Wat test probe card life cycle management method and system

Abstract

This invention belongs to the field of semiconductor testing technology and discloses a method and system for lifecycle management of WAT test probe cards. The method includes: collecting multi-source data throughout the probe card's lifecycle and performing unified management to construct a digital twin data model of the probe card; performing change point detection and segmentation on the monitoring time-series data to construct a stable segment library; extracting a unified feature vector containing electrical contacts, test distribution, spatial structure, and lifetime load; calculating the probe card health index based on a machine learning regression model; predicting the remaining lifespan and failure probability using a proportional hazards survival model; generating maintenance decision suggestions and recommending execution times based on the health index, remaining lifespan, and failure probability; and updating the model through maintenance result feedback to achieve closed-loop optimization. Through the above methods, accurate assessment and predictive maintenance of the probe card's health status are achieved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing technology, and in particular to a WAT ​​test probe card lifecycle management method and system. Background Technology

[0002] The main methods currently used for wafer acceptance testing (WAT) and their corresponding drawbacks are as follows:

[0003] 1. Original diagnostic software from the testing machine manufacturer

[0004] It can monitor key parameters of the pin card in real time, such as contact resistance and leakage current, and provide trend charts and rule-based alarms.

[0005] Disadvantages: Incomplete lifecycle management: Primarily focuses on the "under testing" status, lacking deep integration with front-end and back-end data such as maintenance records and procurement / scrapping data. Weak predictive ability: Essentially based on passive threshold-based alerts rather than model-based proactive prediction. Unable to predict remaining lifespan. Data silos: Data is typically locked within a single machine or a single brand ecosystem, making it difficult to create a unified view across machines and platforms.

[0006] 2. Independent third-party data analysis and management software

[0007] They can integrate data from multiple testing machines to perform more advanced multivariate analysis. They can track the replacement history and performance of pin cards as "device components" and identify abnormal trends through SPC or basic machine learning models.

[0008] Disadvantages: Lack of specialization: These are general-purpose data analysis platforms, with the pin SIM card being just one of many monitored device components. They lack specific failure physics models and dedicated analysis tools for pin SIM cards. General-purpose models: The predictive models used may be quite general and not deeply optimized for the small sample size and multimodal data characteristics of pin SIM cards. Insufficient closed-loop processing: Analysis results often require manual intervention for decision-making.

[0009] 3. Rule-based spreadsheet + human experience management

[0010] Excel is used to record the number of times the pin cards have been used, their maintenance history, and yield performance. Engineers then decide on maintenance and disposal based on their experience.

[0011] Disadvantages: Inefficient, reliant on personal experience, and prone to errors.

[0012] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0013] The main objective of this invention is to provide a WAT ​​test probe card lifecycle management method and system, which aims to solve the technical problems in the prior art such as inaccurate probe card health status assessment, reliance on manual experience for maintenance, and inability to achieve predictive maintenance.

[0014] To achieve the above objectives, the present invention provides a WAT ​​test probe card lifecycle management method, which includes the following steps:

[0015] Collect multi-source data related to the entire lifecycle of the probe card from at least one test machine, and perform unified governance and structured organization on the multi-source data to establish a probe card digital twin data model. The multi-source data includes at least probe card configuration data, test process monitoring timing data, wafer test result data, and probe card maintenance event data.

[0016] Change point detection and segmentation are performed on at least one key parameter sequence in the time series data of the test process monitoring to form multiple time-continuous running segments. Multivariate summary features are extracted from each running segment, and a stable segment library is constructed based on the extracted multivariate summary features.

[0017] Based on the test process monitoring timing data, wafer test result data and cumulative usage data of the pin card corresponding to each running segment in the stable segment library, a unified feature vector is constructed. The unified feature vector includes electrical and contact quality signal features, WAT test result distribution and yield risk signal features, spatial structure signal features, and lifetime and usage intensity signal features.

[0018] The unified feature vector is input into the machine learning regression model to obtain the needle card health index for each running segment. The training sample labels of the machine learning regression model are defined based on the needle card life cycle events and production line performance in historical data.

[0019] The unified feature vector is input into the proportional risk survival model to obtain the remaining service life of the probe card and the probability of failure within a specified future time window.

[0020] Based on the needle card's health index, remaining lifespan, and failure probability, and combined with preset quantitative triggering rules, maintenance and usage strategy decisions are generated. The maintenance and usage strategy decisions include at least the needle card's maintenance, downgrade or scrapping replacement suggestions, and recommended execution time points.

[0021] Based on the maintenance and usage strategy decisions, WAT test probe card lifecycle management is implemented, and maintenance operation results and performance retest data are written back to the probe card digital twin data model, and used to update the machine learning regression model and the proportional risk survival model.

[0022] In one embodiment, the electrical and contact quality signal characteristics are calculated based on the contact resistance or leakage current time-series data within the operating segment, including statistics describing the central trend, dispersion, tail risk, and trend of change of the data distribution; the WAT test result distribution and yield risk signal characteristics are calculated based on the result distribution of wafer test parameters within the operating segment, including indicators describing the distribution shape, dispersion, proportion of specification limits, and difference from historical benchmark distribution.

[0023] In one embodiment, the extraction process of the spatial structure signal features includes: when there is a test value distribution map in units of grains, representing the test values, test residuals or normalized deviation values ​​as a two-dimensional matrix; performing at least one of the following operations on the two-dimensional matrix to generate numerical features: calculating spatial autocorrelation statistics, identifying and quantifying the size and intensity of local anomalous clusters, analyzing gradients or bending trends in the row and column directions, and extracting texture energy distribution features through two-dimensional frequency domain transformation.

[0024] In one embodiment, the definition of the training sample label includes marking the running segment that meets the preset yield standard and has not experienced any needle card-related failure events within a preset future observation window as a first health state sample; and marking the running segment located within a preset early warning window before the occurrence of needle card-related maintenance or scrapping events, or the running segment whose test indicators continue to deteriorate and have not recovered after maintenance, as a second health state sample.

[0025] In one embodiment, the proportional hazards survival model is a deep learning-based Cox proportional hazards model; the training samples of the model are constructed in the form of triples, and the unified feature vector of the corresponding running segment is used as the time-varying covariate within the time interval; wherein, the triple form includes the segment start time, the segment end time and the event marker, and the failure event in the event marker is defined as a preset event type that causes the needle card to be terminated, including needle card scrapping, replacement of key components and maintenance failure.

[0026] In one embodiment, the preset quantitative triggering rule is as follows: when the health index of the needle card is continuously lower than the health threshold and the remaining service life is shorter than the maintenance lead time, a maintenance recommendation is triggered; when the failure probability within a specified future time window exceeds the risk threshold, a maintenance or replacement recommendation is triggered; when the proportion of recently running segments identified as abnormal exceeds the abnormality threshold, a downgrade recommendation is triggered.

[0027] In one embodiment, the method further includes:

[0028] Obtain the set of future available maintenance periods from the manufacturing execution system;

[0029] Determine the comprehensive decision cost for each available maintenance period, wherein the comprehensive decision cost includes the production cost of downtime during that period and the expected risk cost corresponding to delayed maintenance, and the expected risk cost is calculated based on the failure probability;

[0030] The available maintenance period with the lowest overall decision-making cost is selected as the recommended execution time.

[0031] Furthermore, to achieve the above objectives, this invention also proposes a WAT ​​test probe card lifecycle management system, which is applied to the WAT test probe card lifecycle management method described above. The system includes:

[0032] The acquisition module is used to acquire multi-source data related to the entire life cycle of the probe card from at least one test machine, and to perform unified governance and structured organization of the multi-source data to establish a probe card digital twin data model. The multi-source data includes at least probe card configuration data, test process monitoring timing data, wafer test result data, and probe card maintenance event data.

[0033] The processing module is used to perform change point detection and segmentation on at least one key parameter sequence in the time series data of the test process monitoring, form multiple time-continuous running segments, extract multivariate summary features from each running segment, and construct a stable segment library based on the extracted multivariate summary features.

[0034] The construction module is used to construct a unified feature vector based on the test process monitoring timing data, wafer test result data and cumulative usage data of the pin card corresponding to each running segment in the stable segment library. The unified feature vector includes electrical and contact quality signal features, WAT test result distribution and yield risk signal features, spatial structure signal features and lifetime and usage intensity signal features.

[0035] The input module is used to input the unified feature vector into the machine learning regression model to obtain the needle card health index for each running segment. The training sample labels of the machine learning regression model are defined based on the needle card life cycle events and production line performance in historical data.

[0036] The input module is used to input the unified feature vector into the proportional risk survival model to obtain the remaining service life of the probe card and the probability of failure within a specified time window in the future.

[0037] The generation module is used to generate maintenance and usage strategy decisions based on the health index, remaining service life and failure probability of the needle card and in combination with preset quantitative triggering rules. The maintenance and usage strategy decisions include at least the maintenance, downgrade or scrapping replacement suggestions for the needle card and the recommended execution time.

[0038] The management module is used to implement WAT test probe card lifecycle management based on the maintenance and usage strategy decisions, write maintenance operation results and performance retest data back to the probe card digital twin data model, and update the machine learning regression model and the proportional risk survival model.

[0039] Furthermore, to achieve the above objectives, the present invention also proposes a WAT ​​test probe card lifecycle management device, which includes: a memory, a processor, and a WAT ​​test probe card lifecycle management program stored in the memory and executable on the processor. The WAT test probe card lifecycle management program is configured to implement the steps of the WAT test probe card lifecycle management method described above.

[0040] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a WAT ​​test probe card lifecycle management program, wherein when the WAT test probe card lifecycle management program is executed by a processor, it implements the steps of the WAT test probe card lifecycle management method described above.

[0041] This invention collects and manages multi-source data throughout the entire lifecycle of probe cards, constructing a digital twin data model for the probe cards. It performs change point detection and segmentation on the monitoring time-series data, building a stable segment library. It extracts a unified feature vector containing electrical contacts, test distribution, spatial structure, and lifespan load. Based on a machine learning regression model, it calculates the probe card health index. A proportional hazards survival model is used to predict the remaining service life and failure probability. Based on the health index, remaining service life, and failure probability, it generates maintenance decision suggestions and recommends execution times. The model is updated through maintenance result feedback, achieving closed-loop optimization. Through these methods, accurate assessment and predictive maintenance of the probe card's health status are achieved. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the first embodiment of the WAT test probe card lifecycle management method of the present invention;

[0043] Figure 2 This is a structural block diagram of the first embodiment of the WAT test probe card lifecycle management system of the present invention.

[0044] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0045] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0046] This invention provides a method for managing the lifecycle of a WAT ​​test probe card, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of a WAT ​​test probe card lifecycle management method according to the present invention.

[0047] In this embodiment, the WAT test probe card lifecycle management method includes the following steps:

[0048] Step S10: Collect multi-source data related to the entire life cycle of the probe card from at least one test machine, and perform unified governance and structured organization on the multi-source data to establish a probe card digital twin data model.

[0049] In this embodiment, the executing entity is a WAT ​​test probe card lifecycle management device. This WAT test probe card lifecycle management device has functions such as data processing, data communication, and program execution. The WAT test probe card lifecycle management device can be a computer terminal device or other network device, or other devices with similar functions. This embodiment does not limit this.

[0050] It's important to note that the current main methods used for Wafer Acceptance Testing (WAT) and their corresponding shortcomings are as follows: 1. The diagnostic software provided by the original manufacturer of the testing equipment can monitor key parameters of the pin cards in real time, such as contact resistance and leakage current, providing trend charts and rule-based alarms. Disadvantages: Not covering the entire lifecycle: It mainly focuses on the "under testing" status and lacks deep integration with front-end and back-end data such as maintenance records and procurement scrapping. Weak predictive ability: Essentially, it is a passive alarm based on thresholds, rather than a proactive prediction based on models. It cannot predict the remaining lifespan. Data silos: Data is usually locked within a single testing equipment or a single brand ecosystem, making it difficult to form a unified view across equipment and platforms. 2. Independent third-party data analysis and management software can integrate data from multiple testing equipment to perform more advanced multivariate analysis. They can track the replacement history and performance of pin cards as "equipment components" and discover abnormal trends through SPC or basic machine learning models. Disadvantages: Not specialized: They are general-purpose data analysis platforms, and pin cards are just one of many equipment components they monitor. There is a lack of specific failure physics models and dedicated analysis tools for needle cards. The models used are generic and not deeply optimized for the small sample size and multimodal data of needle cards. The closed-loop mechanism is insufficient: analysis results often require human intervention for decision-making. 3. Rule-based spreadsheet + human experience management: Excel is used to record the number of times needle cards are used, maintenance history, and yield performance. Engineers decide on maintenance and disposal based on experience. Disadvantages: Inefficient, reliant on personal experience, and prone to errors.

[0051] To address the aforementioned technical challenges, this embodiment collects and manages multi-source data throughout the probe card's lifecycle, constructing a digital twin data model for the probe card. It performs change point detection and segmentation on the monitoring time-series data, building a stable segment library. It extracts a unified feature vector encompassing electrical contacts, test distribution, spatial structure, and lifetime load. A machine learning regression model is used to calculate the probe card's health index. A proportional hazards survival model is employed to predict remaining service life and failure probability. Maintenance decision suggestions and recommended execution times are generated based on the health index, remaining service life, and failure probability. The model is updated through maintenance result feedback, achieving closed-loop optimization. Through these methods, accurate assessment and predictive maintenance of the probe card's health status are achieved, which can be implemented as follows.

[0052] In a specific implementation, multi-source data related to the entire lifecycle of the probe card is collected from at least one test machine, and the multi-source data is uniformly managed and structured to establish a probe card digital twin data model. The multi-source data includes at least probe card configuration data, test process monitoring timing data, wafer test result data, and probe card maintenance event data.

[0053] It should be noted that the system gathers all data related to the use of the needle card, covering monitoring signals and test output performance during the testing process. It also includes management events such as maintenance, verification, replacement, and scrapping of the needle card before and after use, as well as the corresponding machine and process condition information, forming a traceable data source system.

[0054] In the specific implementation, the collected data includes: pin card configuration data: supplier information, brand / model, pin card profile (number of pins, layout, applicable products, etc.); pin card status data: collected health parameters of the pin card itself (such as contact resistance, leakage current, etc.); test data: real-time collection of test results for each wafer, including the distribution of test values ​​for key parameters, test time, etc.; and pin card usage data: maintenance logs (cleaning, pin repair, replacement), pin card on / off records, and responsible personnel. The hardware interfaces used for data collection include: subscribing to device events and variables via the SECS / GEM protocol, obtaining real-time parameters via the original manufacturer's dedicated API, or parsing log / CSV / JSON files generated by the machine. To support differences between different brands / versions, the system adopts a unified access architecture of "adapter plugin + template configuration".

[0055] The process of accessing test equipment data from different brands through the expandable adapter module includes: acquiring raw data from the equipment protocol interface, dedicated API, or log file through the collector; parsing the raw data into an internal intermediate format through the parser; mapping the parsed data fields to a unified data dictionary according to a predefined configuration template through the mapper to achieve the unification of time base, units, and key identifiers; and performing quality verification and anomaly marking on the mapped data through the validator. For example, an adapter should include at least a Collector, a Parser, a Mapper, and a Validator: The Collector is responsible for establishing connections with SECS / GEM sessions or APIs, or monitoring and reading logs from specified directories / message queues; the Parser parses the raw messages / files into an internal intermediate format (unified key-value pairs / table records), retaining the original fields and units; the Mapper maps fields from different sources to a unified data dictionary according to the "data mapping rule configuration," for example, unifying the timestamp field of different machines to UTC milliseconds, the contact resistance field to ohms, and the wafer number to wafer_id, and establishing primary key associations such as programmer_card_id, equipment_id, recipe_id, and lot_id; the Validator performs quality checks and anomaly marking, including missing rate, duplicate records, reverse time, inconsistent units, out-of-bounds values, and outliers (MAD or IQR rules can be used), and outputs a quality score. The above mapping rules configuration can be described using YAML / JSON template files, including field names, data types, unit conversions, enumeration mappings, primary key concatenation rules, and version numbers; when adding a machine or field, it can be extended simply by adding / modifying the template or loading a new plugin.

[0056] Furthermore, in unified governance and structured organization, data processing includes unified governance of multi-source data, such as unifying time and units, aligning field definitions, handling duplicates and missing data, marking outliers and conducting quality assessments, and standardizing the mapping of data differences across different machines / platforms to ensure consistency and comparability of subsequent analysis and model inputs. Data organization, for example, uses a "needle card—usage interval—event chain" approach, archiving test process data according to continuous usage intervals and associating events such as maintenance, verification, machine setup / disconnection, and replacement with corresponding intervals, forming a data chain that runs through the front and back ends of use, providing continuous and structured input for health assessment and lifespan prediction.

[0057] Furthermore, this embodiment also involves data storage and digital twin. Among them, the real-time database stores high-frequency test event streams. The historical database / data warehouse stores all historical test data, maintenance records, and needle card profiles (number of needles, layout, applicable products, etc.) in a structured manner. The needle card digital twin creates a virtual model for each physical needle card, associates all data in its entire life cycle, and is the core data model of the system.

[0058] Step S20: Perform change point detection and segmentation on at least one key parameter sequence in the monitored timing data of the test process to form multiple continuously running segments, extract multivariate summary features from each running segment, and construct a stable segment library based on the extracted multivariate summary features.

[0059] It should be noted that for change point detection and segmentation, for example, perform Page-Hinkley change point detection on at least one key sequence such as contact resistance, leakage current, and key test time, obtain the change point set {τk}, and accordingly divide the time line into running segments Pk = [τk, τk+1).

[0060] Fragment length constraint and transient processing, for example, to avoid interference of short transients on subsequent modeling, the system sets a minimum length Lmin and a maximum length Lmax for the segments. When the segment length < Lmin, perform one of the following: a) Merge with the adjacent segment (if the distance between segment summary features is less than the threshold ε); b) Mark as a "transient segment" and downweight or exclude it during modeling; when the segment length > Lmax, it can be further subdivided according to a fixed window to ensure statistical stability.

[0061] The segment summary feature is to construct a multivariate summary feature for each segment, including but not limited to robust statistics (median, IQR, P95 / P99), short-window slope / second-order difference, distribution drift distance (PSI / KS), tail risk (proportion close to the spec boundary), and fluctuation amplification coefficient, etc.

[0062] Recognition of hidden abnormal segments, for example, input the segment summary feature into Isolation Forest to obtain the abnormal score score(Pk). When score(Pk) exceeds the threshold θa, it is determined as an abnormal segment; where θa can take the q quantile of the abnormal scores in the training set (for example, q = 0.95), or be automatically given by the contamination parameter. The abnormal segment will be assigned an abnormal label and confidence level, and downweighted in model training (for example, weight w = 0.2 - 0.5), and "this segment may be affected by accidental working conditions" will be prompted in the online output.

[0063] The fragment library output consists of fragment library entries, which include at least the fragment start and end times, a summary of applicable recipes / temperature conditions, a stability score, anomaly tags, and pointers related to maintenance / logout events.

[0064] Step S30: Based on the test process monitoring timing data, wafer test result data, and cumulative usage data of the pin card corresponding to each running segment in the stable segment library, construct a unified feature vector.

[0065] In this embodiment, the unified feature vector includes electrical and contact quality signal features, WAT test result distribution and yield risk signal features, spatial structure signal features, and lifespan and usage intensity signal features.

[0066] In the micro-contact physical model between the needle tip and the pad, the electrical and contact quality signal characteristics are determined by the "actual metal contact area / number of contact points" and the "conductivity characteristics of the interface film (contamination / oxidation)." Needle tip wear / passivation causes an irreversible decrease in microstructure and actual contact area, while contamination / oxidation accumulation introduces additional film resistance and stronger random conduction / breakdown behavior at the interface. Insufficient contact force due to probe arm fatigue or elastic decay also amplifies these effects. This mechanism manifests in test observations as a gradual increase in the contact resistance baseline, increased dispersion, and an increase in occasional "contact failure" events. Contamination / oxidation is more likely to cause tail-end aggravation and abnormal spikes, while wear / elastic decay tends to show a slow overall increase with limited recovery after maintenance and cleaning. Based on the above physical model and observational phenomena, this invention calculates the median, IQR, P95 / P99, extreme value proportion, sliding window slope, fluctuation amplification factor, and anomaly frequency for monitoring items such as contact resistance and leakage current according to segments (stable working intervals formed by condition switching identification). This forms a feature vector that simultaneously reflects gradual degradation (baseline rise, positive slope, and increased fluctuation factor) and occasional instability (tail position rise, increased extreme value proportion, and increased anomaly frequency). This vector is used to characterize the evolution of physical states such as decrease in actual contact area, film thickness increase, and contact instability.

[0067] The distribution of WAT test results and the characteristics of yield risk signals are as follows: In the measurement link model, changes and instabilities in contact resistance are projected onto the WAT parameter measurement results in the form of "systematic bias + random noise". When tip wear / elastic decay leads to insufficient contact force or contamination of the film layer leads to inconsistent conduction states, the measured values ​​will exhibit overall drift, variance amplification, and increased tail risk. If local pin failure or local contact deterioration occurs, some dies or structures will fall near the specification boundary more frequently, resulting in a significant increase in the "proportion of samples close to the specification". This mechanism is observed to manifest as changes in the statistical distribution pattern (changes in mean / variance / tail shape) under the same recipe / similar structure, and is more pronounced in repeated measurements or comparisons between adjacent batches, thus directly affecting yield risk and boundary sample density. Based on this, the present invention forms the distribution center (mean / median), dispersion (standard deviation / IQR), tail risk density (proportion close to the upper and lower limits of spec) and drift distance (PSI, KS statistic) of each test item according to the pin / wafer dimension. When only summary statistics are available and the original sample distribution is lacking, quantile vectors can be used to replace the original distribution to calculate the drift and tail risk approximation, which statistically characterizes the measurement offset and noise increase caused by contact degradation, and enhances the sensitivity to the increase of boundary samples caused by local failure.

[0068] The spatial structure signal characteristics (based on a die-map numerical algorithm) are as follows: In the geometric-mechanical consistency model, local needle position offsets, local wear / contamination accumulations, and changes in card surface flatness (warping, tilting) of the probe card lead to systematic inconsistencies in contact conditions at different spatial locations, thus transforming failures from "random scattered points" into "spatially correlated structures" on the die-map. Specifically, height / positional anomalies of single needles or local needle groups are more likely to form relatively stable hot spot clusters, while large-scale geometric changes such as flatness / warping are more likely to form gradient or strip-like defects along the row and column directions. When degradation is dominated by local mechanisms, anomalies exhibit strong spatial autocorrelation and connectivity; when degradation is dominated by global mechanisms, it manifests as an enhanced spatial low-frequency trend. To transform the aforementioned spatial observations into learnable numerical features, when a die-map exists, it is represented as a two-dimensional matrix M (where each element represents the test value, residual, or standardized deviation of an item on the die), and one or a combination of the following is performed to complete the "die-map → numerical feature" transformation: First, robust centering is performed to reduce global drift interference (e.g., calculating R = M). median(M) or quantile scaling of M to obtain a robust residual matrix); then (a) calculate the spatial autocorrelation index: using die adjacency as the weight matrix W (four-neighborhood or eight-neighborhood), calculate Moran's I or Geary's C on R to quantify whether the anomalies are clustered rather than random; (b) calculate the local anomaly cluster index: on |R| Thresholding is performed to obtain a binary matrix B (the threshold can be the quantile of |R| or k·MAD based on spec / noise estimation). Then, a connected component algorithm (four-connected or eight-connected) is used on B to identify cluster sets and output features such as the number of clusters, the maximum cluster area, the cluster area distribution statistics (mean / variance), and the cluster intensity mean (the average value of |R| within the cluster or the integral of the absolute value of R) to characterize hot spot defects caused by "local needle position / local contamination"; (c) Calculate row and column gradients and bending indices: Calculate the mean of R by row and column to obtain the row mean curve and column mean curve, respectively. Fit the first term of the curve to obtain the slope (global tilt) and calculate the second difference or quadratic term coefficient to obtain the bending / warping intensity (reflecting the systematic spatial trend caused by the change in card surface flatness); (d) Calculate frequency domain texture indices: Perform 2D FFT on R to obtain the spectral energy distribution, calculate the low-frequency / mid-frequency energy ratio, the energy concentration in the main direction, etc., to capture stripes, periodic textures or directional anomalies (typically due to flatness changes or systematic contact inhomogeneity). When there is no die-map, spatial structure features can be degraded to summarizing statistics by region (center / edge / quadrant) (regional mean difference, regional variance difference, edge-center offset, etc.) to maintain a minimum ability to represent large-scale spatial inhomogeneity.

[0069] Furthermore, the process of extracting spatial structure signal features includes: when there is a test value distribution map in units of grains, representing the test values, test residuals, or normalized deviation values ​​as a two-dimensional matrix; performing at least one of the following operations on the two-dimensional matrix to generate numerical features: calculating spatial autocorrelation statistics, identifying and quantifying the size and intensity of local anomalous clusters, analyzing gradients or bending trends in the row and column directions, and extracting texture energy distribution features through two-dimensional frequency domain transformation.

[0070] In the cumulative damage and recoverability model, the characteristics of lifespan and usage intensity signals are as follows: probe arm fatigue, plastic deformation, and elastic decay are typically driven by cyclic loading and time effects, with their degradation degree showing a monotonically cumulative relationship with the cumulative number of touchdowns, machine usage time, temperature / process stress exposure, and maintenance history. In contrast, contamination / oxidation buildup is more consistent with a reversible process of "significant recovery after cleaning as usage and exposure time increases." Therefore, in observation, material fatigue / elastic decay often manifests as health-related signals continuously deteriorating with load and the recovery amplitude gradually decreasing after maintenance, while contamination-driven degradation exhibits a periodic pattern of "the longer the time since the last cleaning → the heavier the tail / fluctuation → a significant decline after cleaning." Based on this, the present invention constructs lifespan and usage intensity characteristics, including cumulative touchdown, cumulative machine usage time, time interval since the last cleaning / maintenance, number of maintenance, and the recovery magnitude (Δ recovery) and recovery decay trend of key indicators before and after maintenance, in order to characterize the cumulative law of irreversible fatigue-type degradation and the recovery law of reversible pollution-type degradation, and provide time-varying load and recoverability evidence for subsequent health index and lifespan prediction.

[0071] Step S40: Input the unified feature vector into the machine learning regression model to obtain the needle health index for each running segment.

[0072] It should be noted that the monitoring signal is generated based on lifecycle events and production line performance. The training sample labels of the machine learning regression model are defined based on the lifecycle events of the pin card and the production line performance in historical data. The first health status sample is a high health status sample, for example, within the usage period of the segment, the yield of the key test item is ≥ Ygood (e.g., 99% or set by the product), and the key monitoring items (contact resistance / leakage current) have not shown K consecutive abnormal segments, and no maintenance / downtime event triggered by the pin card has occurred in the future Wpre window. The second health status sample is a low health status sample, for example, meeting one of the following: i) It is located in the Wpre window before the maintenance / repair / scrap event occurs, and the cause field of the event is related to the pin card (contamination, poor contact, pin position damage, etc.); ii) The risk density at the end of the segment exceeds Rbad or PSI / KS exceeds Dbad and lasts for K segments; iii) The retest in the Wpost window after maintenance still cannot recover to the stable threshold (defined as maintenance invalid).

[0073] Furthermore, to make engineering applications more direct, this solution uniformly maps the health status of the SIM card to a Health Index (HI) (0–100) and provides a continuous curve over time. This embodiment employs a machine learning regression model, which demonstrates stable performance on multivariate, nonlinear, highly interactive, and complex tabular data with categorical conditions, and exhibits fast inference speed, making it suitable for online deployment. The model input consists of fragment-level features and conditional summaries, and the output is the PCHI score. The model can be trained in groups by device / product or share parameters using a multi-task approach. The PCHI supervision signal comes from the structured definition of historical lifecycle events: the system marks "stable and usable fragments" as high health and "frequent anomalies / fragments significantly deteriorated before maintenance / windows before failure or scrapping" as low health, and introduces consistency constraints for maintenance events: if the same SIM card shows significant improvement after maintenance, the model should learn a PCHI rebound; if it continues to deteriorate after maintenance, the model should output a persistently low PCHI. Through this "event-driven + consistency constraint" approach, PCHI is not merely a classifier output, but a quantitative health indicator with engineering semantics. When running online, the system provides interpretable output. It updates the PCHI curve according to fixed time windows or segment boundaries and provides "health trends, inflection points, and abnormal segment markers" on the interface, allowing engineers to intuitively see the process of the needle card from health to deterioration, rather than passively waiting for threshold triggers.

[0074] Step S50: Input the unified feature vector into the proportional hazards survival model to obtain the remaining service life of the probe card and the probability of failure within a specified time window in the future.

[0075] In practical implementation, health scores alone (such as PCHI) are insufficient to support planned maintenance. Therefore, Remaining Life Prediction (RUL) is introduced based on health assessment: for each SIM card, the failure probability curve, survival function curve, and RUL estimate are output in the dimensions of future time (hours / days) or future touchdowns, providing a quantitative basis for maintenance scheduling, spare parts preparation, and risk control. This solution uses DeepSurv (Deep Cox Proportional Risk Survival Model) as the RUL prediction model. This model uses "time to event occurrence" as the modeling object, adapting to the characteristics of SIM cards where "failure events are relatively sparse, but process signals are rich." Its core advantage lies in its ability to correctly utilize truncated data (SIM cards that have not failed by the end of the observation period) through survival analysis, avoiding the bias introduced by crudely treating non-failed samples as negative samples. Simultaneously, it can integrate multi-source features, unifying electrical, distribution, spatial, and usage intensity inputs to output stable and interpretable risk estimates.

[0076] It should be noted that the survival analysis model is a Cox proportional hazards model based on deep learning; the training samples of the model are constructed in the form of triples, and the unified feature vector of the corresponding running segment is used as the time-varying covariate within the time interval; wherein, the triple form includes the segment start time, the segment end time and the event marker, and the failure event in the event marker is defined as a preset event type that causes the needle card to be terminated, including needle card scrapping, replacement of key components and maintenance failure.

[0077] Furthermore, the system defines a failure event as a set of termination events where the probe card exits a stable production state and requires discontinuation or major maintenance. By default, it includes the following event types, which can be configured as single or combined events according to factory management standards: E1 Scrap (probe card retirement / unusable); E2 Overhaul or replacement of critical components (e.g., replacing probes, probe bundles, probe arms, etc., representing a degradation stage that cannot be recovered through routine cleaning); E3 Unplanned shutdown due to probe card issues, where maintenance fails to restore the device to a stable threshold within N wafers (or N segments) (e.g., insufficient PCHI recovery or critical contact indicators still exceeding limits), considered an "unrecoverable failure due to ineffective maintenance." When different termination causes need to be distinguished, a cause-specific survival model or competing risk extension can be used, but this does not change the target and scope of the "failure probability curve and RUL estimation" output in this step.

[0078] For truncated data processing and time-varying covariate input, DeepSurv natively supports truncated data. For SIM cards that have experienced a termination event, the event time is recorded as T and marked as event=1; for SIM cards that have not failed by the end of the observation period, the last observable time is recorded as C and marked as event=0 (truncated). The model incorporates these as truncated samples into survival modeling, thus correctly utilizing the information that "they have survived to the current time but it is unknown whether they will fail in the future". For feature input, the system not only uses a single health score but also introduces time-varying covariates to characterize the degradation process: using the stable usage segment Pk (or usage interval) obtained in step S2 as the time alignment unit, a feature vector xk is calculated for each segment (consistent with the PCHI input, including segment features, conditional summary, usage intensity features, etc., and can be further integrated with electrical, WAT distribution, spatial structure, and lifetime load features), and samples in the form of (start, stop, xk, event) are constructed, where start / stop is the segment start and end time (or the cumulative value of start and end touchdown); if the above failure event occurs at the end of the segment, event=1, otherwise event=0; if there is still no failure by the end of the observation period, the last segment event=0 is used as the truncation. Through this construction, the evolution of segment features over time is used as a time-varying covariate input to DeepSurv, realizing survival modeling of the "process signal-rich" needle card degradation trajectory.

[0079] Model output, risk threshold and RUL calculation, confidence level determination: DeepSurv outputs the risk function λ(t|x) and survival function S(t|x). The system can define the time axis as natural time (hours / days) or touchdowns, and further calculate the failure probability within a specified window H, such as the failure probability within the next 24 / 48 / 72 hours or the next T touchdowns.

[0080]

[0081] Based on this, the system implements executable maintenance triggering and RUL estimation through risk thresholds: given a risk threshold pth or a survival threshold Sth, RUL is defined as satisfying... The minimum H, or satisfying The minimum H. pth and Sth are calibrated and optimized on the validation set using historical data. For example, under the constraint of the target false positive rate α (or acceptable false stop rate), the threshold that maximizes recall or minimizes overall cost can be selected. Differentiated thresholds can be set according to product / recipe / process segment to adapt to different risk baselines.

[0082] To provide "predicted reliability" for engineering decision-making, the system also outputs a confidence level: by estimating the prediction uncertainty through methods such as integrated models / bootstrap or Monte Carlo dropout, the confidence interval width CI of Pfail(H) under the same window H is obtained, and graded accordingly: CI ≤ c1 is high confidence, c1 < CI ≤ c2 is medium confidence, CI > c2 is low confidence; where c1 and c2 can also be calibrated on the validation set to ensure that the actual hit rate and calibration error under different confidence levels meet the engineering requirements.

[0083] Step S60: Generate a maintenance and usage strategy decision based on the needle card health index, remaining service life, and failure probability, and in combination with a preset quantitative trigger rule.

[0084] It should be noted that the preset quantitative trigger rule is that when the needle card health index is continuously lower than the health threshold and the remaining service life is shorter than the maintenance lead time, a maintenance recommendation is triggered; when the failure probability within a specified future time window exceeds the risk threshold, a maintenance or replacement recommendation is triggered; when the proportion of segments identified as abnormal in the recent operation exceeds the abnormality threshold, a downgraded usage recommendation is triggered. In specific implementation, for example, a) Maintenance recommendation: PCHI is lower than the threshold Tmaint for K consecutive segments, and RUL ≤ Rmaint, or Pfail(H) ≥ Pmaint; b) Downgraded usage recommendation: The proportion of abnormal segments Aseg exceeds Ath within the most recent window, or the tail risk density Rtail of the key test item exceeds Rdeg but has not reached the scrapping threshold; c) Scrapping / replacement recommendation: PCHI is lower than Tscrap and persists for Kscrap segments, or Pfail(H) ≥ Pscrap, or the number of ineffective maintenance times ≥ Mfail. Among them, Tmaint, Rmaint, Pmaint, Ath, Rdeg, Tscrap, Pscrap, K, Kscrap, and Mfail are all configurable parameters.

[0085] In one embodiment, a set of future available maintenance periods from the Manufacturing Execution System (MES) is obtained; the comprehensive decision cost for each available maintenance period is determined, wherein the comprehensive decision cost includes the production cost of downtime during the period and the expected risk cost corresponding to delayed maintenance, the expected risk cost being calculated based on the failure probability; the available maintenance period with the minimum comprehensive decision cost is selected as the recommended execution time. In this embodiment, the recommended execution time is when the system can obtain MES / scheduling information, using the set of future available maintenance windows W={wi} as a constraint, calculating the comprehensive cost C(wi) = downtime cost D(wi) + expected risk cost Risk(wi) for each window, and selecting the window that minimizes C as the recommended time; when the schedule cannot be obtained, the system adopts a rule-based strategy: prioritizing the natural downtime window of the next batch / temperature / recipe change, or selecting the nearest available maintenance shift window before RUL enters Rmaint.

[0086] Step S70: Based on the maintenance and usage strategy decision, implement WAT test probe card lifecycle management, write back the maintenance operation results and performance retest data to the probe card digital twin data model, and use them to update the machine learning regression model and the proportional risk survival model.

[0087] In the specific implementation, this embodiment also supports closed-loop feedback and model update, writing back the maintenance operation results and performance retest data to the needle card digital twin data model, and using it to update the machine learning regression model and the proportional risk survival model. Specific implementations include (1) writing back the history: writing back the actual maintenance type, time, responsible person, post-maintenance retest improvement range, recurrence status, etc. to the needle card digital twin; (2) threshold and rule self-calibration: dynamically adjusting thresholds such as Tmaint and Pmaint according to the maintenance effectiveness statistics, or forming differentiated rules according to the product family tree; (3) model update: incrementally training / retraining the PCHI and DeepSurv models according to the cycle or trigger conditions, and updating the baseline for new machines / new products through transfer learning or domain adaptation.

[0088] In this embodiment, multi-source data from the entire lifecycle of the probe card is collected and uniformly managed to construct a digital twin data model of the probe card; change point detection and segmentation are performed on the monitoring time-series data to build a stable segment library; a unified feature vector containing electrical contacts, test distribution, spatial structure, and lifespan load is extracted; the probe card health index is calculated based on a machine learning regression model; the remaining service life and failure probability are predicted using a proportional hazards survival model; maintenance decision suggestions and recommended execution times are generated based on the health index, remaining service life, and failure probability; and the model is updated through maintenance result feedback to achieve closed-loop optimization. Through the above methods, accurate assessment and predictive maintenance of the probe card health status are achieved.

[0089] Furthermore, this embodiment of the invention also proposes a storage medium storing a WAT ​​test probe card lifecycle management program, which, when executed by a processor, implements the steps of the WAT test probe card lifecycle management method described above.

[0090] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the WAT test probe card lifecycle management system of the present invention.

[0091] like Figure 2 As shown, the WAT test probe card lifecycle management system proposed in this embodiment of the invention includes:

[0092] The acquisition module 10 is used to acquire multi-source data related to the entire life cycle of the probe card from at least one test machine, and to perform unified governance and structured organization of the multi-source data to establish a probe card digital twin data model. The multi-source data includes at least probe card configuration data, test process monitoring timing data, wafer test result data, and probe card maintenance event data.

[0093] Processing module 20 is used to perform change point detection and segmentation on at least one key parameter sequence in the time series data of the test process monitoring, form multiple time-continuous running segments, extract multivariate summary features from each running segment, and construct a stable segment library based on the extracted multivariate summary features;

[0094] The construction module 30 is used to construct a unified feature vector based on the test process monitoring timing data, wafer test result data and cumulative usage data of the pin card corresponding to each running segment in the stable segment library. The unified feature vector includes electrical and contact quality signal features, WAT test result distribution and yield risk signal features, spatial structure signal features and lifespan and usage intensity signal features.

[0095] Input module 40 is used to input the unified feature vector into the machine learning regression model to obtain the needle card health index for each running segment, wherein the training sample labels of the machine learning regression model are defined based on the needle card life cycle events and production line performance in historical data;

[0096] The input module 40 is used to input the unified feature vector into the proportional risk survival model to obtain the remaining service life of the probe card and the probability of failure within a specified time window in the future.

[0097] The generation module 50 is used to generate maintenance and usage strategy decisions based on the health index, remaining service life and failure probability of the needle card and in combination with preset quantitative triggering rules. The maintenance and usage strategy decisions include at least the maintenance, downgrade or scrapping replacement suggestions for the needle card and the recommended execution time.

[0098] The management module 60 is used to implement WAT test probe card lifecycle management based on the maintenance and usage strategy decision, write back the maintenance operation results and performance retest data to the probe card digital twin data model, and update the machine learning regression model and the proportional risk survival model.

[0099] In this embodiment, multi-source data from the entire lifecycle of the probe card is collected and uniformly managed to construct a digital twin data model of the probe card; change point detection and segmentation are performed on the monitoring time-series data to build a stable segment library; a unified feature vector containing electrical contacts, test distribution, spatial structure, and lifespan load is extracted; the probe card health index is calculated based on a machine learning regression model; the remaining service life and failure probability are predicted using a proportional hazards survival model; maintenance decision suggestions and recommended execution times are generated based on the health index, remaining service life, and failure probability; and the model is updated through maintenance result feedback to achieve closed-loop optimization. Through the above methods, accurate assessment and predictive maintenance of the probe card health status are achieved.

[0100] This application embodiment also provides a WAT ​​test probe card lifecycle management device, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other through the communication bus. The memory is used to store the WAT test probe card lifecycle management program. When the processor executes the program stored in the memory, it implements the above-mentioned WAT test probe card lifecycle management method.

[0101] The communication bus mentioned in the WAT test probe card lifecycle management device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0102] The communication interface is used for communication between the aforementioned WAT test probe card lifecycle management device and other devices.

[0103] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0104] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0105] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0106] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0107] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0108] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0109] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0110] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0111] In addition, for technical details not described in detail in this embodiment, please refer to the WAT test probe card lifecycle management method provided in any embodiment of the present invention, which will not be repeated here.

[0112] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0113] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0115] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

[0116] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

Claims

1. A method for managing the lifecycle of a WAT ​​test probe card, characterized in that, The WAT test probe card lifecycle management method includes: Collect multi-source data related to the entire lifecycle of the probe card from at least one test machine, and perform unified governance and structured organization on the multi-source data to establish a probe card digital twin data model. The multi-source data includes at least probe card configuration data, test process monitoring timing data, wafer test result data, and probe card maintenance event data. Change point detection and segmentation are performed on at least one key parameter sequence in the time series data of the test process monitoring to form multiple time-continuous running segments. Multivariate summary features are extracted from each running segment, and a stable segment library is constructed based on the extracted multivariate summary features. Based on the test process monitoring timing data, wafer test result data and cumulative usage data of the pin card corresponding to each running segment in the stable segment library, a unified feature vector is constructed. The unified feature vector includes electrical and contact quality signal features, WAT test result distribution and yield risk signal features, spatial structure signal features, and lifetime and usage intensity signal features. The unified feature vector is input into the machine learning regression model to obtain the needle card health index for each running segment. The training sample labels of the machine learning regression model are defined based on the needle card life cycle events and production line performance in historical data. The unified feature vector is input into the proportional risk survival model to obtain the remaining service life of the probe card and the probability of failure within a specified future time window. Based on the needle card's health index, remaining lifespan, and failure probability, and combined with preset quantitative triggering rules, maintenance and usage strategy decisions are generated. The maintenance and usage strategy decisions include at least the needle card's maintenance, downgrade or scrapping replacement suggestions, and recommended execution time points. Based on the maintenance and usage strategy decision, WAT test probe card lifecycle management is implemented, maintenance operation results and performance retest data are written back to the probe card digital twin data model, and used to update the machine learning regression model and the proportional risk survival model; The extraction process of the spatial structure signal features includes: when there is a test value distribution map in units of grains, representing the test values, test residuals or normalized deviation values ​​as a two-dimensional matrix; performing at least one of the following operations on the two-dimensional matrix to generate numerical features: calculating spatial autocorrelation statistics, identifying and quantifying the size and intensity of local anomaly clusters, analyzing the gradient or bending trend in the row and column directions, and extracting texture energy distribution features through two-dimensional frequency domain transformation.

2. The WAT test probe card lifecycle management method as described in claim 1, characterized in that, The electrical and contact quality signal characteristics are calculated based on the contact resistance or leakage current time series data within the operating segment, including statistics describing the central trend, dispersion, tail risk, and changing trend of the data distribution; the WAT test result distribution and yield risk signal characteristics are calculated based on the result distribution of wafer test parameters within the operating segment, including indicators describing the distribution shape, dispersion, proportion of specification limits, and difference from historical benchmark distribution.

3. The WAT test probe card lifecycle management method as described in claim 1, characterized in that, The definition of the training sample labels includes marking the running segments that meet the preset yield standard and do not experience any needle card-related failure events within a preset future observation window as first health status samples; and marking the running segments that are located within a preset warning window before the occurrence of needle card-related maintenance or scrapping events, or the running segments whose test indicators continue to deteriorate and have not recovered after maintenance, as second health status samples.

4. The WAT test probe card lifecycle management method as described in claim 1, characterized in that, The proportional hazards survival model is a deep learning-based Cox proportional hazards model; the training samples of the model are constructed in the form of triples, and the unified feature vector of the corresponding running segment is used as the time-varying covariate within the time interval; wherein, the triple form includes the segment start time, the segment end time and the event marker, and the failure event in the event marker is defined as a preset event type that causes the needle card to be terminated, including needle card scrapping, key component replacement and maintenance failure.

5. The WAT test probe card lifecycle management method as described in claim 1, characterized in that, The preset quantitative triggering rules are as follows: when the health index of the needle card is continuously lower than the health threshold and the remaining service life is shorter than the maintenance lead time, a maintenance recommendation is triggered; when the failure probability within a specified future time window exceeds the risk threshold, a maintenance or replacement recommendation is triggered; when the proportion of recently run segments identified as abnormal exceeds the abnormality threshold, a downgrade recommendation is triggered.

6. The WAT test probe card lifecycle management method as described in any one of claims 1 to 5, characterized in that, The method further includes: Obtain the set of future available maintenance periods from the manufacturing execution system; Determine the comprehensive decision cost for each available maintenance period, wherein the comprehensive decision cost includes the production cost of downtime during that period and the expected risk cost corresponding to delayed maintenance, and the expected risk cost is calculated based on the failure probability; The available maintenance period with the lowest overall decision-making cost is selected as the recommended execution time.

7. A WAT test probe card lifecycle management system, characterized in that, The WAT test probe card lifecycle management system is applied to the WAT test probe card lifecycle management method as described in any one of claims 1 to 6, the system comprising: The acquisition module is used to acquire multi-source data related to the entire life cycle of the probe card from at least one test machine, and to perform unified governance and structured organization of the multi-source data to establish a probe card digital twin data model. The multi-source data includes at least probe card configuration data, test process monitoring timing data, wafer test result data, and probe card maintenance event data. The processing module is used to perform change point detection and segmentation on at least one key parameter sequence in the time series data of the test process monitoring, form multiple time-continuous running segments, extract multivariate summary features from each running segment, and construct a stable segment library based on the extracted multivariate summary features. The construction module is used to construct a unified feature vector based on the test process monitoring timing data, wafer test result data and cumulative usage data of the pin card corresponding to each running segment in the stable segment library. The unified feature vector includes electrical and contact quality signal features, WAT test result distribution and yield risk signal features, spatial structure signal features and lifetime and usage intensity signal features. The input module is used to input the unified feature vector into the machine learning regression model to obtain the needle card health index for each running segment. The training sample labels of the machine learning regression model are defined based on the needle card life cycle events and production line performance in historical data. The input module is used to input the unified feature vector into the proportional risk survival model to obtain the remaining service life of the probe card and the probability of failure within a specified time window in the future. The generation module is used to generate maintenance and usage strategy decisions based on the health index, remaining service life and failure probability of the needle card and in combination with preset quantitative triggering rules. The maintenance and usage strategy decisions include at least the maintenance, downgrade or scrapping replacement suggestions for the needle card and the recommended execution time. The management module is used to implement WAT test probe card lifecycle management based on the maintenance and usage strategy decisions, write maintenance operation results and performance retest data back to the probe card digital twin data model, and update the machine learning regression model and the proportional risk survival model.

8. A WAT test probe card lifecycle management device, characterized in that, The WAT test probe card lifecycle management device includes: a memory, a processor, and a WAT ​​test probe card lifecycle management program stored in the memory and executable on the processor, wherein the WAT test probe card lifecycle management program is configured to implement the steps of the WAT test probe card lifecycle management method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a WAT ​​test probe card lifecycle management program, which, when executed by a processor, implements the steps of the WAT test probe card lifecycle management method as described in any one of claims 1 to 6.

Images