DDR wafer test anomaly detection method and device based on machine learning

Abstract

The invention provides a DDR wafer test anomaly detection method and device based on machine learning, and the method comprises the steps: obtaining a time sequence eye pattern synchronously collected in a DDR wafer test process and time sequence waveform data related to total jitter, deterministic jitter and random jitter, and carrying out the clock edge phase segmentation alignment of differential time sequence sampling points in the time sequence waveform data, phase domain dense sampling data is constructed, adaptive sparse coding is completed in combination with DDR jitter physical priori, noise and normal process fluctuation related components are removed, defect sensitive sparse components are obtained, variational mode decomposition is performed, components related to process drift, normal process fluctuation and abrupt change abnormity are obtained, and the defect sensitive sparse components are subjected to sparse coding; and the abrupt change anomaly correlation component is extracted as target detection data, a preset machine learning model is input, an anomaly mode correlation result is generated, a DDR wafer test anomaly detection result is output based on the result, and a corresponding test setting item adjustment instruction is generated. According to the invention, the overall efficiency and the result reliability of the DDR wafer test can be improved.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and data processing, and more specifically, to a method and device for detecting anomalies in DDR wafer testing based on machine learning. Background Technology

[0002] Anomaly detection in DDR wafer testing is a core step in ensuring product yield during the mass production of DDR chips. By analyzing various signal data generated during the testing process, potential anomalies in the wafer manufacturing or design stages can be identified, providing a basis for process optimization and product classification and screening.

[0003] Currently, the industry typically collects timing waveform data during DDR wafer testing. After preprocessing the data using common algorithms such as signal filtering and mode decomposition, it combines traditional machine learning models to identify anomalies. The identified anomalies are mainly used to mark problematic wafers. Some solutions provide broad suggestions for process adjustments based on the identification results. However, these solutions do not incorporate the specific signal fluctuation patterns of DDR during the data acquisition phase, which can easily introduce a large amount of invalid data with low correlation to anomalies. This significantly increases the load on subsequent data processing stages. The general algorithms used in the signal processing stage cannot accurately adapt to the signal characteristics of DDR, making it difficult to accurately distinguish between normal process fluctuations and real abnormal signals. This leads to frequent misjudgments and omissions in anomaly identification. Furthermore, the identification results lack compatibility with the calibration logic of DDR wafer testing equipment, making it impossible to quickly translate them into directly executable test setting adjustments. This makes it difficult to efficiently guide the real-time calibration of testing equipment, ultimately affecting the overall efficiency of DDR wafer testing and the accuracy of yield control. Summary of the Invention

[0004] This invention provides a method and device for detecting anomalies in DDR wafer testing based on machine learning.

[0005] In a first aspect, embodiments of the present invention provide a machine learning-based method for detecting anomalies in DDR wafer testing, comprising: acquiring timing eye diagrams and timing waveform data related to total jitter, deterministic jitter, and random jitter synchronously collected during DDR wafer testing, wherein the timing waveform data includes a sequence of continuously sampled differential signal sampling points with timestamps; performing clock edge phase segmentation alignment on the differential timing sampling points in the timing waveform data to construct phase domain dense sampling data, combining DDR jitter physical priors to complete adaptive sparse coding, removing noise and normal process fluctuation-related components to obtain defect-sensitive sparse components; performing variational mode decomposition on the defect-sensitive sparse components to separate components related to process drift, normal process fluctuations, and abrupt anomalies, and extracting abrupt anomaly-related components as target detection data; inputting the target detection data into a preset machine learning model, generating anomaly pattern association results through association modeling and anomaly pattern matching processing of the target detection data; outputting DDR wafer testing anomaly detection results based on the anomaly pattern association results, generating corresponding test setting adjustment instructions, and sending the test setting adjustment instructions to the DDR wafer testing equipment for test setting calibration operations.

[0006] In a second aspect, embodiments of the present invention provide a computer device, including: a memory storing a computer program; and a processor for loading the computer program to implement the machine learning-based DDR wafer test anomaly detection method as described above.

[0007] This invention precisely identifies core signal dimensions highly correlated with anomalies by synchronously acquiring timing eye diagrams and timestamped differential sampling point sequences related to three types of core jitter during DDR wafer testing, avoiding resource consumption from invalid data. It constructs dense sampling data in the phase domain by aligning clock edge phase segments, and combines this with DDR jitter physical priors to complete adaptive sparse coding, accurately filtering noise and normal process fluctuation components, focusing on key defect-related information, and improving subsequent detection efficiency. Variational mode decomposition is performed on defect-sensitive sparse components to accurately separate abrupt anomaly-related components as detection targets, avoiding misjudging normal process changes as anomalies and reducing the probability of misjudgment. The target data is input into a preset machine learning model, and correlation results are generated through association modeling and anomaly pattern matching to deeply explore internal data correlations and accurately identify the essence of anomalies. Based on the correlation results, detection results are output and equipment calibration instructions are generated, achieving efficient integration of anomaly detection and equipment adjustment, reducing manual intervention, and improving the overall efficiency and reliability of DDR wafer testing results. Attached Figure Description

[0008] Figure 1 This is a flowchart of a machine learning-based DDR wafer testing anomaly detection method provided in an embodiment of the present invention.

[0009] Figure 2This is a schematic diagram of the composition of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Please see Figure 1 This is a flowchart of a machine learning-based anomaly detection method for DDR wafer testing, provided by an embodiment of the present invention. The method can be executed by a computer device and may include the following steps: Step S100: Acquire the timing eye diagram and timing waveform data related to total jitter, deterministic jitter and random jitter synchronously collected during the DDR wafer testing process. The timing waveform data includes a sequence of continuously sampled differential signal sampling points with timestamps.

[0012] In DDR wafer testing scenarios, timing eye diagrams are a key visualization tool for evaluating signal quality. They are created by sampling the signal waveform multiple times over multiple bit periods and superimposing these sampled waveforms. During transmission, signals are affected by various factors such as noise, interference, and inter-symbol interference (ISI), causing waveform distortion. By observing the timing eye diagram, key information such as rising edges, falling edges, zero-crossing points, and noise margins can be visually observed, allowing for the assessment of signal distortion and timing characteristics. Total jitter is an indicator that measures the total deviation of a signal from its ideal position during transmission, representing a combination of deterministic and random jitter. Deterministic jitter is caused by predictable factors, including clock skew, ISI, and power supply ripple. Clock skew is the deviation between the actual and ideal phase of the clock signal, leading to inaccuracies in data signal sampling. ISI is the interference between adjacent data bits, causing waveform distortion. Power supply ripple is fluctuation in power supply voltage, affecting signal amplitude and phase. Random jitter is caused by unpredictable factors, mainly including thermal noise and shot noise. Thermal noise is generated by the thermal motion of electrons in a conductor, while shot noise is generated by the dispersion of charge.

[0013] To acquire this data, testing equipment such as an oscilloscope can be used. The oscilloscope samples the signal at a set sampling frequency, converting the signal voltage value into a digital signal and adding a timestamp to each sampling point. The sampling frequency is the number of times the oscilloscope samples the signal per second. The higher the sampling frequency, the richer the signal details are, but it also increases the amount of data and the processing difficulty.

[0014] Step S200: Align the differential timing sampling points in the timing waveform data by clock edge phase segmentation, construct dense sampling data in the phase domain, and complete adaptive sparse coding by combining DDR jitter physical priors to remove noise and normal process fluctuation related components, thereby obtaining defect-sensitive sparse components.

[0015] In an optional implementation, step S200 may specifically include the following steps S210 to S260: Step S210: Extract the phase nodes of the clock signal and the differential timing sampling points in the timing waveform data from the timing waveform data, trace the conduction dependency path of each differential timing sampling point in the clock phase advancement process, analyze its conduction correlation range with phase change, and generate a sampling point-phase conduction dependency correlation table, which includes the conduction start phase, dependency node range and conduction path characteristics of each sampling point.

[0016] In the timing waveform data of DDR wafer testing, the phase nodes of the clock signal are the positions of the clock signal at different phases. These nodes usually correspond to the rising or falling edge of the clock signal and are key reference points for signal synchronization and timing analysis. Differential timing sampling points are a series of data points obtained by sampling the differential signal. These points reflect the state of the differential signal at different times.

[0017] Tracing the conduction dependency path of each differential timing sampling point during clock phase progression involves analyzing the conduction relationships and dependencies between each sampling point and other nodes or signals as the clock signal phase changes. For example, a differential timing sampling point may depend on a signal state from the previous clock cycle or have some conduction correlation with adjacent sampling points. By tracing these conduction dependency paths, we can understand the propagation patterns and interactions of signals during clock phase progression.

[0018] Analyzing the conduction correlation range as it changes with phase determines the degree and extent of correlation between each sampling point and other nodes or signals at different phases. Different sampling points may have different conduction correlation ranges at different phases; some sampling points may only have a strong correlation with a few adjacent nodes, while others may have correlations with a wider range of nodes. By analyzing the conduction correlation range, a more comprehensive understanding of the characteristics and interrelationships of signals at different phases can be obtained.

[0019] Generating a sampling point-phase conduction dependency table involves organizing and recording the conduction start phase, dependent node range, and conduction path characteristics of each sampling point into a table. The conduction start phase is the clock phase at which the sampling point begins to establish a conduction relationship with other nodes or signals; the dependent node range is the range of nodes that have a conduction relationship with the sampling point; and the conduction path characteristics describe some features of the conduction path of the sampling point, such as conduction direction and conduction strength.

[0020] For example, a clock recovery circuit can be used to extract the phase nodes of the clock signal from timing waveform data. The clock recovery circuit can accurately recover the phase information of the clock signal by processing the signal through filtering, phase-locked loops, etc. For tracking and conducting correlation analysis of differential timing sampling points, correlation analysis algorithms can be used to determine the degree and range of conducting correlation between sampling points by calculating the correlation coefficient between them.

[0021] Step S220: Based on the sampling point-phase conduction dependency association table, connect the sampling points with the same source conduction start phase and continuous dependency path in the phase advancement order to form a phase conduction dependency link. Each link corresponds to a set of continuous clock phase conduction processes, and retain the conduction timing and dependency relationship of the sampling points in the link.

[0022] In an optional implementation, step S220 may specifically include the following steps S221 to S226: Step S221: Extract the conduction initiation phase and dependency path features of each sampling point from the sampling point-phase conduction dependency association table to obtain the conduction homology identification imprint and generate the phase conduction homology dependency imprint sequence, which contains the conduction homology identifier and dependency path feature description of each sampling point.

[0023] The conduction origin identification mark is a marker used to identify whether sampling points have the same conduction start phase and similar dependency path characteristics. Sampling points with the same conduction origin identification mark can be considered to have the same conduction origin, starting conduction at the same phase of the clock signal, and having similar conduction path characteristics.

[0024] Generating a phase-transduction homology dependency imprint sequence involves arranging the conduction homology identifier and dependency path feature description of each sampling point in a specific order to form a sequence. The conduction homology identifier is used to distinguish different conduction homology groups, and the sampling points within each group have the same conduction initiation phase; the dependency path feature description details the dependency path characteristics of each sampling point.

[0025] For example, database queries can be used to extract the required conduction initiation phase and dependency path feature data from a sampling point-phase conduction dependency association table. This data can then be processed and transformed programmatically to generate conduction homology identification imprints and phase conduction homology dependency imprint sequences.

[0026] Step S222: Perform homology clustering on the phase conduction homology clustering sequence, group the sampling point imprints with the same conduction initiation phase and similar dependency path characteristics into a group, and generate homology conduction dependency cluster groups, each group corresponding to a continuous phase conduction process.

[0027] By using homology dependency clustering, sampling point imprints can be divided into different groups, each group corresponding to a type of continuous phase propagation process. Such grouping helps to more clearly observe and analyze the characteristics and patterns of signals in different propagation processes. For example, in a test of DDR memory, it may be found that some sampling points begin propagation at a specific phase of the clock signal, and their dependency paths are all related to a certain circuit module. These sampling points can be grouped together, representing a phase propagation process.

[0028] Homology dependency clustering can utilize clustering algorithms such as K-means clustering and hierarchical clustering. Taking K-means clustering as an example, K centroids are randomly selected. Each sample imprint is then assigned to the group containing the nearest centroid. The centroids of each group are recalculated, and this process is repeated until the centroids no longer change or a preset number of iterations is reached. In DDR wafer testing, the distance between sample imprints can be calculated based on the similarity of conduction initiation phase and dependency path characteristics, grouping closely spaced sample imprints into a single group.

[0029] Step S223: Fit the dependency path of the sampling points in each homologous conduction dependency cluster to generate a conduction trajectory baseline sequence of the sampling points in the group. The conduction trajectory baseline sequence retains a unified description of the conduction timing and dependency relationship of all sampling points in the group.

[0030] In DDR wafer testing, fitting the dependency paths of sampling points within each homologous conduction dependency cluster is done to obtain a baseline sequence representing the conduction characteristics of that group of sampling points. Homologous conduction dependency clusters are obtained through the preceding clustering operation; sampling points within a group share the same conduction initiation phase and similar dependency path characteristics. Fitting the dependency paths of sampling points can be done using curve fitting algorithms, such as the least squares method. The least squares method is a commonly used curve fitting method; its basic idea is to determine the parameters of the fitted curve by minimizing the sum of squared errors between the actual data points and the fitted curve. In DDR wafer testing, the dependency path data of the sampling points can be used as the actual data points, and the least squares method can be used to fit a curve as a baseline sequence of conduction trajectories.

[0031] Step S224: Identify the location where the dependent path in the conduction trajectory reference sequence breaks, and mark it as a conduction trajectory fault. The fault location corresponds to the area of ​​missing conduction information caused by the sampling gap.

[0032] A break in the dependency path means that the conduction correlation at a certain location in the conduction trajectory reference sequence is discontinuous. This break may be caused by sampling gaps, which occur when some data points are missing during the sampling process due to various reasons (such as insufficient sampling frequency, signal interference, etc.). When sampling gaps occur, conduction information is missing at these locations, thus causing a break in the dependency path.

[0033] Marking these as conduction trajectory faults involves clearly identifying the locations where these dependent paths break. The fault locations correspond to areas where conduction information is missing due to sampling gaps. By marking these fault locations, a basis can be provided for subsequent completion and repair operations.

[0034] Specifically, abrupt changes in the conduction trajectory baseline sequence can be detected by calculating the derivative or difference; these abrupt changes may indicate breaks in the dependent path. Alternatively, machine learning algorithms, such as anomaly detection algorithms, can be used to identify anomalous patterns in the conduction trajectory baseline sequence, which may correspond to conduction trajectory faults.

[0035] Step S225: Perform pre-complete of the fault location of the conduction trajectory, deduce the basic conduction properties of the fault location based on the dependency rules of adjacent normal conduction segments within the group, and generate a reference sequence of the pre-completed conduction trajectory of the fault.

[0036] In an optional implementation, step S225 may specifically include the following steps S2251 to S2256: Step S2251: Extract the normal conduction segment dependency features on both sides of the fault from the conduction trajectory reference sequence corresponding to the conduction trajectory fault, and generate a set of conduction dependency features on both sides of the fault.

[0037] Dependency features of normal conduction segments include conduction direction, dependency node association patterns, and phase progression rhythm. The conduction direction is the direction of signal propagation during conduction, which can be unidirectional or bidirectional. Dependency node association patterns refer to the ways and patterns in which sampling points are associated with other nodes or signals during conduction; for example, some sampling points may only be associated with specific nodes, or have complex associations with multiple nodes. Phase progression rhythm is the rhythm of signal changes during clock phase progression, reflecting the conduction speed and changes of the signal at different phases. By extracting these dependency features, a set of conduction dependency features on both sides of the fault can be generated. This set contains information such as the conduction direction, dependency node association patterns, and phase progression rhythm of the normal conduction segments on both sides of the fault, providing a foundation for subsequent interpolation and derivation.

[0038] For example, data extraction and analysis algorithms can be used. For instance, the normal conduction segments on both sides of the fault can be separated by segmenting the baseline sequence of the conduction trajectory, and then features can be extracted from these normal conduction segments. The conduction direction can be determined by analyzing the conduction correlation between sampling points; the correlation patterns of dependent nodes can be summarized by statistically analyzing the frequency and manner of association between sampling points and other nodes; and the phase progression rhythm can be analyzed by calculating the rate of change of sampling points at different phases.

[0039] Step S2252: Perform continuous dependency interpolation on the set of conduction dependency features on both sides of the fault, and generate a conduction dependency interpolation sequence from one side to the other based on the transition law of the features on both sides. The conduction dependency interpolation sequence contains the complete process of continuous change of dependency features.

[0040] The set of conduction dependency features on both sides of the fault contains information such as the conduction direction of the normal conduction segments on both sides of the fault, the correlation rules of dependency nodes, and the phase advancement rhythm. Based on the transition rules of the features on both sides, a conduction dependency interpolation sequence from one side to the other is generated. This is done by inserting a series of continuous dependency feature values ​​at the fault location according to the changing trends and rules of the features on both sides of the fault, so that the conduction dependency features can transition smoothly at the fault location.

[0041] The conduction dependency interpolation sequence encompasses the complete process of continuous change in dependency features, gradually transitioning from the normal conduction segment features on one side of the fault to the normal conduction segment features on the other side. To achieve continuous dependency interpolation, common interpolation algorithms such as linear interpolation and spline interpolation can be used. Linear interpolation assumes that the features change linearly between two known data points; spline interpolation, on the other hand, can more accurately estimate unknown data points by fitting a smooth curve. In DDR wafer testing, the appropriate interpolation algorithm can be selected based on the specific characteristics of the conduction dependency features on both sides of the fault. For example, in memory testing, when the changes in conduction dependency features on both sides of the fault are relatively gentle, a linear interpolation algorithm can be used to generate the conduction dependency interpolation sequence; when the feature changes are more complex, a spline interpolation algorithm can be used to obtain more accurate results.

[0042] Step S2253: Match the conduction-dependent interpolation sequence with the conduction law of the DDR jitter physical prior, adjust the characteristic parameters of the interpolation sequence to make it fully fit the DDR jitter physical characteristics, and generate the calibrated conduction-dependent interpolation sequence.

[0043] DDR jitter physics priors are a priori knowledge derived from the physical principles and characteristics of DDR jitter, including the jitter generation mechanism, jitter distribution patterns, and the relationship between jitter and signal parameters. Conduction laws are some basic laws and characteristics of signal transmission in DDR chips, such as transmission speed, transmission direction, and transmission delay.

[0044] Matching the conduction-dependent interpolation sequence with the prior physical characteristics of DDR jitter is a way to check whether the conduction-dependent interpolation sequence conforms to the physical characteristics of DDR jitter. If some characteristic parameters of the interpolation sequence do not conform to the prior physical characteristics of DDR jitter, these characteristic parameters need to be adjusted. The purpose of the adjustment is to make the characteristic parameters of the interpolation sequence perfectly match the physical characteristics of DDR jitter, thereby more accurately reflecting the signal propagation at the fault location.

[0045] The generated calibrated propagation-dependent interpolation sequence is a new interpolation sequence obtained after adjusting the feature parameters. This calibrated sequence better conforms to the physical characteristics of DDR (Dual Generated Dependency) jitter. Specifically, model matching and parameter optimization algorithms can be used. For example, a DDR jitter physical model can be established, the propagation-dependent interpolation sequence can be input into the model, and the propagation law of the model output and the DDR jitter physical prior can be compared to identify the feature parameters that need adjustment. Then, optimization algorithms, such as gradient descent, can be used to adjust these feature parameters so that the interpolation sequence gradually conforms to the physical characteristics of DDR jitter.

[0046] Step S2254: Based on the calibrated conduction-dependent interpolation sequence, derive the conduction properties of each phase point at the fault location and generate a set of conduction-dependent properties at the fault location.

[0047] Conduction properties include, for example, direction of conduction, dependency node association, and phase advance rhythm. Direction of conduction is the direction of signal propagation during the conduction process; dependency node association is the way and relationship between the sampling point and other nodes or signals during the conduction process; phase advance rhythm is the rhythm of signal change during the clock phase advance process.

[0048] The conduction properties of each phase point at the fault location are derived by using the characteristic values ​​of each phase point in the calibrated conduction-dependent interpolation sequence, combined with the physical characteristics of DDR jitter and conduction laws, to calculate the conduction direction, dependent node association, and phase advance rhythm of that phase point. This derivation can fill the gaps in conduction information at the fault location caused by sampling intervals.

[0049] Generating a set of conduction-dependent attributes for the fault location involves organizing and recording the conduction attributes of each phase point obtained from the derivation, forming a set. This set contains conduction attribute information for all phase points at the fault location, providing detailed data support for subsequent completion and repair operations.

[0050] For example, the direction of conduction can be determined based on the amplitude and phase changes of adjacent phase points in the calibrated conduction dependency interpolation sequence, combined with the typical conduction paths of signals under different circuit structures in the DDR jitter physical prior. If the amplitude of the subsequent phase point increases relative to the previous phase point and the phase increases, and considering the positive conduction pattern of the corresponding circuit part of the signal in the DDR jitter physical prior, the direction of conduction can be determined to be positive. If the amplitude decreases and the phase increases, considering the possible reverse conduction or signal attenuation in the physical prior, the direction of conduction may be reversed or in the attenuation stage. For the correlation of dependent nodes, it can be deduced by analyzing the feature similarity between each phase point in the calibrated conduction dependency interpolation sequence and other known nodes. The similarity between the feature vector of the current phase point and the feature vector of the known node is calculated. If the similarity exceeds a preset threshold, it is considered that the phase point and the corresponding node have a dependency correlation. The feature vector can contain information such as the amplitude, frequency, and phase of the phase point. The derivation of the phase advancement rhythm can be achieved by first statistically analyzing the time interval change pattern between adjacent phase points in the calibrated conduction dependency interpolation sequence. If the time interval is relatively stable, it indicates that the phase advancement rhythm is relatively uniform; if the time interval varies greatly, exhibiting periodic or non-periodic fluctuations, further analysis of the causes and patterns of these fluctuations is needed. The stability of the rhythm can be measured by calculating the standard deviation of the time interval between adjacent phase points; the smaller the standard deviation, the more uniform the rhythm. Simultaneously, comparisons and adjustments should be made based on typical rhythm patterns of clock and data signal phase advancements from DDR jitter priors.

[0051] In practice, these derivations can be achieved using data processing and computational algorithms. For example, correlation analysis algorithms can be used to calculate the similarity between phase point feature vectors; statistical analysis algorithms can be used to calculate the standard deviation of the time interval between adjacent phase points, etc.

[0052] Step S2255: Insert the set of conduction-dependent attributes of the fault location into the corresponding conduction trajectory reference sequence to fill the missing conduction information of the fault location and generate the conduction trajectory reference sequence of the inserted fault.

[0053] The set of conduction-dependent attributes at fault locations contains the conduction attributes of each phase point at the fault location, such as conduction direction, dependency node associations, and phase advancement rhythm. Inserting this set into the conduction trajectory baseline sequence fills in these conduction attribute information at the fault locations within the sequence, thus filling in areas where conduction information is missing. Generating the inserted fault conduction trajectory baseline sequence is the result of inserting the set of conduction-dependent attributes at the fault locations. To implement the insertion operation, data insertion and merging algorithms can be used. For example, data from the set of conduction-dependent attributes at the fault locations can be inserted into the corresponding positions based on the index of the fault location in the conduction trajectory baseline sequence.

[0054] Step S2256: Verify the conduction timing coherence of the inserted fault conduction trajectory reference sequence to ensure that the conduction timing of the inserted sequence is consistent with that of the original reference sequence, and generate a fault pre-complete conduction trajectory reference sequence.

[0055] The transmission timing sequence refers to the temporal order and rhythm of a signal during transmission, reflecting the transmission speed and changes at different phases. If the transmission timing sequence of the inserted sequence differs from the original reference sequence, it may lead to misinterpretation of the transmission information and inaccurate subsequent analysis.

[0056] Conductivity time series coherence verification checks whether the inserted sequence is consistent with the original reference sequence in terms of conductivity time series. If inconsistencies are found, the conductivity dependency attributes of the inserted sequence can be adjusted to match the conductivity time series of the original reference sequence. The generated tomographic pre-complete conduction trajectory reference sequence is the final sequence obtained after conductivity time series coherence verification and adjustment. To achieve conductivity time series coherence verification, time series analysis and verification algorithms can be used. For example, the timestamps and conduction order of sampling points in the inserted tomographic trajectory reference sequence and the original reference sequence can be compared, and their differences can be calculated. If the difference exceeds a certain threshold, the conductivity time series is considered inconsistent and adjustment is required.

[0057] Step S226: Integrate all fault pre-complete conduction trajectory reference sequences according to the progression order of the conduction initiation phase to generate phase conduction dependent links, each link corresponding to a set of continuous and complete phase conduction processes.

[0058] Phase-propagation dependent links consist of integrated tomographic pre-complete conduction trajectory reference sequences, with each link corresponding to a continuous and complete phase conduction process. Through integration, scattered conduction trajectory reference sequences can be connected to form a complete conduction link, more clearly demonstrating the conduction patterns and interactions of signals during clock phase progression. Sorting and merging algorithms, such as quicksort, can be used to sort all tomographic pre-complete conduction trajectory reference sequences according to the conduction initiation phase, and then the sorted sequences can be merged sequentially.

[0059] Step S230: Complete the location of the conduction path break in each phase conduction dependent link, deduce the conduction properties of the break location based on the dependency rules of adjacent conduction segments in the link, generate a complete phase conduction dependent link, and fill the conduction information missing caused by the sampling gap.

[0060] First, the conduction attributes are quantified. The conduction direction can be set as two states: forward and reverse, represented by designated symbols or codes. An association matrix is ​​constructed for dependent nodes, with matrix elements reflecting the strength of the association between nodes. The phase progression rhythm can be measured by the time interval between adjacent phase points. Next, features of adjacent conduction segments are extracted. For the conduction direction, its changing trend is analyzed, such as whether it is continuous forward or reverse conduction, or whether there are periodic direction changes, recording the location and frequency of these changes. Regarding dependent node association, changes in the elements of the association matrix are studied to see if the association strength increases, decreases, or remains stable, and whether any new or fewer associated nodes are added. For the phase progression rhythm, the statistical characteristics of the time interval between adjacent conduction segments are calculated, such as the mean and standard deviation, to understand the stability and fluctuation of the rhythm.

[0061] In actual DDR wafer testing scenarios, at the break point of a phase-dependent transmission path, adjacent transmission segments may exhibit certain patterns of transmission direction changes. For example, multiple consecutive sampling points may maintain forward transmission before reverse transmission occurs, or a periodic alternation of forward and reverse directions may occur. Regarding the correlation of dependent nodes, the correlation matrix of adjacent transmission segments may show that the correlation strength between some nodes gradually weakens, while new nodes join the correlation. In terms of phase advancement rhythm, the time interval between adjacent transmission segments may show a gradual increasing or decreasing trend, or fluctuate relatively stably.

[0062] Based on these characteristics, the conduction properties are deduced. Regarding conduction direction, if adjacent conduction segments exhibit a periodic direction change pattern, and the fracture location is at a specific position within the period, the conduction direction at the fracture location can be inferred according to this periodic pattern. For example, if it is found that adjacent conduction segments exhibit a pattern of forward conduction followed by reverse conduction after a period, and the fracture location is at the periodic position corresponding to the forward conduction, the conduction direction is inferred to be forward. Regarding dependency node association, if the association strength in adjacent conduction segments exhibits a certain trend, it can be extrapolated using appropriate mathematical methods. For example, if the association strength exhibits a linear decreasing trend, the association strength at the fracture location can be inferred based on the decreasing slope. Regarding phase progression rhythm, if the time interval between adjacent conduction segments exhibits a certain changing trend, the time interval at the fracture location can be predicted based on this trend. For example, if the time interval gradually increases, the time interval at the fracture location can be inferred based on the increasing trend.

[0063] To achieve the derivation and completion operations, regression algorithms from machine learning, such as linear regression and multinomial regression, can be used. Taking linear regression as an example, the conduction attribute data of adjacent conduction segments are used as the training set to construct a linear relationship model between conduction attributes and location. For the conduction direction, location can be used as the independent variable and the conduction direction label as the dependent variable. A linear equation is fitted using the least squares method, and then the location information of the break point is substituted into the equation to obtain the predicted conduction direction. A similar method is used for dependent node associations and phase advancement rhythms, respectively constructing corresponding linear regression models for prediction. In practical applications, Bayesian inference algorithms can also be combined, using prior knowledge and data from adjacent conduction segments to calculate the posterior probability of the conduction attribute at the break point, further improving the accuracy of the derivation. Through these algorithms and methods, the break point of the conduction path is finally completed, generating a complete phase conduction dependency link and filling the conduction information gap caused by sampling gaps.

[0064] Step S240: Compare all complete phase propagation dependent links with the propagation path rules of the DDR jitter physical prior, identify link segments whose propagation paths do not conform to the DDR jitter physical prior rules, and generate a set of abnormal propagation dependent segments.

[0065] The comparison involves comparing the conduction path in the complete phase-dependent link with the prior physical conduction path pattern of the DDR jitter mechanism to check for consistency. If a link segment's conduction path deviates from the prior physical pattern of DDR jitter, it indicates a potential anomaly in that segment. Identifying link segments with inconsistent conduction paths is achieved through the comparison process, pinpointing those non-compliant segments. These segments may be due to chip manufacturing defects, signal interference, or other reasons.

[0066] To enable comparison and identification, pattern matching and rule-based judgment algorithms can be used. For example, a physical prior model of DDR shaking can be established, and the complete phase-dependent transmission link can be input into the model. By comparing the model output with the actual transmission path, segments that do not conform to the rules can be identified.

[0067] Step S250: Encode the set of abnormal conduction dependency segments, convert each conduction segment into a conduction dependency coding sequence containing phase dependency relationships, and obtain a set of conduction dependency coding sequences. Each coding unit in the conduction dependency coding sequence corresponds to a conduction node and dependency path information in the link.

[0068] Converting each conduction segment into a conduction dependency encoded sequence that includes phase dependencies involves quantizing and encoding the phase dependencies within the conduction segment to form a sequence. Phase dependencies are the conduction associations and dependencies between sampling points in a conduction segment at different phases.

[0069] Each coding unit in the transmission dependency coding sequence corresponds to a transmission node and dependency path information in the link. A transmission node is a key node in the transmission process, and the dependency path information describes the transmission association and path characteristics between this node and other nodes. This coding method simplifies and organizes the complex information of the transmission segment, facilitating subsequent analysis and processing.

[0070] To achieve encoding conversion, encoding algorithms can be used, such as Huffman coding and run-length encoding. Taking Huffman coding as an example, this algorithm constructs a Huffman tree, representing frequently occurring symbols with shorter codes and less frequently occurring symbols with longer codes, thereby achieving data compression and encoding.

[0071] Step S260: Remove the conduction-dependent coding sequence corresponding to the normal test state of the DDR wafer from the set of conduction-dependent coding sequences, retain the remaining coding sequences and integrate them into defect-sensitive sparse components. Each retained coding sequence corresponds to phase conduction information that deviates from the physical prior.

[0072] Removing conduction-dependent coding sequences corresponding to normal test states of the DDR wafer from the set of conduction-dependent coding sequences involves removing these normal coding sequences from the set and retaining only those coding sequences that may be related to defects. Retaining the remaining coding sequences and integrating them into a defect-sensitive sparse component involves merging and rearranging the remaining coding sequences after removing the normal coding sequences to form a new component. Each retained coding sequence in the defect-sensitive sparse component corresponds to phase conduction information deviating from the physical prior, which may be related to defects or anomalies in the DDR wafer.

[0073] For example, data filtering and integration algorithms can be used. For instance, the characteristics and ranges of the conduction-dependent coding sequences corresponding to the normal test state of a DDR wafer can be predefined. By comparing each coding sequence in the set of conduction-dependent coding sequences with these characteristics and ranges, coding sequences that conform to the normal state are eliminated. Then, data merging and sorting algorithms can be used to integrate the remaining coding sequences into defect-sensitive sparse components.

[0074] Step S300: Perform variational mode decomposition on the defect-sensitive sparse components to separate the components related to process drift, normal process fluctuations and abrupt changes, and extract the abrupt change-related components as target detection data.

[0075] In an optional implementation, step S300 may specifically include the following steps S310 to S360: Step S310: Divide the defect-sensitive sparse component into multiple time segments according to three consecutive clock cycles, analyze the amplitude-frequency co-evolution trend of the signal in each time segment, and generate an amplitude-frequency co-evolution sequence. The sequence contains the amplitude change rate and frequency propagation trajectory of each time segment.

[0076] In an optional implementation, step S310 may specifically include the following steps S311 to S316: Step S311: Extract the amplitude and corresponding frequency values ​​of all sampling points from each three-period time series of the defect-sensitive sparse component, arrange them in the sampling order to generate a time series amplitude-frequency subsequence, and retain the phase correlation and co-evolution markers of the sampling points in the time series amplitude-frequency subsequence.

[0077] Generating a time-series amplitude-frequency subsequence by arranging the extracted amplitude and frequency values ​​according to the order of sampling is to form a subsequence. This arrangement method can preserve the order and trend of signal changes within the time segment.

[0078] The amplitude-frequency subsequence of the time segment retains the phase correlation and co-evolution markers of the sampling points. The phase correlation is the relationship between the sampling points in the clock phase, which reflects the position of the sampling points within the clock signal period; the co-evolution marker is a marker used to identify the co-evolution relationship between amplitude and frequency, through which the co-evolution of the signal amplitude and frequency at the sampling points can be understood.

[0079] For example, signal processing and data extraction algorithms can be used. For instance, the Fast Fourier Transform (FFT) algorithm can be used to perform spectral analysis on the signal within a time series to obtain the frequency value at each sampling point; simultaneously, amplitude information can be directly extracted from the signal's sampled data. Then, the amplitude and frequency values ​​can be sorted and organized through programming to generate a time series amplitude-frequency subsequence.

[0080] Step S312: Analyze the correlation between each sampling point in the amplitude-frequency subsequence of the time segment and the amplitude-frequency changes of the previous three and the last three sampling points, calculate the continuous change slope of the amplitude and the continuous conduction velocity of the frequency, and generate a slope-velocity change sequence. The slope-velocity change sequence contains the dual change information of each sampling point.

[0081] By analyzing the correlation between the amplitude and frequency changes of each sampling point and the preceding and following three sampling points, we can understand the trends and interrelationships of amplitude and frequency changes between adjacent sampling points. For example, the amplitude and frequency changes of a sampling point may be affected by the preceding and following sampling points; analyzing this correlation can reveal the dynamic characteristics of the signal.

[0082] The slope of the continuous change in amplitude and the continuous propagation rate of frequency are calculated. The slope of the continuous change in amplitude is the rate of change of amplitude between adjacent sampling points, reflecting the speed and direction of amplitude change; the continuous propagation rate of frequency is the rate of change of frequency between adjacent sampling points, reflecting the trend of frequency change. By calculating these two parameters, the changes in signal amplitude and frequency between adjacent sampling points can be described more accurately. Generating a slope-rate change sequence involves organizing and recording the slope of the continuous change in amplitude and the continuous propagation rate of frequency at each sampling point to form a sequence. This sequence contains dual change information for each sampling point, namely the slope of the amplitude change and the propagation rate of frequency, providing more detailed signal change characteristics for subsequent analysis and processing. For example, data calculation and analysis algorithms can be used. For instance, the slope of the continuous change in amplitude and the continuous propagation rate of frequency can be obtained through differential calculation.

[0083] Step S313: Connect the sampling point segments in the slope-rate change sequence where the slope and rate changes are continuous and consistent into a co-evolution sub-chain. Each sub-chain corresponds to a change process in amplitude-frequency co-stable, and the phase, slope and rate correlation of the sampling points in the sub-chain are preserved.

[0084] The continuous and consistent slope and rate of change mean that within this sampling point segment, the amplitude and frequency of the signal change in a relatively stable manner. For example, if the slope of the amplitude change always remains positive and the frequency propagation rate always remains within a stable range, then the amplitude and frequency of the signal can be considered to be changing stably and in a coordinated manner.

[0085] Each subchain corresponds to a stable amplitude-frequency change process. By separating these subchains, the coordinated change characteristics of the signal at different stages can be observed and analyzed more clearly. Preserving the phase, slope, and rate correlation of the sampling points in the subchains is to record the position of these sampling points on the clock phase and the correlation between their amplitude change slope and frequency propagation rate, providing a basis for subsequent analysis and comparison.

[0086] For example, sequence analysis and clustering algorithms can be used. For instance, by traversing the slope-rate change sequence, sampling points with continuous slope and rate changes can be identified, and these segments can be connected to form co-evolutionary subchains.

[0087] Step S314: Identify sampling point segments in the co-evolutionary subchains where the slope or rate changes exhibit third-order abrupt changes, and mark them as anomalous co-evolutionary segments. The amplitude-frequency co-changes within these segments are discontinuous and differ from those in adjacent subchains.

[0088] In an optional implementation, step S314 may specifically include the following steps S3141 to S3146: Step S3141: Extract the slope-rate change sequence from the co-evolutionary subchain, arrange them in the sampling order to generate a double change sequence, which contains the amplitude change slope, frequency conduction rate and phase correlation information of each sampling point.

[0089] When identifying sampling points in a co-evolutionary subchain where slope or rate changes exhibit third-order abrupt changes, the first step is to extract the slope-rate change sequence from the co-evolutionary subchain. A co-evolutionary subchain is formed by connecting sampling points with continuous and consistent slope and rate changes; the slope-rate change sequence records the amplitude change slope and frequency propagation rate at each sampling point.

[0090] The dual-change sequence is generated by arranging the extracted slope-rate change sequences according to the order of sampling, forming a new sequence. The dual-change sequence contains the amplitude change slope, frequency propagation rate, and phase correlation information for each sampling point. The phase correlation information is the position of the sampling point on the clock phase, reflecting its relative position within the clock signal period.

[0091] Step S3142: Calculate the slope difference and rate difference between each sampling point in the dual change sequence and the previous three and the last three sampling points to generate a third-order difference change sequence. The third-order difference change sequence contains the third-order slope difference and third-order rate difference information of each sampling point.

[0092] The slope difference is the difference between the amplitude change slope of the current sampling point and the amplitude change slope of the previous three and the next three sampling points. The rate difference is the difference between the frequency conduction rate of the current sampling point and the frequency conduction rate of the previous three and the next three sampling points.

[0093] Calculating the third-order difference sequence involves further difference calculations on the slope difference and rate difference. The third-order slope difference is the difference of the slope difference, and the third-order rate difference is the difference of the rate difference. By calculating the third-order difference, sudden changes in slope and rate can be captured more sensitively because the third-order difference can highlight the changes in the changes, i.e., the acceleration of the changes.

[0094] The third-order difference sequence contains the third-order slope difference and third-order rate difference information for each sampling point. This sequence can be used for subsequent mutation detection and analysis. If the third-order slope difference or third-order rate difference of a sampling point has a large value, it indicates that a third-order mutation has occurred in the slope or rate of that sampling point.

[0095] Step S3143: Compare each slope difference in the third-order differential change sequence with the range of third-order slope differences under normal DDR wafer testing conditions, and compare each rate difference with the range of third-order rate differences under normal conditions to confirm whether each difference is within the corresponding normal coordinated change range.

[0096] The third-order slope difference range and third-order rate difference range under normal DDR wafer testing conditions were obtained through statistical analysis of a large amount of normal test data. They represent the range of variation of the third-order difference of the signal slope and rate under normal conditions.

[0097] The purpose of the comparison is to confirm whether each difference is within the corresponding normal range of coordinated change. If the third-order slope difference or third-order rate difference of a certain sampling point exceeds the normal range, it indicates that the slope or rate change at that sampling point is abnormal, which may indicate a problem with the signal.

[0098] Step S3144: Mark the sampling points corresponding to the differences that exceed the normal range with mutations. The marking content includes the magnitude, direction and corresponding phase position of the mutation, and generate a set of mutation sampling points.

[0099] The magnitude of the mutation is the change in slope or rate, which reflects the size of the mutation; the direction of the mutation is whether the change in slope or rate is increasing or decreasing; the corresponding phase position is the position of the sampling point on the clock phase, which can help determine the time point when the mutation occurred.

[0100] Generating a mutation sampling point set involves organizing and recording all sampling points marked as mutations into a single set. This set contains information on all sampling points where the slope or rate undergoes a third-order mutation, providing a target for subsequent analysis and processing.

[0101] Step S3145: Connect consecutive sampling points in the mutation sampling point set to form mutation segments. Each mutation segment corresponds to a continuous change process of amplitude-frequency coordinated mutation, and the temporal order and dual change information of the sampling points within the segment are preserved.

[0102] Each abrupt change segment corresponds to a continuous abrupt change in amplitude and frequency, during which the signal amplitude and frequency change in a continuous and anomalous manner. Preserving the timing sequence and dual-change information of the sampling points within the segment is to record the order of these sampling points in the clock phase, as well as their amplitude change slope and frequency propagation rate, providing a basis for subsequent analysis and comparison.

[0103] For example, sequence analysis and merging algorithms can be used. For instance, by traversing the set of mutation sampling points, consecutive sampling point segments can be found, and then these segments can be connected to form mutation segments.

[0104] Step S3146: Integrate all mutation segments into a set of anomalous co-evolutionary segments, each segment corresponding to an anomalous mutation process in amplitude-frequency co-evolution.

[0105] Each segment corresponds to an anomalous mutation process in the amplitude-frequency co-evolution, which may be caused by chip failure, external interference, or other reasons. The set of anomalous co-evolution segments contains information on all segments where anomalous mutations occur in the amplitude-frequency co-evolution. To achieve integration, data merging and organization algorithms can be used. For example, the data of all mutation segments can be summarized and organized programmatically to form a unified set of anomalous co-evolution segments.

[0106] Step S315: Fit the cooperative trend of the abnormal cooperative evolution segment, and fit the corrected cooperative trend of the abnormal segment based on the slope and rate of the adjacent normal cooperative evolution sub-chain to generate the corrected cooperative evolution sub-chain.

[0107] The corrected co-evolutionary trend of anomalous segments is derived from fitting the slopes and rates of adjacent normal co-evolutionary subchains. This is because the slopes and rates of adjacent normal co-evolutionary subchains reflect the amplitude-frequency co-evolutionary pattern of the signal under normal conditions. By referencing these normal patterns, the co-evolutionary trend of anomalous segments can be corrected. For example, if the amplitude change slope and frequency propagation rate of adjacent normal co-evolutionary subchains exhibit a linear relationship, it can be assumed that the anomalous segment, after correction, also follows a similar linear relationship, thus allowing for fitting.

[0108] The generation of the corrected co-evolutionary subchain involves merging the corrected co-evolutionary segment with the original normal co-evolutionary subchain after fitting the corrected co-evolutionary trend of the anomalous segment, forming a new co-evolutionary subchain. This new subchain has been corrected in the anomalous segment portion, and thus better reflects the true co-evolutionary changes of the signal.

[0109] To achieve co-evolutionary trend fitting, curve fitting algorithms, such as the least squares method, can be used. The least squares method determines the parameters of the fitted curve by minimizing the sum of squared errors between the actual data points and the fitted curve. In DDR wafer testing, data from anomalous co-evolutionary segments can be used as actual data points. The type of fitted curve is determined based on the slope and rate of adjacent normal co-evolutionary subchains, and then the least squares method is used for fitting. For example, for an anomalous co-evolutionary segment, by referring to the slope and rate of adjacent normal co-evolutionary subchains, a linear function is determined for fitting. Then, the parameters of the linear function are calculated using the least squares method to obtain the corrected co-evolutionary trend, and subsequently, the corrected co-evolutionary subchain is generated.

[0110] Step S316: Integrate all the corrected co-evolution subchains according to the sampling order to generate the amplitude-frequency co-evolution sequence for each time segment. The sequence contains the complete amplitude-frequency co-evolution trend and the corrected anomalous segment information within the time segment.

[0111] A magnitude-frequency co-evolution sequence is generated for each time segment. This sequence contains the complete magnitude-frequency co-evolution trend within the time segment and the corrected anomalous segment information. The magnitude-frequency co-evolution trend describes the overall trend of the signal's amplitude and frequency within the time segment, while the corrected anomalous segment information records the specific details of the corrected anomalous co-evolution segments.

[0112] Step S320: Perform third-order dependency analysis on each trend segment in the amplitude-frequency co-evolution sequence to identify trend segment groups with continuous co-evolutionary associations. Each group of trend segments covers the co-evolution process of three time segments, and preserves the temporal order and co-evolutionary association of the trend segments within the group.

[0113] Identifying trend segments with continuous co-evolutionary relationships involves using third-order dependency analysis to find those trend segments that exhibit continuous co-evolutionary relationships in amplitude, frequency, and time, and then grouping them together. Each group of trend segments covers the co-evolutionary process across three time periods, which helps to more comprehensively observe and analyze the co-evolutionary characteristics of signals over a relatively long period.

[0114] Preserving the temporal sequence and co-correlation of trend segments within a group allows for a clearer understanding of the co-changing patterns of signals across different trend segments. The temporal sequence reflects the sequential order of trend segments over time, while the co-correlation describes the relationships between trend segments in terms of amplitude and frequency. By retaining this information, it is possible to gain a clearer understanding of the co-changing patterns of signals across different trend segments.

[0115] To perform third-order dependency analysis and identify trend segment groups, data analysis and clustering algorithms can be used. For example, the degree of co-evolutionary association between trend segments can be determined by calculating indicators such as correlation coefficients and distances. Then, trend segments with continuous co-evolutionary associations can be grouped together based on these indicators.

[0116] Step S330: Extract the cross-period evolution features of each trend segment group, deduce the continuous evolution trajectory across time periods based on the amplitude-frequency co-evolution rate within the trend segment, and generate a set of cross-period evolution trajectories containing the amplitude-frequency co-change paths of each trend segment.

[0117] Specifically, the amplitude-frequency co-evolution sequence within the trend segment group is first analyzed, which records the coordinated changes in signal amplitude and frequency over time within the trend segment. The rate of amplitude change can be calculated by differential processing of the amplitude data in the amplitude-frequency co-evolution sequence. Differential processing involves calculating the difference in amplitude between adjacent time points; this difference reflects the amount of amplitude change within that short time interval. Dividing this change by the corresponding time interval yields the rate of amplitude change during that time period. For example, within a short time interval, by comparing the amplitude values ​​of adjacent sampling points, the increase or decrease in amplitude can be calculated, and combined with the sampling time interval, the instantaneous rate of amplitude change can be obtained. The calculation of the rate of frequency change follows a similar principle, but due to the characteristics of frequency data, a more complex method may be required. Spectral analysis techniques, such as the Fast Fourier Transform (FFT), can be used to convert the time-domain signal to the frequency domain, thereby obtaining more accurate frequency information. Then, differential calculation is performed on the frequency data at different time points to obtain the change in frequency between adjacent time points, which is then divided by the time interval to obtain the rate of frequency change.

[0118] After obtaining the rates of change of amplitude and frequency, the continuous evolution trajectory across time periods can be derived using polynomial fitting. Based on the calculated rates of change and known amplitude-frequency co-evolution sequence data, a polynomial function is fitted. This polynomial function can approximately represent the variation patterns of amplitude and frequency throughout the entire trend segment group. Using this polynomial function, the amplitude and frequency values ​​at any time point within the trend segment group can be predicted, thus constructing the continuous evolution trajectory across time periods. During the fitting process, an appropriate polynomial order must be selected. If the order is too low, it may not accurately fit the data's changing trend; if the order is too high, it may lead to overfitting, reducing the model's generalization ability. The optimal polynomial order can be determined using methods such as cross-validation. Furthermore, the derivation process can be optimized by incorporating prior knowledge of DDR (Digital Frequency Reduction) propagation. DDR propagation prior knowledge includes information such as the amplitude and frequency variation range and trends of the signal under normal conditions. When deriving the cross-time period evolution trajectory, this prior knowledge is used as a constraint to adjust the fitted polynomial function. If the amplitude or frequency predicted by the fitted function exceeds the normal range specified by the physical prior, the parameters of the function are corrected to make the derived evolution trajectory more consistent with the actual physical situation.

[0119] Step S340: Compare the set of cross-time evolution trajectories with the preset cross-time evolution trajectories corresponding to process drift and normal process fluctuations, and mark the trend segment groups belonging to the two types of trajectories.

[0120] The purpose of the comparison is to identify which trend segments in the cross-period evolution trajectory set match the cross-period evolution trajectory corresponding to process drift or normal process fluctuation. If the trajectory of a certain trend segment matches one of these types of trajectories, it indicates that the signal change corresponding to that trend segment may be caused by process drift or normal process fluctuation.

[0121] The labeling of trend segment groups belonging to the two types of trajectories involves clearly identifying trend segment groups that match the cross-period evolution trajectory corresponding to process drift or the cross-period evolution trajectory corresponding to normal process fluctuations. This allows these trend segment groups to be distinguished from other trend segment groups that may be abnormal.

[0122] Pattern matching and labeling algorithms can be used. For example, the similarity between the trajectory of each trend segment group in the cross-time evolution trajectory set and a preset trajectory can be calculated. When the similarity reaches a certain threshold, they are considered to match, and the trend segment group is labeled.

[0123] Step S350: Remove the marked trend segment groups from the cross-time evolution trajectory set, retain the time series corresponding to the remaining trend segment groups, convert the remaining time series into time domain signals, and generate a mutation anomaly co-evolution component.

[0124] The remaining time series segments corresponding to the trend segments are retained, as the signal changes in these segments may be caused by abrupt anomalies. The remaining time series segments are converted to time-domain signals because time-domain signals are easier to analyze and process subsequently. A time-domain signal is a signal representation where time is the independent variable and signal amplitude is the dependent variable.

[0125] Generating a mutation anomaly co-evolutionary components involves organizing and recording the transformed time-domain signal to form a new component. This component contains amplitude-frequency co-evolutionary information that may be related to mutation anomalies.

[0126] Step S360: Integrate all mutational anomalous co-evolutionary components into mutational anomalous related components in their original temporal order, which serve as target detection data. This data includes cross-time-period anomalous co-evolutionary information in both amplitude and frequency dimensions.

[0127] As target detection data, the mutation anomaly-related component contains cross-temporal anomaly co-evolution information in both amplitude and frequency dimensions. The amplitude dimension describes the changes in signal amplitude, the frequency dimension describes the changes in signal frequency, and the cross-temporal dimension indicates that this anomaly co-evolution information is accumulated and manifested over multiple time periods.

[0128] Step S400: Input the target detection data into the preset machine learning model, and generate the abnormal pattern association result by association modeling and abnormal pattern matching processing of the target detection data.

[0129] In an optional implementation, step S400 may specifically include the following steps S410 to S460: Step S410: Divide the target detection data into multiple time units in groups of three consecutive clock cycles. Each unit corresponds to a set of three-cycle abnormal co-evolutionary components, and retain the amplitude-frequency co-evolutionary correlation and multi-order transmission information within the unit.

[0130] Before inputting the target detection data into the preset machine learning model, further processing is required. The target detection data is divided into multiple time-series units, grouped into sets of three consecutive clock cycles. This division helps to structure the target detection data, facilitating analysis and modeling. Each unit corresponds to a set of three-cycle anomalous co-evolutionary components, which contain amplitude-frequency co-evolutionary information of the signal within three consecutive clock cycles. The amplitude-frequency co-evolutionary correlation and multi-order conduction information within the unit are preserved. The amplitude-frequency co-evolutionary correlation describes the cooperative change relationship between amplitude and frequency within the unit, while the multi-order conduction information reflects the conduction and evolution relationship of the signal between different time segments.

[0131] Step S420: Input the time series unit into the preset machine learning model, trigger the cross-unit third-order dependency modeling operation inside the model, explore the co-evolutionary relationship between each time series unit and the first three groups and the last three groups of time series units, and generate a set of cross-group dependency relationships between units.

[0132] After obtaining multiple temporal units, they are fed into a pre-defined machine learning model. The pre-defined machine learning model can be a deep learning model, such as a recurrent neural network (RNN) or a long short-term memory network (LSTM). These models have the ability to process sequential data and can extract temporal correlation information from the data.

[0133] This triggers cross-unit third-order dependency modeling operations within the model. Cross-unit third-order dependency modeling analyzes the dependency relationships between each temporal unit and the preceding three groups and the following three groups of temporal units in three dimensions: amplitude, frequency, and time. Through this modeling operation, the co-evolutionary correlations between temporal units can be uncovered.

[0134] In an optional implementation, step S420 may specifically include the following steps S421 to S426: Step S421: Extract the mined cross-group evolutionary relationships of units from the preset machine learning model and generate the original set of relationships. The original set of relationships contains the evolutionary direction, rate and cross-group transmission relationship information of each time unit with other units.

[0135] After completing the cross-unit third-order dependency modeling operation using the pre-defined machine learning model, it is necessary to extract the mined cross-group evolutionary relationships between units from the model. These relationships reflect the mutual relationships between each temporal unit and the preceding three and following three temporal units in terms of amplitude, frequency, and time. An initial set of relationships is generated, containing the evolutionary direction, rate, and cross-group transmission information of each temporal unit with other units. The evolutionary direction describes the direction of change of the temporal unit in terms of amplitude and frequency, the rate represents the speed of change, and the cross-group transmission information reflects the transmission and dependency relationships between temporal units.

[0136] Step S422: Analyze the cross-group transitivity of each association in the original set of associations, identify cross-group association paths that can be continuously transmitted between multiple groups of units, and generate a set of continuous cross-group association paths. The set of continuous cross-group association paths includes the unit transmission order and cooperative label of each path.

[0137] Cross-group transportability refers to whether a relationship can be continuously propagated across multiple time units; that is, whether a change in one time unit can affect other time units. The goal is to identify cross-group transportability paths that can propagate continuously across multiple time units. By analyzing the cross-group transportability of these relationships, paths that can form continuous transport between multiple time units are identified. A set of continuous cross-group transportability paths is generated, containing the unit transport order and coordination markers for each path. The unit transport order describes the sequential order of time units within the path, while the coordination markers identify the cooperative changes between time units within the path.

[0138] To perform analysis and identification operations, graph theory and path search algorithms can be used. For example, the original set of associations can be represented as a graph, where temporal units are nodes and associations are edges, and then depth-first search or breadth-first search algorithms can be used to find paths that can propagate continuously between multiple sets of units.

[0139] Step S423: Assign weights to each path in the set of continuous cross-group association paths. Different transmission weights are assigned based on the number of continuous cross-group transmissions of the association relationship in the path. The more transmissions, the higher the weight, thus generating a weighted cross-group association path.

[0140] In an optional implementation, step S423 may specifically include the following steps S4231 to S4236: Step S4231: Select the target cross-group association path from the set of continuous cross-group association paths, extract the number of association relationships contained in the path, the number of units continuously transmitted across groups and the co-evolution marker, and record the transmission length and co-evolution density of the path.

[0141] Extract the number of associations, the number of units continuously propagating across groups, and the co-evolutionary markers contained in the path. The number of associations reflects the extent of associations contained in the path; the number of units continuously propagating across groups indicates the number of temporal units that can propagate continuously through the path; and the co-evolutionary markers are used to identify the co-evolutionary relationships between temporal units in the path. Record the path's propagation length and co-evolutionary density. Propagation length is the propagation distance between temporal units in the path, while co-evolutionary density describes the tightness of co-evolutionary changes between temporal units in the path. To implement the extraction and recording operations, data extraction and recording algorithms can be used. For example, by programming, a target cross-group association path can be selected from a set of continuous cross-group association paths, and then the number of associations, the number of units continuously propagating across groups, and the co-evolutionary markers can be extracted from the path data, and the path's propagation length and co-evolutionary density can be recorded.

[0142] Step S4232: Based on the cross-group propagation law of the corresponding path in the DDR shaking prior, set the correspondence between the number of consecutive cross-group propagation and the weight. Each time the number of consecutive propagation increases, the weight is increased according to a preset ratio, and the weight assignment rule is generated.

[0143] The DDR jitter prior is derived from extensive DDR wafer testing experiments and theoretical analysis. It reflects the basic laws and characteristics of signal propagation across groups in a DDR chip under normal conditions. These laws and characteristics include the intrinsic relationship between propagation stability, propagation speed, propagation direction, and the number of consecutive cross-group propagations. When setting the correspondence between the number of consecutive cross-group propagations and the weight, different propagation paths in the DDR chip have different physical structures and electrical characteristics, which leads to differences in their cross-group propagation laws. For example, some paths may be more stable in signal propagation and have a stronger ability to propagate across groups due to their rational circuit layout; while other paths may be subject to more interference and have a relatively fewer consecutive propagation times. Through statistical analysis of a large amount of experimental data, the relationship between the number of consecutive cross-group propagations and factors such as propagation stability and speed under different paths can be summarized. Based on these relationships, the functional relationship between the number of consecutive propagations and the weight can be determined. For example, the more consecutive cross-group propagations, the better the propagation performance of the path and the greater its contribution to the overall signal cross-group evolution; therefore, it should be given a higher weight. A base weight value can be set, and the weight is the base value when the number of consecutive propagations is 1. As the number of consecutive conduction cycles increases, the weights are increased according to a preset ratio. Determining this preset ratio requires comprehensive consideration of the DDR chip's performance requirements and actual testing conditions. If the preset ratio is too large, the weight increase may be too drastic, resulting in excessively high weights for some paths and affecting the overall balance; if the preset ratio is too small, the weight differentiation will be unclear, failing to accurately reflect the differences in conduction performance between different paths.

[0144] For example, analysis of the physical priors of DDR jitter revealed a linear relationship between the increase in conduction stability and speed for a certain type of path as the number of consecutive cross-group conductions increases. Based on this, a linear weighting rule can be established: for every 1 consecutive conduction iteration, the weight increases by 1.1. That is, when the number of consecutive conductions is 1, the weight is 1; when the number of consecutive conductions is 2, the weight is 1 × 1.1 = 1.1; when the number of consecutive conductions is 3, the weight is 1.1 × 1.1 = 1.21, and so on.

[0145] In practical applications, adjustments are made based on the specific circumstances of different paths. For some special paths, non-linear weighting rules can be used. If the conduction performance of certain paths changes abruptly after the number of consecutive conductions reaches a certain threshold, a piecewise function can be used to set the weights. When the number of consecutive conductions is less than the threshold, the weights are increased according to one ratio; when the number of consecutive conductions is greater than or equal to the threshold, the weights are increased according to another ratio.

[0146] Step S4233: Calculate the number of consecutive cross-group propagation times for each relationship in the target cross-group association path, assign propagation weights to each relationship based on the weight assignment rules, and generate the weight distribution of a single path.

[0147] Associations reflect the transmission connections between time units, and the number of consecutive cross-group transmissions reflects the ability of this connection to persist across multiple time units. The continuity of each association during cross-group transmission can be recorded by traversing the target cross-group association path. Starting from the beginning of the path, each association is checked to see if it can continuously propagate from one time unit to the next. If it can, the number of consecutive cross-group transmissions is incremented by 1; if the transmission is interrupted, the counting restarts. For example, in a path consisting of multiple time units, if an association starts from the first time unit and propagates sequentially to the second and third time units without interruption, then the number of consecutive cross-group transmissions for this association is 3.

[0148] After obtaining the number of consecutive cross-group propagation times for each association, a propagation weight is assigned to each association based on the weight assignment rule generated in step S4232. The weight assignment rule clarifies the correspondence between the number of consecutive cross-group propagation times and the weight. According to this rule, the corresponding propagation weight can be calculated by substituting the number of consecutive cross-group propagation times for each association. Assuming the weight assignment rule is that the weight increases by 1.1 for every 1 increase in the number of consecutive propagation times, when the number of consecutive cross-group propagation times for a certain association is 2, its propagation weight is the base weight value multiplied by 1.1 to the power of 2.

[0149] After assigning propagation weights to each relationship in the target cross-group association path, these propagation weights are organized and recorded to generate a weight distribution for a single path. This weight distribution details the propagation weight of each relationship in the path.

[0150] Step S4234: Accumulate the transmission weights of all relationships in a single path to generate the total transmission weight of the target cross-group associated path. The total weight reflects the stability and synergistic consistency of the cross-group continuous transmission of the path.

[0151] Step S4235: Repeat the above assignment process for all target cross-group association paths to generate the total transmission weight of each path and the transmission weight of a single association relationship, and generate a weighted cross-group association path fragment.

[0152] Step S4236: Integrate all weighted cross-group association path segments in path order to generate a weighted cross-group association path set, which includes the transmission weight distribution, total weight and co-evolutionary label information of each path.

[0153] After generating weighted cross-group association path fragments, they are integrated in path order. The path order reflects the transmission order of the path between time units, and integrating them in path order ensures that the generated set of weighted cross-group association paths accurately reflects the transmission relationships between the paths.

[0154] A weighted set of cross-group association paths is generated, which includes the transmission weight distribution, total weight, and co-evolutionary marker information for each path. The transmission weight distribution describes the transmission weight of each association in the path, the total weight reflects the cross-group continuous transmission stability and co-evolutionary consistency of the path, and the co-evolutionary marker is used to identify the co-evolutionary relationships between temporal units in the path.

[0155] Step S424: Identify path segments in the weighted cross-group association path where the transmission weights exhibit a third-order abrupt change, and mark them as anomalous weight transmission segments. The transmission weights of the association relationships within these segments are discontinuous and differ from those in the normal path.

[0156] A third-order mutation occurs when the transmission weights change significantly and discontinuously at a certain position. Such a mutation may indicate an anomaly in the path.

[0157] Identifying path segments in weighted cross-group association paths where the transmission weights exhibit third-order abrupt changes is achieved by performing differential calculations and analyses on the transmission weights in the weighted cross-group association paths to find those path segments with large rates of change of the rate of change of the transmission weights (i.e., third-order differences).

[0158] Marking segments as anomalous weight transmission segments involves explicitly identifying these path segments exhibiting third-order abrupt changes. The transmission weights of the relationships within these segments differ discontinuously from those in normal paths. This discontinuity indicates that the transmission weight characteristics of these segments differ from those of normal paths, potentially suggesting anomalies. By marking these anomalous weight transmission segments, we can provide targets for further analysis and processing.

[0159] To enable identification and labeling operations, differential calculation and anomaly detection algorithms can be used. For example, third-order difference calculations can be performed on the propagation weights in weighted cross-group association paths through programming. Then, based on a preset threshold, it can be determined which path segments have third-order differences exceeding the threshold, and these path segments can be marked as anomalous weight propagation segments.

[0160] Step S425: Correct the weight of the abnormal weight transmission segment. Based on the weight distribution of the adjacent normal weighted cross-group association path, correct the transmission weight of the abnormal segment and generate the corrected weighted cross-group association path.

[0161] Weight correction aims to eliminate third-order abrupt changes in the transmission weights of anomalous weighted segments, making the transmission weights of the path more consistent with normal variation patterns. The transmission weights of anomalous segments are corrected based on the weight distribution of adjacent normal weighted cross-group association paths. The weight distribution of adjacent normal weighted cross-group association paths reflects the variation pattern of transmission weights under normal conditions. By referring to these normal weight distributions, the transmission weights of anomalous segments can be adjusted. For example, interpolation or fitting methods can be used to estimate the transmission weights of anomalous segments based on the weight distribution of adjacent normal paths.

[0162] The generation of the corrected weighted cross-group association path involves adjusting the propagation weights of the outlier segments and merging the corrected path with the original normal path to form a new weighted cross-group association path. This new path has been corrected in the outlier segment portion, thus better reflecting the true propagation weights of the path.

[0163] To implement weight correction operations, interpolation, fitting, and data update algorithms can be used. For example, by programming linear interpolation or polynomial fitting, the propagation weights of outlier segments can be corrected based on the weight distribution of adjacent normal weighted cross-group association paths, and then the data of the weighted cross-group association paths can be updated.

[0164] Step S426: Integrate all the modified weighted cross-group association paths into a unit cross-group dependency association set, which contains the modified cross-group evolutionary association relationships and transmission weight information.

[0165] The intergroup dependency association set contains the modified intergroup evolutionary association and transmission weight information. The intergroup evolutionary association describes the mutual relationship between temporal units in terms of amplitude, frequency and time, while the transmission weight information reflects the importance and stability of the path in intergroup transmission.

[0166] Step S430: Perform feature transformation on the set of cross-group dependency associations of units, and convert each cross-group evolutionary association into a pattern fragment containing cross-group collaborative information. Each fragment corresponds to a cross-unit collaborative evolutionary association form.

[0167] Feature transformation is the process of converting cross-group evolutionary relationships in a set of cross-group dependency relationships from one form to another form that is easier to analyze and match.

[0168] Each cross-group evolutionary association is converted into a pattern fragment containing cross-group coordination information, which is the information on the coordinated changes in amplitude, frequency, and time between temporal units. Each pattern fragment corresponds to a cross-unit coordinated evolutionary association pattern, which can be a defined amplitude-frequency coordinated change pattern, a temporal transmission pattern, etc.

[0169] Step S440: Perform cross-group collaborative matching between the mode fragment and the modes in the preset DDR wafer test abnormal mode library, compare the consistency of the timing evolution trajectory of the mode fragment and the first three groups and the last three groups of the modes in the library, and generate a cross-group matching result list.

[0170] By comparing the temporal evolution trajectories of the pattern fragment with the first three and last three groups of patterns in the library, we can determine whether the pattern fragment matches the patterns in the library. The temporal evolution trajectory describes the changes of the pattern over time.

[0171] A cross-group matching result list is generated, which records the matching status of pattern fragments with patterns in the library, including information such as the matched pattern pairs and the degree of matching. By analyzing this list, abnormal patterns that match the pattern fragments can be identified, thereby determining potential anomalies in DDR wafer testing.

[0172] Step S450: Filter the list of cross-group matching results, retain the pattern pairs with completely consistent cross-group evolutionary trajectories, generate a set of cross-group matching pattern pairs, and generate an identifier for no cross-group matching patterns if the set is empty.

[0173] Pattern pairs with completely identical cross-group evolutionary trajectories are retained. By comparing the pattern pairs in the cross-group matching result list, only those pattern pairs with completely identical cross-group evolutionary trajectories are retained.

[0174] Generate a set of cross-group matching pattern pairs, which contains all pattern pairs with completely identical cross-group evolutionary trajectories. If the set is empty, it means that no abnormal pattern with completely identical cross-group evolutionary trajectory was found, and an identifier for no cross-group matching pattern is generated.

[0175] To perform filtering and set generation operations, data filtering and label generation algorithms can be used. For example, by programming, the list of cross-group matching results can be traversed, the cross-group evolution trajectory of each pattern pair can be compared, and pattern pairs with completely identical cross-group evolution trajectories can be added to the set. If the set is empty, a label for a pattern with no cross-group matching can be generated.

[0176] Step S460: Integrate the identifiers of cross-group matching pattern pairs or no cross-group matching patterns into anomaly pattern association results. The anomaly pattern association results include the matched cross-group anomaly evolution patterns and the corresponding multi-level collaborative association details.

[0177] The anomaly pattern association results include the matched cross-group anomaly evolution patterns and the corresponding multi-level co-association details. If a set of cross-group matching pattern pairs exists, the pattern pairs represent the matched cross-group anomaly evolution patterns, and the corresponding multi-level co-association details describe the co-variation relationships of these patterns in terms of amplitude, frequency, and time. If there is no identifier for a cross-group matching pattern, it means that no matching anomaly pattern was found.

[0178] Step S500: Output the DDR wafer test anomaly detection result based on the anomaly mode association result, generate the corresponding test setting adjustment instruction, and send the test setting adjustment instruction to the DDR wafer test equipment to perform test setting calibration operation.

[0179] In an optional implementation, step S500 may specifically include the following steps S510 to S560: Step S510: Analyze the abnormal pattern association results, extract the identifiers of the matched cross-group abnormal evolution patterns or those without cross-group matching patterns, and generate a set of core information on abnormal evolution. The set of core information on abnormal evolution includes the cross-group evolution trajectory, transmission path, and collaborative unit information of the abnormality.

[0180] The results of anomaly pattern association may exist in the form of a defined data table, database records, or structured files. These files contain matching information between pattern fragments and patterns in a predefined anomaly pattern library, such as matched pattern pairs, degree of matching, and temporal evolution trajectory. Data parsing tools and programming languages ​​can be used to read and analyze the anomaly pattern association results row by row and field by field, according to predefined data format rules.

[0181] During the analysis, we focus on whether there are matching cross-group anomalous evolution patterns. If matching patterns exist, these patterns are extracted from the anomalous pattern association results. These cross-group anomalous evolution patterns describe in detail the anomalous amplitude-frequency coordinated changes occurring across multiple time units, and the temporal and spatial evolution of these changes. For example, some patterns may show that in a given combination of time units, the amplitude suddenly increases or decreases, or the frequency exhibits anomalous fluctuations.

[0182] If no abnormal pattern is matched, extract the identifier for the pattern that does not match across groups. This identifier can be a code, symbol, or text message that explicitly indicates that no case matching the preset abnormal pattern was found in this test.

[0183] After extracting the identifiers of matching cross-group anomaly evolution patterns or those without matching cross-group patterns, further mining is performed to uncover the anomaly's cross-group evolution trajectory, propagation path, and cooperating unit information. The cross-group evolution trajectory describes the development process of the anomaly across multiple timing units, including the trends and rates of change in amplitude and frequency. By analyzing the timing evolution trajectory data in the matching patterns, the specific manifestations and variation patterns of the anomaly in different timing units can be extracted. The propagation path reflects the propagation route of the anomaly signal in the DDR wafer circuit. By analyzing the nodes and relationships involved in the pattern, the propagation process of the anomaly signal from the starting point to the ending point can be traced. Cooperating unit information records the timing units participating in the anomaly's cooperating evolution, as well as their interactions and relationships. By analyzing the cooperating markers and timing correlation information in the pattern, it can be determined which timing units play a key role in the anomaly evolution process and how they cooperate and change.

[0184] The extracted cross-group evolutionary trajectories, transmission paths, and cooperative unit information of the anomalies are sorted and integrated to generate a core information set of anomaly evolution.

[0185] Step S520: Map the core information set of abnormal evolution to the preset DDR wafer test abnormal-calibration correlation map to generate the corresponding test setting cross-group calibration requirements. The requirements include the cross-group timing nodes of calibration, evolution direction and collaborative adjustment correlation.

[0186] The preset DDR wafer test anomaly-calibration correlation map is pre-established, which records the relationship between different types of anomalies and the corresponding test setting calibration requirements.

[0187] Through mapping operations, based on the characteristics and types of anomalies in the core information set of anomaly evolution, the corresponding test setting calibration requirements are identified from the correlation graph. These requirements include cross-group time-series nodes for calibration, evolution directions, and collaborative adjustment correlations. Cross-group time-series nodes are the locations of time-series units that need calibration, evolution directions describe the evolution trend of the anomaly, and collaborative adjustment correlations reflect the collaborative adjustment relationships between different time-series units.

[0188] To implement mapping and requirement generation operations, data mapping and requirement generation algorithms can be used. For example, by programming, the abnormal features in the core information set of abnormal evolution can be matched with the abnormal types in the correlation graph to find the corresponding calibration requirements, and then these requirements can be organized into cross-group calibration requirements for test settings.

[0189] Step S530: Integrate the core information set of anomaly evolution with cross-group calibration requirements into DDR wafer test anomaly detection results. The results include details of the anomaly's cross-group evolution, calibration requirements, and generation time.

[0190] The core information set on anomaly evolution records in detail the specific characteristics of anomalies that occur during DDR wafer testing, such as the cross-group evolution trajectory, conduction path, and cooperative unit information of the anomaly; the cross-group calibration requirements are derived from the analysis of the anomaly and are specific requirements for adjusting test settings to eliminate or reduce the anomaly.

[0191] Integrating these two parts into anomaly detection results requires a deep understanding and analysis of the core information set of anomaly evolution and cross-group calibration requirements. The core information set of anomaly evolution provides detailed information about the occurrence of the anomaly, while the cross-group calibration requirements clarify the direction and specific content of adjustments needed to address the anomaly. First, these two parts of information are categorized and organized, and related information is correlated and matched. For example, the problematic time-series unit in the anomaly evolution trajectory is mapped to the adjustment requirements for that time-series unit in the cross-group calibration requirements. The cross-group evolution details describe the specific development process of the anomaly across multiple time-series units, including anomalous changes in amplitude and frequency, the speed of change, and the direction. Simultaneously, the cross-group calibration requirements are also recorded in detail in the detection results, including the cross-group time-series nodes to be calibrated, the evolution direction, and the correlation with collaborative adjustments.

[0192] The generation time records the exact moment the test result was generated, which facilitates tracking the development of anomalies and evaluating the effectiveness of calibration measures. The generation time can be recorded using system time or a predefined timestamp and accurately added to the test result.

[0193] Step S540: Based on the cross-group calibration requirements, generate a draft cross-group timing calibration instruction for the test settings. The draft includes the adjustment content of each calibration node, the cross-group timing sequence, and the collaborative constraints of the associated test units.

[0194] The draft includes the adjustment details for each calibration node, the cross-group timing sequence, and the coordination constraints of associated test units. The adjustment details for each calibration node describe the specific adjustments required for the test settings of that node. The cross-group timing sequence specifies the execution order of calibration operations among multiple timing units, and the coordination constraints of associated test units reflect the coordination relationship between different test units during the calibration process.

[0195] To achieve the instruction draft generation operation, instruction generation and constraint addition algorithms can be used. For example, by programming, based on information such as calibration nodes, evolution direction, and collaborative adjustment associations in cross-group calibration requirements, the adjustment content of each calibration node can be generated, the cross-group timing sequence can be determined, and collaborative constraints of associated test units can be added to form a cross-group timing calibration instruction draft for test settings.

[0196] Step S550: Conduct cross-group evolution conflict investigation on the draft cross-group timing calibration instructions, analyze whether the cross-group timing adjustment of the calibration instructions conflicts with the cross-group evolution trajectory of the current test process, and investigate whether the adjustment will trigger new cross-group abnormal evolution.

[0197] In an optional implementation, step S550 may specifically include the following steps S551 to S556: Step S551: Extract the adjustment content, cross-group timing order and coordination constraints of all calibration nodes from the cross-group timing calibration instruction draft, and generate a calibration node adjustment sequence. The sequence includes the adjustment direction, cross-group association and coordination mark of each calibration node.

[0198] The adjustment content of the calibration node describes the specific adjustment operations that need to be made to the test settings of the node. The cross-group timing sequence specifies the execution order of the calibration operation among multiple timing units. The coordination constraint reflects the coordination relationship between different test units in the calibration process.

[0199] Generate a calibration node adjustment sequence, which includes the adjustment direction, cross-group association, and coordination marker for each calibration node. The adjustment direction describes whether the calibration node is adjusted to increase or decrease the value of the test setting item. The cross-group association indicates the relationship between the calibration node and other timing units. The coordination marker is used to identify the role of the calibration node in the coordinated adjustment process.

[0200] To perform the extraction and sequence generation operations, data extraction and sequence construction algorithms can be used. For example, the adjustment content of calibration nodes, cross-group timing order, and cooperative constraints can be extracted from a cross-group time-series calibration instruction draft by programming, and then this information can be organized into a calibration node adjustment sequence.

[0201] Step S552: Obtain real-time cross-group evolution trajectory data of the current test process from the DDR wafer test equipment, extract the cross-group evolution direction, rate and conduction path of the test unit corresponding to the calibration node, and generate a set of real-time cross-group evolution trajectories.

[0202] In an optional implementation, step S552 may specifically include the following steps S5521 to S5526: Step S5521: Generate a real-time cross-group evolution trajectory query request compatible with the DDR wafer test equipment communication protocol. The request content includes the cross-group evolution direction, rate and conduction path query identifier of the test unit, as well as the cross-group association mark of the corresponding calibration node.

[0203] To obtain real-time cross-group evolution trajectory data of the current test process from the DDR wafer test equipment, a real-time cross-group evolution trajectory query request compatible with the equipment's communication protocol must first be generated. The communication protocol is the rule and standard for communication between the equipment and external systems, ensuring that the request can be correctly received and processed by the equipment.

[0204] The request includes cross-group evolution direction, rate, and conduction path query identifiers for the test units. These identifiers specify the exact content of the query. The cross-group association markers for the corresponding calibration nodes identify the test units associated with the calibration nodes, enabling the equipment to accurately provide relevant evolution trajectory data.

[0205] To implement the request generation operation, protocol encapsulation and data encoding algorithms can be used. For example, according to the communication protocol requirements of DDR wafer test equipment, the cross-group evolution direction, rate and conduction path query identifier of the test unit, as well as the cross-group association mark of the corresponding calibration node, are encapsulated and encoded to generate a query request that conforms to the protocol requirements.

[0206] Step S5522: Send the real-time cross-group evolution trajectory query request to the DDR wafer test equipment through a preset bidirectional communication channel, and record the sending time, unique identifier and calibration node association mark of the request for subsequent response matching and data alignment.

[0207] The system records the request's sending time, unique identifier, and calibration node association tag. The sending time records when the request was sent, the unique identifier distinguishes different requests, and the calibration node association tag identifies the relationship between the request and the calibration node. This information is crucial for subsequent response matching and data alignment.

[0208] To enable request sending and information recording operations, communication sending and information recording algorithms can be used. For example, real-time cross-group evolution trajectory query requests can be sent to the device via a bidirectional communication channel through programming, and the sending time, unique identifier, and calibration node association marker of the request can be recorded.

[0209] Step S5523: Monitor the response information of the two-way communication channel. If the response data returned by the device is received within a preset time, extract the real-time cross-group evolution trajectory content from the response data and generate the original trajectory data set, which contains the cross-group association information of the test unit.

[0210] Step S5524: Perform cross-group timing alignment processing on the original trajectory data set, match the timestamps of the trajectory data with the three-cycle cross-group clock sequence of the current test process, and generate an aligned cross-group trajectory data set.

[0211] The timestamps of the trajectory data are matched with the three-cycle cross-group clock sequence of the current test process. The timestamps record the acquisition time of the trajectory data, and the three-cycle cross-group clock sequence serves as the clock reference sequence for the current test process. By matching, the time of the trajectory data can be aligned with the clock sequence.

[0212] Generate an aligned cross-group trajectory data set. The trajectory data in this set is consistent with the clock sequence of the current test process in time, which makes it easier to analyze the cross-group evolution of the test unit.

[0213] To achieve cross-group timing alignment, the timestamps in the original trajectory data set can be compared with the three-cycle cross-group clock sequence. The trajectory data can be adjusted based on the comparison results to generate an aligned cross-group trajectory data set.

[0214] Step S5525: Classify the aligned cross-group trajectory data set according to test units. Each test unit corresponds to a set of cross-group evolution trajectory data, generating a classified real-time cross-group evolution trajectory set.

[0215] Each test unit corresponds to a set of cross-group evolutionary trajectory data. Through classification, trajectory data belonging to the same test unit can be grouped together. A classified real-time cross-group evolutionary trajectory set is generated, containing the cross-group evolutionary trajectory data for each test unit, thus more clearly demonstrating the evolution of each test unit. For example, the aligned cross-group trajectory data set can be classified according to the test unit's identifier, adding trajectory data belonging to the same test unit to the same set, generating a classified real-time cross-group evolutionary trajectory set.

[0216] Step S5526: Perform an integrity check on the classified real-time cross-group evolution trajectory set to confirm that there is no missing cross-group evolution direction, rate and propagation path information for each group of trajectory data.

[0217] It is confirmed that the cross-group evolution direction, rate, and propagation path information of each trajectory data set is complete. By checking the completeness of this information, potential problems in the data can be identified. If missing information is found, further processing is required, such as supplementing data or re-collecting data. For example, the real-time cross-group evolution trajectory set after classification is traversed to check whether the cross-group evolution direction, rate, and propagation path information of each trajectory data set is complete.

[0218] Step S553: ​​Perform cross-group evolution comparison between each calibration node in the calibration node adjustment sequence and the corresponding test unit in the real-time cross-group evolution trajectory set, and analyze the third-order difference between the adjusted cross-group evolution trajectory and the current evolution trajectory.

[0219] Analyzing the third-order difference between the adjusted cross-group evolutionary trajectory and the current evolutionary trajectory reveals that the third-order difference can more sensitively capture changes in trajectory, i.e., the acceleration of those changes. By analyzing the third-order difference, it can be determined whether the adjustment of the calibration node will cause significant changes in the cross-group evolutionary trajectory of the test unit.

[0220] To enable comparison and analysis, data alignment and difference calculation algorithms can be used. For example, each calibration node in the calibration node adjustment sequence can be compared with the corresponding test unit in the real-time cross-group evolutionary trajectory set by programming, and the third-order difference between the adjusted cross-group evolutionary trajectory and the current evolutionary trajectory can be calculated.

[0221] Step S554: Identify calibration nodes in the third-order differential that cause the cross-group evolution direction of the test unit to deviate from the normal range, mark them as conflict calibration nodes, and adjust the nodes to trigger new cross-group abnormal evolution risks.

[0222] For example, a threshold judgment and marking algorithm can be used. By setting a threshold for third-order difference, when the third-order difference of a calibration node exceeds this threshold, the node is considered to cause the cross-group evolution direction of the test unit to deviate from the normal range, and it is marked as a conflict calibration node.

[0223] Step S555: Adjust the conflict calibration node adjustment scheme. Based on the third-order transmission relationship of the current real-time cross-group evolution trajectory, adjust the adjustment direction, magnitude and cross-group correlation to generate the corrected calibration node adjustment sequence.

[0224] Corrections to the adjustment direction can be made by referring to the third-order transmission relationship of the real-time cross-group evolution trajectory to determine whether the current adjustment direction will cause the cross-group evolution direction of the test unit to deviate further from the normal range. If so, it needs to be adjusted to the opposite direction or a more subtle adjustment needs to be made. For example, if the real-time cross-group evolution trajectory shows that the evolution direction of a certain test unit has already shifted in a certain direction, and the adjustment direction of the original calibration node will exacerbate this shift, then it is necessary to reverse the adjustment direction or reduce the adjustment magnitude.

[0225] The adjustment magnitude also depends on the third-order transmission relationship. By analyzing the rate and trend of change in real-time cross-group evolutionary trajectories, an appropriate adjustment magnitude can be determined. If the third-order transmission relationship shows that the evolutionary changes of the test unit are relatively drastic, then the adjustment magnitude should be relatively small to avoid over-adjustment; conversely, if the changes are relatively gradual, the adjustment magnitude can be appropriately increased.

[0226] Correcting cross-group associations requires considering the collaborative relationships between calibration nodes and other test units. The third-order transmission relationship of real-time cross-group evolution trajectories can reveal the mutual influence and transmission patterns between different test units. Based on these patterns, the cross-group associations of calibration nodes and other test units are adjusted to ensure that calibration operations do not disrupt the collaborative relationships between test units.

[0227] Generating a revised calibration node adjustment sequence involves updating the original calibration node adjustment sequence with the revised adjustment direction, magnitude, and cross-group correlation information. This revised sequence will serve as the basis for subsequent calibration operations, more effectively avoiding cross-group evolution conflicts and improving the accuracy and stability of the test.

[0228] Step S556: Integrate the revised calibration node adjustment sequence with other contents of the original instruction draft to generate a revised cross-group timing calibration instruction draft. The cross-group timing adjustment in the draft has no third-order conflict with the current test cross-group evolution trajectory.

[0229] The purpose of integration is to incorporate the revised calibration node adjustment scheme into the original draft instructions, forming a complete draft of cross-group timing calibration instructions without third-order conflicts. This integration ensures that the calibration instructions will not conflict with the current cross-group evolution trajectory during implementation, thus guaranteeing the smooth progress of testing.

[0230] A revised draft of cross-group timing calibration instructions has been generated. The cross-group timing adjustments in this draft have been revised and do not conflict with the current cross-group evolution trajectory. This will more accurately guide the DDR wafer test equipment to perform test setting calibration operations, thereby improving the accuracy and reliability of the test.

[0231] To facilitate integration, data merging and sorting algorithms can be used. For example, the revised calibration node adjustment sequence can be merged with other parts of the original draft directive to ensure consistency and completeness of information across different sections.

[0232] Step S560: Convert the approved cross-group timing calibration instruction draft into an instruction format compatible with the DDR wafer test equipment, and generate test setting adjustment instructions. The instructions contain timing nodes for cross-group calibration and related information for collaborative adjustment.

[0233] DDR wafer testing equipment has its own instruction format standard, which differs from the original form of the cross-group timing calibration instruction draft, so a conversion operation is required.

[0234] The draft cross-group timing calibration instructions are generated based on the analysis results of the preceding processes and include detailed calibration schemes, such as the specific adjustments for each calibration node and the collaborative constraints between test units. However, this information is usually recorded in a data structure that is easy for humans to understand and analyze, such as tables or lists. But DDR wafer testing equipment cannot directly recognize and execute such a draft; the equipment can only accept instructions in a format that conforms to its internal communication protocols and operating specifications.

[0235] The instruction format requirements for DDR wafer testing equipment can be determined by consulting the equipment's technical documentation. Then, the information in the draft cross-group timing calibration instructions should be organized and filtered. Based on the equipment instruction format requirements, necessary data items should be extracted, such as timing nodes for cross-group calibration, specific adjustment parameters, and collaborative adjustment information. After extracting key information, this information should be encoded according to the equipment's encoding rules. Different data items may require different encoding methods. For timing node information, it may be necessary to convert it to a timestamp format that the equipment can recognize; for adjustment parameters, it may be necessary to convert them to binary or hexadecimal encoding to meet the equipment's requirements for data accuracy and transmission efficiency.

[0236] In addition, necessary header and footer information can be added to the instruction to conform to the device's instruction structure. The instruction header typically contains information such as the instruction type identifier and version number, which the device uses to determine the instruction's purpose and format version. The instruction footer may contain information such as a checksum to ensure that no errors occur during transmission. After encoding and adding the necessary information, the test setting adjustment instruction is generated. Finally, this instruction is sent to the DDR wafer testing equipment so that the equipment can calibrate the test settings according to the instruction content, thereby correcting any anomalies that may occur during testing and improving the accuracy and reliability of the test.

[0237] Please see Figure 2 , Figure 2This is a schematic diagram of a computer device provided in an embodiment of the present invention. The computer device includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer device, capable of parsing various instructions and processing various data within the computer device. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer device. The memory 103 is a storage device in the computer device used to store programs and data. It is understood that the memory 103 here can include the computer device's built-in memory, or it can include extended memory supported by the computer device. The memory 103 provides storage space, which stores the computer device's operating system; this invention does not limit the storage space.

[0238] In one embodiment, the processor 101 executes the machine learning-based DDR wafer test anomaly detection method provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A machine learning-based method for detecting anomalies in DDR wafer testing, characterized in that, The method includes: Acquire timing eye diagrams and timing waveform data related to total jitter, deterministic jitter, and random jitter synchronously collected during DDR wafer testing. The timing waveform data includes a sequence of continuously sampled differential signal sampling points with timestamps. The differential timing sampling points in the timing waveform data are aligned by clock edge phase segmentation to construct phase domain dense sampling data. Adaptive sparse coding is completed by combining DDR jitter physical prior, and noise and normal process fluctuation related components are removed to obtain defect-sensitive sparse components. Variational mode decomposition is performed on the defect-sensitive sparse components to separate the components related to process drift, normal process fluctuations and abrupt anomalies, and the abrupt anomaly-related components are extracted as target detection data. The target detection data is input into a preset machine learning model, and anomaly pattern association results are generated through correlation modeling and anomaly pattern matching of the target detection data. Based on the abnormal mode association results, the DDR wafer test abnormality detection results are output, and corresponding test setting adjustment instructions are generated. The test setting adjustment instructions are sent to the DDR wafer test equipment to perform test setting calibration operations.

2. The method according to claim 1, characterized in that, The differential timing sampling points in the timing waveform data are aligned by clock edge phase segmentation to construct dense sampling data in the phase domain. Adaptive sparse coding is then performed using DDR jitter priors to remove noise and normal process fluctuation-related components, resulting in defect-sensitive sparse components, including: The clock signal phase nodes and differential timing sampling points in the timing waveform data are extracted from the timing waveform data. The conduction dependency path of each differential timing sampling point in the clock phase advancement process is tracked, and its conduction correlation range with phase change is analyzed to generate a sampling point-phase conduction dependency correlation table, which includes the conduction start phase, dependency node range and conduction path characteristics of each sampling point. Based on the sampling point-phase conduction dependency association table, sampling points with the same origin conduction start phase and continuous dependency path are connected in phase progression order to form phase conduction dependency links. Each link corresponds to a set of continuous clock phase conduction processes, and the conduction timing and dependency relationship of the sampling points in the link are preserved. The location of the conduction path break in each phase conduction dependent link is filled in. The conduction properties of the break location are deduced based on the dependency rules of adjacent conduction segments in the link, and a complete phase conduction dependent link is generated to fill the conduction information loss caused by the sampling gap. All complete phase-transduction dependent links are compared with the conduction path rules of the DDR jitter physical prior, and link segments whose conduction paths do not conform to the DDR jitter physical prior rules are identified, generating a set of abnormal conduction dependent segments. The abnormal conduction dependency fragment set is encoded and converted into a conduction dependency coding sequence containing phase dependency relationship, resulting in a set of conduction dependency coding sequences. Each coding unit in the conduction dependency coding sequence corresponds to a conduction node and dependency path information in the link. The conduction-dependent coding sequence corresponding to the normal test state of the DDR wafer is removed from the set of conduction-dependent coding sequences. The remaining coding sequences are retained and integrated into defect-sensitive sparse components. Each retained coding sequence corresponds to phase conduction information that deviates from the physical prior.

3. The method according to claim 2, characterized in that, The step of connecting sampling points with the same origin propagation initiation phase and continuous dependency path into a phase propagation dependency link based on the sampling point-phase conduction dependency association table includes: Extract the conduction initiation phase and dependency path features of each sampling point from the sampling point-phase conduction dependency association table to obtain the conduction homology identification imprint and generate the phase conduction homology dependency imprint sequence, which contains the conduction homology identifier and dependency path feature description of each sampling point; The phase conduction homology dependency imprint sequence is subjected to homology dependency clustering. Imprints with the same conduction initiation phase and similar dependency path characteristics are grouped together to generate homology conduction dependency cluster groups. Each group corresponds to a continuous phase conduction process. The dependency paths of the sampling points in each homologous conduction dependency cluster are fitted to generate a conduction trajectory reference sequence of the sampling points in the group. The conduction trajectory reference sequence retains a unified description of the conduction timing and dependency relationship of all sampling points in the group. The location where the dependent path breaks in the conduction trajectory reference sequence is identified and marked as a conduction trajectory fault. The fault location corresponds to the area of ​​missing conduction information caused by the sampling gap. The fault location of the conduction trajectory is pre-completed, and the basic conduction properties of the fault location are deduced based on the dependency rules of adjacent normal conduction segments within the group, generating a reference sequence of fault pre-completed conduction trajectory. All fault pre-complete conduction trajectory reference sequences are integrated according to the progression order of the conduction initiation phase to generate phase conduction dependent links, each link corresponding to a set of continuous and complete phase conduction processes.

4. The method according to claim 3, characterized in that, The pre-completion of the conduction trajectory fault location, based on the dependency relationship between adjacent normal conduction segments within the group, derives the basic conduction properties of the fault location, generating a fault pre-completion conduction trajectory reference sequence, including: Extract the normal conduction segment dependency features on both sides of the fault from the conduction trajectory reference sequence corresponding to the conduction trajectory fault, and generate a set of conduction dependency features on both sides of the fault. Continuous dependency interpolation is performed on the set of conduction dependency features on both sides of the fault. Based on the transition law of the features on both sides, a conduction dependency interpolation sequence from one side to the other is generated. The conduction dependency interpolation sequence contains the complete process of continuous change of dependency features. The conduction-dependent interpolation sequence is matched with the conduction law of the DDR jitter physical prior, and the characteristic parameters of the interpolation sequence are adjusted to make it perfectly match the DDR jitter physical characteristics to generate a calibrated conduction-dependent interpolation sequence. Based on the calibrated conduction-dependent interpolation sequence, the conduction properties of each phase point at the fault location are derived, and a set of conduction-dependent properties at the fault location is generated. The set of conduction-dependent attributes at the fault location is inserted into the corresponding conduction trajectory reference sequence to fill the missing conduction information at the fault location and generate the conduction trajectory reference sequence with the inserted fault. The conduction timing coherence of the inserted fault conduction trajectory reference sequence is verified to ensure that the conduction timing of the inserted sequence is consistent with that of the original reference sequence, thereby generating a fault pre-complete conduction trajectory reference sequence.

5. The method according to claim 1, characterized in that, The variational mode decomposition of the defect-sensitive sparse components separates components related to process drift, normal process fluctuations, and abrupt anomalies. The components related to abrupt anomalies are then extracted as target detection data, including: The defect-sensitive sparse component is divided into multiple time segments according to three consecutive clock cycles. The amplitude-frequency co-evolution trend of the signal in each time segment is analyzed to generate an amplitude-frequency co-evolution sequence. The sequence contains the amplitude change rate and frequency propagation trajectory of each time segment. A third-order dependency analysis is performed on each trend segment in the amplitude-frequency co-evolution sequence to identify trend segment groups with continuous co-evolutionary associations. Each group of trend segments covers the co-evolution process of three time segments, and the temporal order and co-evolutionary association of the trend segments within the group are preserved. Extract the cross-period evolution features of each trend segment group, deduce the continuous evolution trajectory across time periods based on the amplitude-frequency co-evolution rate within the trend segment, and generate a set of cross-period evolution trajectories, which contains the amplitude-frequency co-change path of each trend segment group. The set of cross-time evolution trajectories is compared with the preset cross-time evolution trajectories corresponding to process drift and normal process fluctuations, and the trend segment groups belonging to the two types of trajectories are marked. Remove the marked trend segment groups from the cross-time evolution trajectory set, retain the time series corresponding to the remaining trend segment groups, convert the remaining time series into time domain signals, and generate a mutation anomaly co-evolution component. All the aforementioned mutational anomaly co-evolutionary components are integrated into mutational anomaly-related components in their original temporal order, which serve as target detection data. This data includes cross-time-period anomaly co-evolutionary information in both amplitude and frequency dimensions.

6. The method according to claim 5, characterized in that, The analysis examines the amplitude-frequency co-evolution trend of signals within each time segment, generating an amplitude-frequency co-evolution sequence, including: The amplitude and corresponding frequency values ​​of all sampling points are extracted from each three-period time series of the defect-sensitive sparse component, and the time series amplitude-frequency subsequence is generated by arranging them in the sampling order. The time series amplitude-frequency subsequence retains the phase correlation and co-evolution markers of the sampling points. The correlation between each sampling point in the amplitude-frequency subsequence of the time segment and the amplitude-frequency changes of the preceding three and following three sampling points is analyzed. The continuous change slope of the amplitude and the continuous conduction rate of the frequency are calculated to generate a slope-rate change sequence. The slope-rate change sequence contains dual change information for each sampling point. The sampling point segments in the slope-rate change sequence with continuous slope and rate change are connected to form a co-evolution sub-chain. Each sub-chain corresponds to a change process with amplitude-frequency co-stable, and the phase, slope and rate correlation of the sampling points in the sub-chain are preserved. Identify sampling points in the co-evolutionary subchain where the slope or rate changes exhibit third-order abrupt changes, and mark them as anomalous co-evolutionary segments. The amplitude-frequency co-evolutionary changes within these segments are discontinuous and differ from those in adjacent subchains. The abnormal co-evolution segment is fitted with a co-evolution trend. Based on the slope and rate of the adjacent normal co-evolution sub-chain, the modified co-evolution sub-chain is fitted with the modified co-evolution trend of the abnormal segment, and the modified co-evolution sub-chain is generated. All the corrected co-evolutionary subchains are integrated in the sampling order to generate an amplitude-frequency co-evolutionary sequence for each time segment. The sequence contains the complete amplitude-frequency co-evolution trend within the time segment and the corrected anomalous segment information.

7. The method according to claim 6, characterized in that, The identification of sampling points in the co-evolutionary subchain where the slope or rate change exhibits a third-order abrupt change includes: Slope-rate change sequences are extracted from the co-evolutionary subchain and arranged in the sampling order to generate a dual change sequence. The dual change sequence contains the amplitude change slope, frequency conduction rate and phase correlation information of each sampling point. Calculate the slope difference and rate difference between each sampling point and the previous three and the last three sampling points in the dual change sequence to generate a third-order difference change sequence. The third-order difference change sequence contains the third-order slope difference and third-order rate difference information of each sampling point. Each slope difference in the third-order differential change sequence is compared with the range of third-order slope differences under normal DDR wafer testing conditions. At the same time, each rate difference is compared with the range of third-order rate differences under normal conditions to confirm whether each difference is within the corresponding normal coordinated change range. The sampling points corresponding to the differences that exceed the normal range are marked with abrupt changes. The marking content includes the amplitude, direction and corresponding phase position of the abrupt change, and a set of abrupt change sampling points is generated. Connect consecutive sampling points in the set of mutation sampling points to form a mutation segment. Each mutation segment corresponds to a continuous change process of amplitude-frequency coordinated mutation, and the temporal order and dual change information of the sampling points within the segment are preserved. All mutation segments are integrated into a set of anomalous co-evolutionary segments, with each segment corresponding to an anomalous mutation process in amplitude-frequency co-evolution.

8. The method according to claim 1, characterized in that, The step of inputting the target detection data into a preset machine learning model and generating anomaly pattern association results through correlation modeling and anomaly pattern matching of the target detection data includes: The target detection data is divided into multiple time-series units in groups of three consecutive clock cycles. Each unit corresponds to a set of three-cycle abnormal co-evolutionary components, and the amplitude-frequency co-evolutionary correlation and multi-order transmission information within the unit are preserved. The time series units are input into a preset machine learning model, triggering cross-unit third-order dependency modeling operations within the model to explore the co-evolutionary relationship between each time series unit and the first three and last three time series units, generating a set of cross-unit dependency relationships. The cross-group dependency association set of the units is transformed by feature conversion, and each cross-group evolutionary association relationship is converted into a pattern fragment containing cross-group collaborative information. Each fragment corresponds to a cross-unit collaborative evolutionary association form. The pattern fragment is matched with the patterns in the preset DDR wafer test abnormal pattern library across groups. The consistency of the timing evolution trajectory of the pattern fragment with the first three groups and the last three groups of the patterns in the library is compared, and a list of cross-group matching results is generated. The cross-group matching result list is filtered, and pattern pairs with completely consistent cross-group evolution trajectories are retained to generate a set of cross-group matching pattern pairs. If the set is empty, an identifier with no cross-group matching pattern is generated. The identifiers of the cross-group matching pattern pairs or no cross-group matching patterns are integrated into anomaly pattern association results, which include the matched cross-group anomalous evolution patterns and the corresponding multi-level collaborative association details.

9. The method according to claim 8, characterized in that, The process involves mining the co-evolutionary relationships between each temporal unit and the preceding three groups and the following three groups of temporal units, generating a set of cross-group dependency relationships for units, including: The cross-group evolutionary relationships of units are extracted from the preset machine learning model to generate an original set of relationships. The original set of relationships includes the evolutionary direction, rate and cross-group transmission relationship information of each time unit with other units. Analyze the cross-group transitivity of each relationship in the original set of relationships, identify cross-group relationship paths that can be continuously transmitted between multiple groups of units, and generate a set of continuous cross-group relationship paths, which includes the unit transmission order and collaboration tag for each path; Each path in the set of continuous cross-group associated paths is assigned a weight. Different transmission weights are assigned based on the number of consecutive cross-group transmissions of the association relationship in the path. The more transmissions, the higher the weight, thus generating a weighted cross-group associated path. Identify path segments in the weighted cross-group association path where the transmission weights exhibit a third-order abrupt change and mark them as abnormal weight transmission segments. The transmission weights of the association relationships within these segments are discontinuous and differ from those in normal paths. The abnormal weight propagation segment is corrected by adjusting the propagation weight of the abnormal segment based on the weight distribution of the adjacent normal weighted cross-group association path, and a corrected weighted cross-group association path is generated. All the modified weighted cross-group association paths are integrated into a unit cross-group dependency association set, which includes the modified cross-group evolutionary association relationship and transmission weight information; The step of assigning weights to each path in the set of consecutive cross-group associated paths, and assigning different propagation weights based on the number of consecutive cross-group propagations of the association relationship in the path, includes: Select a target cross-group association path from the set of continuous cross-group association paths, extract the number of associations contained in the path, the number of units continuously transmitted across groups and the co-evolution marker, and record the transmission length and co-evolution density of the path; Based on the cross-group propagation law of the corresponding path in the DDR shaking physical prior, the correspondence between the number of consecutive cross-group propagation and the weight is set. For each increase in the number of consecutive propagation, the weight is increased according to a preset ratio, and the weight assignment rule is generated. Calculate the number of consecutive cross-group propagation times for each relationship in the target cross-group association path, assign propagation weights to each relationship based on the weight assignment rules, and generate the weight distribution of a single path; The transmission weights of all relationships in a single path are summed to generate the total transmission weight of the target cross-group associated path. The total weight reflects the stability and synergistic consistency of the cross-group continuous transmission of the path. Repeat the above assignment process for all target cross-group association paths to generate the total transmission weight of each path and the transmission weight of a single association, thus generating a weighted cross-group association path fragment; All weighted cross-group association path segments are integrated in path order to generate a weighted cross-group association path set, which includes the transmission weight distribution, total weight, and co-evolutionary label information for each path.

10. A computer device, characterized in that, include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the machine learning-based DDR wafer test anomaly detection method as described in any one of claims 1-9.

Images