feature set corresponding to any two adjacent load states. The local transition feature set includes local transition parameter record data corresponding to multiple state transition events. The steady-state anomaly analysis module is used to perform sliding window analysis based on state coordination characteristics and quantile difference characteristics on each local steady-state feature set, extract the state coordination matrix and quantile difference matrix of each local steady-state feature set, and construct the first fault detection strategy for steady-state anomaly detection of the chip. The state transition analysis module is used to perform fluctuation feature analysis based on load change for each local transition feature set, generate multiple fluctuation feature matrices for each local transition feature set, perform fluctuation co-occurrence analysis on multiple sets of local transition parameter record data in the local transition feature set, generate fluctuation co-occurrence matrix for each set of local transition parameter record data, and construct a second fault detection strategy for detecting state transition anomalies of the chip based on multiple fluctuation feature matrices and fluctuation co-occurrence matrices. The fault detection module is used to achieve real-time fault detection of the chip through a first fault detection strategy and a second fault detection strategy.

Images