Search

CN-121808731-B - DDR wafer test abnormality detection method and equipment based on machine learning

CN121808731BCN 121808731 BCN121808731 BCN 121808731BCN-121808731-B

Abstract

The invention provides a DDR wafer test anomaly detection method and equipment based on machine learning, which are characterized in that a sequential eye pattern and sequential waveform data related to total jitter, deterministic jitter and random jitter which are synchronously acquired in the DDR wafer test process are obtained, clock edge phase segmentation alignment is carried out on differential sequential sampling points in the sequential waveform data, phase domain dense sampling data are constructed, adaptive sparse coding is completed by combining DDR jitter physical priori, noise and normal process fluctuation related components are removed, defect sensitive sparse components are obtained, variation modal decomposition is carried out, components related to process drift, normal process fluctuation and mutation anomaly are obtained, mutation anomaly related components are extracted to serve as target detection data, a preset machine learning model is input, an anomaly mode association result is generated, the DDR wafer test anomaly detection result is output based on the clock edge phase segmentation alignment, and a corresponding test setting item adjustment instruction is generated. The invention can improve the overall efficiency and the result reliability of DDR wafer test.

Inventors

  • HUANG HUI
  • WU YUN
  • Xiang Yuechao

Assignees

  • 深圳市芯片测试技术有限公司

Dates

Publication Date
20260505
Application Date
20260310

Claims (9)

  1. 1. The DDR wafer test abnormality detection method based on machine learning is characterized by comprising the following steps: Acquiring a timing sequence eye diagram and timing sequence waveform data related to total jitter, deterministic jitter and random jitter, which are synchronously acquired in the DDR wafer test process, wherein the timing sequence waveform data comprises a continuously sampled differential timing sequence sampling point sequence with a time stamp mark; Performing clock edge phase segment alignment on differential time sequence sampling points in the time sequence waveform data, constructing phase domain dense sampling data, completing self-adaptive sparse coding by combining DDR jitter physical prior, removing noise and normal process fluctuation related components, and obtaining defect sensitive sparse components; the method specifically comprises the steps of extracting phase nodes of clock signals from time sequence waveform data and differential time sequence sampling points in the time sequence waveform data, tracking conduction dependency paths of each differential time sequence sampling point in a clock phase advancing process, analyzing conduction dependency ranges of the differential time sequence sampling points along with phase changes, generating a sampling point-phase conduction dependency correlation table, wherein the sampling point-phase conduction dependency correlation table comprises conduction initial phases, dependency node ranges and conduction path characteristics of each sampling point, connecting sampling points with homologous conduction initial phases and continuous dependency paths into phase conduction links according to phase advancing sequences based on the sampling point-phase conduction dependency correlation table, each link corresponds to a group of continuous clock phase conducting processes, retaining conduction time sequences and relations of the sampling points in the links, completing conduction path breakage positions in each phase conduction dependency link, deriving conduction dependency links of breakage positions based on dependency rules of adjacent conduction segments in the links, generating complete phase conduction dependency links, filling conduction information loss caused by sampling gaps, comparing conduction paths of all the complete phase conduction dependency links and DDR physical jitter, identifying conduction paths and conducting dependency links, generating abnormal coding and conducting segments, the method comprises the steps of converting each conducting segment into a conducting dependency coding sequence containing phase dependency relationship to obtain a conducting dependency coding sequence set, wherein each coding unit in the conducting dependency coding sequence corresponds to one conducting node and dependency path information in a link; Performing variation modal decomposition on the defect-sensitive sparse component, separating to obtain components related to process drift, normal process fluctuation and mutation abnormality, and extracting mutation abnormality related components as target detection data; Inputting the target detection data into a preset machine learning model, and generating an abnormal mode association result through association modeling and abnormal mode matching processing of the target detection data; Outputting DDR wafer test abnormality detection results based on the abnormality mode association results, generating corresponding test setting item adjustment instructions, and sending the test setting item adjustment instructions to DDR wafer test equipment to perform test setting item calibration operation.
  2. 2. The method of claim 1, wherein the connecting the sampling points having homologous conduction start phases and consecutive dependency paths in phase-advance order as phase-conduction dependency links based on the sampling point-phase-conduction dependency correlation table comprises: Extracting the conduction initial phase and the dependence path characteristics of each sampling point from the sampling point-phase conduction dependence correlation table to obtain conduction homologous identification marks, and generating a phase conduction homologous dependence mark sequence, wherein the phase conduction homologous identification mark sequence comprises a conduction homologous identification and dependence path characteristic description of each sampling point; carrying out homologous dependency clustering on the phase conduction homologous dependency imprinting sequences, classifying sampling point imprints with the same conduction initial phase and similar dependency path characteristics into a group, and generating homologous conduction dependency clustering groups, wherein each group corresponds to a continuous phase conduction process; fitting the dependency paths of the sampling points in each homologous conduction dependency cluster group to generate a conduction track reference sequence of the sampling points in the group, wherein the conduction track reference sequence keeps the unified description of the conduction time sequence and the dependency relationship of all the sampling points in the group; identifying the position of the conducting track reference sequence, at which the dependent path breaks, and marking the position as a conducting track fault, wherein the fault position corresponds to a conducting information missing area caused by a sampling gap; Pre-complementing the fault positions of the conductive tracks, deducing basic conductive properties of the fault positions based on the dependency rules of adjacent normal conductive segments in the group, and generating a fault pre-complementing conductive track reference sequence; and integrating all fault pre-complement conductive track reference sequences according to the advancing sequence of the initial conductive phases to generate phase conductive dependent links, wherein each link corresponds to a group of continuous and complete phase conductive processes.
  3. 3. The method of claim 2, wherein the pre-complementing the conductive trace fault location derives a base conductive property of the fault location based on a dependency rule of adjacent normal conductive segments in the group, generating a fault pre-complementing conductive trace reference sequence comprising: extracting normal conduction section dependency features at two sides of the fault from a conduction track reference sequence corresponding to the conduction track fault to generate a conduction dependency feature set at two sides of the fault; conducting dependency interpolation is carried out on the conducting dependency feature sets at two sides of the fault, a conducting dependency interpolation sequence from one side to the other side is generated based on the transition rule of the features at two sides, and the conducting dependency interpolation sequence comprises a complete process of continuous change of the dependency features; Matching the conduction dependent interpolation sequence with the conduction rule of DDR jitter physical priori, and adjusting characteristic parameters of the interpolation sequence to completely fit DDR jitter physical characteristics to generate a calibrated conduction dependent interpolation sequence; deducing the conduction attribute of each phase point of the fault position based on the calibrated conduction dependency interpolation sequence to generate a fault position conduction dependency attribute set; inserting the fault position conduction dependency attribute set into a corresponding conduction track reference sequence, filling the conduction information loss of the fault position, and generating a conduction track reference sequence inserted into the fault; And conducting time sequence continuity verification is conducted on the conducting track reference sequence inserted into the fault, so that the conducting time sequence of the inserted sequence is consistent with that of the original reference sequence, and a fault pre-complement conducting track reference sequence is generated.
  4. 4. The method of claim 1, wherein the performing a variational modal decomposition on the defect-sensitive sparse component, separating to obtain components related to process drift, normal process fluctuation, and mutation anomaly, and extracting the mutation anomaly related components as target detection data comprises: dividing the defect-sensitive sparse component into a plurality of groups of time sequence segments according to three continuous clock periods, analyzing the amplitude-frequency co-evolution trend of signals in each group of time sequence segments, and generating an amplitude-frequency co-evolution sequence, wherein the sequence comprises the amplitude change rate and the frequency conduction track of each group of time sequence segments; carrying out third-order dependency analysis on each trend segment in the amplitude-frequency co-evolution sequence, identifying trend segment groups with continuous co-evolution association, wherein each trend segment group covers the co-evolution process of three time sequence segments, and reserving the time sequence and the co-association of the trend segments in the groups; extracting cross-period evolution characteristics of each trend segment group, deducing a cross-period continuous evolution track based on amplitude-frequency co-evolution speed in the trend segments, and generating a cross-period evolution track set which comprises amplitude-frequency co-variation paths of each trend segment group; Comparing the inter-period evolution track set with a pre-set inter-period evolution track corresponding to a system Cheng Piaoyi and a normal process fluctuation inter-period evolution track corresponding to a normal process fluctuation, wherein the marks belong to trend segment groups of two types of tracks; Removing marked trend segment groups from the inter-period evolution track set, reserving time sequence segments corresponding to the remaining trend segment groups, converting the remaining time sequence segments into time domain signals, and generating mutation abnormal co-evolution components; And integrating all the mutation anomaly co-evolution components into mutation anomaly related components according to the original time sequence, and taking the mutation anomaly related components as target detection data, wherein the mutation anomaly related components comprise cross-period anomaly co-evolution information with amplitude-frequency dual dimensions.
  5. 5. The method of claim 4, wherein analyzing the magnitude-frequency co-evolution trend of the signal within each set of timing segments to generate the magnitude-frequency co-evolution sequence comprises: extracting the amplitude values and the corresponding frequency values of all sampling points from each group of three-period time sequence segments of the defect-sensitive sparse component, and arranging the amplitude values and the corresponding frequency values according to a sampling sequence to generate a time sequence segment amplitude-frequency subsequence, wherein the time sequence segment amplitude-frequency subsequence keeps the phase association and the co-evolution mark of the sampling points; Analyzing amplitude-frequency change correlation of each sampling point in the time sequence segment amplitude-frequency subsequence and the first three sampling points and the last three sampling points, calculating continuous change slope of the amplitude and continuous conduction rate of the frequency, and generating a slope-rate change sequence, wherein the slope-rate change sequence comprises double change information of each sampling point; Connecting sampling point segments with continuous and consistent slopes and speed changes in the slope-speed change sequence into co-evolution sub-chains, wherein each sub-chain corresponds to a segment of amplitude-frequency co-stable change process, and the phase and slope of the sampling points in the sub-chain are kept to be related to the speed; Identifying a sampling point segment with third-order mutation on slope or rate change in the co-evolution sub-chain, marking the sampling point segment as an abnormal co-evolution segment, wherein the amplitude-frequency co-variation in the segment has discontinuous difference with an adjacent sub-chain; performing cooperative trend fitting on the abnormal cooperative evolution segments, and generating corrected cooperative evolution sub-chains based on the slope of adjacent normal cooperative evolution sub-chains and the corrected cooperative trend of the rate fitting abnormal segments; And integrating all corrected co-evolution sub-chains according to a sampling sequence to generate an amplitude-frequency co-evolution sequence of each group of time sequence segments, wherein the sequence comprises the complete amplitude-frequency co-evolution trend in the time sequence segments and corrected abnormal segment information.
  6. 6. The method of claim 5, wherein identifying the sample point segment in the co-evolution subchain where a third order mutation occurs in a slope or rate change comprises: extracting a slope-rate change sequence from the co-evolution sub-chain, and arranging the slope-rate change sequence according to a sampling sequence to generate a double change sequence, wherein the double change sequence comprises amplitude change slope, frequency conduction rate and phase correlation information of each sampling point; Calculating slope difference and rate difference between each sampling point and the first three sampling points and the last three sampling points in the double change sequence to generate a third-order difference change sequence, wherein the third-order difference change sequence comprises third-order slope difference and third-order rate difference information of each sampling point; Comparing each slope difference in the third-order difference change sequence with a third-order slope difference range in a DDR wafer test normal state, and simultaneously comparing each speed difference with a third-order speed difference range in a normal state to confirm whether each difference is in a corresponding normal cooperative change range; Carrying out mutation marking on sampling points corresponding to the difference beyond the normal range, wherein marking content comprises mutation amplitude, mutation direction and corresponding phase position, and generating a mutation sampling point set; connecting continuous sampling points in the abrupt sampling point set into abrupt sections, wherein each abrupt section corresponds to a section of amplitude-frequency cooperative continuous abrupt change process, and the time sequence and double change information of the sampling points in the section are reserved; and integrating all mutation sections into an abnormal co-evolution section set, wherein each section corresponds to an abnormal mutation process in amplitude-frequency co-evolution.
  7. 7. The method of claim 1, wherein inputting the object detection data into a preset machine learning model, generating an abnormal pattern association result by performing association modeling and abnormal pattern matching processing on the object detection data, comprises: Dividing the target detection data into a plurality of time sequence units according to a group of three continuous clock periods, wherein each unit corresponds to a group of three-period abnormal co-evolution components, and preserving amplitude-frequency co-evolution association and multi-order conduction information in the unit; Inputting the time sequence units into a preset machine learning model, triggering cross-unit third-order dependency modeling operation in the model, excavating the co-evolution association relation between each time sequence unit and the first three groups of time sequence units and the last three groups of time sequence units, and generating a unit cross-group dependency association set; Feature conversion is carried out on the unit cross-group dependency association sets, each cross-group evolution association relation is converted into a mode segment containing cross-group cooperative information, and each segment corresponds to a cross-unit cooperative evolution association form; performing cross-group collaborative matching on the pattern fragments and patterns in a preset DDR wafer test abnormal pattern library, comparing the consistency of the pattern fragments with the first three groups and the last three groups of time sequence evolution tracks of the patterns in the library, and generating a cross-group matching result list; screening the cross-group matching result list, reserving mode pairs with completely consistent cross-group evolution tracks, generating a cross-group fit mode pair set, and generating an identifier of a non-cross-group matching mode if the set is empty; And integrating the identification of the cross-group fit mode pair set or the non-cross-group matching mode into an abnormal mode association result, wherein the abnormal mode association result comprises the matched cross-group abnormal evolution mode and corresponding multi-order collaborative association details.
  8. 8. The method of claim 7, wherein the mining co-evolving associations of each timing unit with the first three groups of timing units and the last three groups of timing units, generating a unit cross-group dependency association set, comprises: Extracting the mined unit cross-group evolution association relation from the preset machine learning model to generate an association relation original set, wherein the association relation original set comprises the evolution direction, the rate and the cross-group conduction association information of each time sequence unit and other units; analyzing the group-crossing conductivity of each association in the original association set, identifying group-crossing association paths capable of conducting among multiple groups of units continuously, and generating a continuous group-crossing association path set, wherein the continuous group-crossing association path set comprises the unit conducting sequence and the collaborative tag of each path; Performing weight assignment on each path in the continuous cross-group association path set, and assigning different conduction weights based on continuous cross-group conduction times of association relations in the paths, wherein the more the conduction times are, the higher the weights are, and generating weighted cross-group association paths; Identifying a path segment with third-order mutation of the conduction weight in the weighted cross-group association path, marking the path segment as an abnormal weight conduction segment, wherein the conduction weight of the association relationship in the segment has discontinuous difference with a normal path; Carrying out weight correction on the abnormal weight conduction segments, and correcting the conduction weights of the abnormal segments based on the weight distribution of the adjacent normal weight cross-group association paths to generate corrected weight cross-group association paths; integrating all the corrected weighted cross-group correlation paths into a unit cross-group dependency correlation set, wherein the unit cross-group dependency correlation set comprises corrected cross-group evolution correlation and conduction weight information; the step of carrying out weight assignment on each path in the continuous cross-group association path set, and giving different conduction weights based on continuous cross-group conduction times of association relations in the paths comprises the following steps: Selecting a target cross-group association path from the continuous cross-group association path set, extracting the association relation quantity, the continuous cross-group conduction unit quantity and the co-evolution mark contained in the path, and recording the conduction length and the co-intensity of the path; Setting a corresponding relation between continuous cross-group conduction times and weights based on the cross-group conduction rule of the corresponding paths in the DDR jitter physical prior, and increasing the weights according to a preset proportion each time the continuous conduction times are increased to generate weight assignment rules; Calculating continuous cross-group conduction times of each association relation in the target cross-group association path, and endowing each association relation with a conduction weight by combining a weight assignment rule to generate weight distribution of a single path; Accumulating the conduction weights of all the association relations in the single path to generate the total conduction weight of the target cross-group association path, wherein the total weight reflects the cross-group continuous conduction stability and the cooperative consistency of the path; repeating the assignment process for all target cross-group association paths to generate the total conduction weight of each path and the conduction weight of a single association relation, and generating a weighted cross-group association path segment; And integrating all the weighted cross-group associated path fragments according to the path sequence to generate a weighted cross-group associated path set, wherein the set comprises the conduction weight distribution, the total weight and the co-evolution mark information of each path.
  9. 9. A computer device, comprising: a memory in which a computer program is stored; A processor for loading the computer program to implement the DDR wafer test anomaly detection method based on machine learning as claimed in any one of claims 1 to 8.

Description

DDR wafer test abnormality detection method and equipment based on machine learning Technical Field The invention relates to the field of machine learning and data processing, in particular to a DDR wafer test abnormality detection method and device based on machine learning. Background DDR wafer test anomaly detection is a core link for guaranteeing the product yield in the DDR chip mass production stage, and potential anomalies in the wafer manufacturing process or design link are identified by analyzing various signal data generated in the test process, so that basis is provided for process optimization and product grading screening. At present, time sequence waveform data in the DDR wafer test process is generally collected in the industry, after data preprocessing is completed by means of general algorithms such as signal filtering and modal decomposition, the abnormal recognition work is carried out by combining a traditional machine learning model, an abnormal result obtained by recognition is mainly used for marking a wafer with problems, a part of schemes can give wide process adjustment direction suggestions based on recognition results, but the schemes do not combine with the DDR specific signal fluctuation rule to carry out directional design in the data collection stage, a large amount of invalid data with low abnormal association degree are easily introduced, the load of a subsequent data processing link is greatly increased, the general algorithm adopted in the signal processing link cannot accurately adapt to the DDR signal characteristic, normal process fluctuation and real abnormal signals are difficult to accurately distinguish, misjudgment and misjudgment of abnormal recognition are frequently caused, the recognition result and calibration logic of DDR wafer test equipment cannot be quickly converted into test setting adjustment operation capable of being directly executed, real-time calibration of the DDR wafer test equipment is difficult to be efficiently guided, and the overall efficiency and the accuracy of quality control of DDR wafer test is finally influenced. Disclosure of Invention The invention provides a DDR wafer test abnormality detection method and device based on machine learning. The embodiment of the invention provides a DDR wafer test anomaly detection method based on machine learning, which comprises the steps of obtaining a timing eye diagram and timing waveform data related to total jitter, deterministic jitter and random jitter, which are synchronously collected in the DDR wafer test process, wherein the timing waveform data comprises a continuously sampled differential signal sampling point sequence with a timestamp mark, carrying out clock edge phase segmentation alignment on the differential timing sampling points in the timing waveform data, constructing phase domain dense sampling data, completing self-adaptive sparse coding by combining DDR jitter physical priori, eliminating noise and normal process fluctuation related components to obtain defect sensitive sparse components, carrying out variation modal decomposition on the defect sensitive sparse components, separating to obtain components related to process drift, normal process fluctuation and mutation anomaly, extracting the mutation anomaly related components to serve as target detection data, inputting the target detection data into a preset machine learning model, generating an anomaly mode association result by carrying out association modeling and anomaly mode matching processing on the target detection data, outputting the DDR wafer test anomaly detection result based on the anomaly mode association result, generating a corresponding test setting item adjustment instruction, and sending the test setting item adjustment instruction to DDR wafer test equipment to carry out test setting item calibration operation. In a second aspect, an embodiment of the present invention provides a computer device, including a memory, where a computer program is stored in the memory, and a processor, where the processor is configured to load the computer program to implement the DDR wafer test anomaly detection method based on machine learning as described above. The method comprises the steps of synchronously collecting a time sequence eye diagram of a DDR wafer test and a time stamp differential sampling point sequence related to three types of core jitter, accurately locking core signal dimensions related to abnormal heights, avoiding invalid data consumption resources, constructing phase domain dense sampling data by means of clock edge phase segmentation alignment, completing self-adaptive sparse coding by combining DDR jitter physical priori, accurately filtering noise and normal process fluctuation components, focusing defect related key information, improving subsequent detection efficiency, carrying out variation modal decomposition on defect sensitive sparse components, accurately separating abrupt change ab