Semiconductor wat test data real-time compensation method and system

Abstract

The present application belongs to the technical field of semiconductor testing, and discloses a semiconductor WAT test data real-time compensation method and system. The method comprises: collecting WAT original measurement values and associated multi-dimensional test condition data in real time to form a time series data stream; outputting a clean data set and a data quality score after preprocessing; selecting an optimal historical segment as a compensation reference based on a stability index; inputting die inherent characteristics and test conditions to train a gradient boosting tree model to obtain a difference model; calculating a predicted difference value between the current and reference conditions as a systematic offset through counterfactual reasoning; and generating a corrected measurement value and outputting the same in combination with a confidence weight. The present application realizes automatic, accurate and real-time compensation and interpretable attribution of systematic offsets caused by multiple test conditions and their interactions, significantly improves test data consistency and reduces manual intervention costs.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing technology, and in particular to a method and system for real-time compensation of semiconductor WAT test data. Background Technology

[0002] In the semiconductor wafer manufacturing process, WAT (Wafer Acceptance Test) is a critical step for process verification and electrical parameter screening. It involves using probe cards to directly contact the die and measure various electrical parameters. In actual production, test results are affected not only by wafer process variations but also by systematic interference from various test conditions, including probe station temperature drift, changes in probe card contact resistance, differences in power board configuration, individual machine deviations, and differences in test program versions. These variations can lead to systematic measurement gaps within the same process batch under different conditions, affecting parameter distribution consistency, risk assessment near specification boundaries, and subsequent yield analysis.

[0003] Traditional compensation methods primarily rely on manual processing by engineers: calculating the average offset of each test item under different combinations of conditions using historical data, then manually creating a compensation table in Excel or a script, and adjusting the measured values ​​or spec limits item by item. Recent practices of this method include some factories introducing simple scripts to automate single-condition statistics, but still require manual specification of condition boundaries and baselines.

[0004] The main drawbacks of existing technologies include: (1) Extremely low efficiency and serious time consumption: Every time the conditions change (such as replacing the probe card, adjusting the temperature setting, or upgrading the power distribution panel), it is necessary to re-screen historical data, manually count the differences and verify them. A single complete calibration often takes several hours to several days, which is difficult to meet the needs of high-speed mass production. (2) Insufficient accuracy: Manual statistics are difficult to fully capture the nonlinear interaction effects between conditions (such as the amplification effect of the probe card contact resistance when the temperature rises), which can easily lead to incomplete compensation or overcompensation, with residual offset still reaching several percentage points. (3) Poor consistency and weak traceability: Different engineers have different choices of benchmark segments and statistical standards (such as window size and filtering rules), which leads to increased deviations in the judgment standards of the same product at different time periods or on different production lines, increasing the risk of misjudgment and batch release. (4) Poor adaptability and difficulty in expansion: When faced with new conditions (such as new probe cards or new temperature ranges), there is a lack of sufficient historical data, making it impossible to quickly establish a reliable compensation amount. It can only rely on experience to estimate or temporarily not compensate, which leads to a decrease in the stability of the transition period test. (5) Lack of interpretability: Traditional methods only output the average offset, which cannot explain which conditions or combinations of interactions cause the difference, which is not conducive to problem tracing and process optimization.

[0005] 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

[0006] The main objective of this invention is to provide a real-time compensation method and system for semiconductor WAT test data, aiming to solve the technical problems in the prior art where systematic offsets in WAT test data caused by multiple test conditions and their interactions rely on manual compensation, resulting in low efficiency, insufficient accuracy, and a lack of interpretability.

[0007] To achieve the above objectives, the present invention provides a real-time compensation method for semiconductor WAT test data, the method comprising the following steps: The raw measurement values ​​of the wafer under test (WAT) and the associated multidimensional test condition data are collected in real time from the test equipment, equipment logs and sensor network to form a time-series data stream. The time-series data stream includes timestamps, measurement values ​​and condition vectors. The multidimensional test conditions include at least probe card identifier, equipment identifier, power distribution panel identifier, test program version, probe station set temperature and actual temperature. The time-series data stream is preprocessed to output a clean dataset corresponding to each data record and a data quality score that characterizes the credibility of the record. The data preprocessing includes timestamp correction, missing value imputation, encoding standardization, and outlier detection. Based on the measurement sequence of key test items in the cleaning dataset, a change point detection algorithm is used to automatically identify the boundaries where test conditions change explicitly or implicitly, divide the data into different condition segments, and assign a unique segment identifier to each segment. Based on the clean dataset and the fragment identifier, the stability score of each historical condition fragment is determined according to the preset stability index. The fragment with the highest score is selected as the compensation benchmark, a unique benchmark identifier is assigned, and the condition distribution summary and time range corresponding to the fragment with the highest score are frozen. Using the inherent characteristics of the tube core in the cleaning dataset and the multidimensional test condition data as input features, and the original measurement value as the prediction target, a gradient boosting decision tree model is trained to obtain a difference model for predicting the expected measurement value under any test condition. For the current die to be calibrated, the difference between the model prediction under the current actual test conditions and the benchmark condition under the benchmark condition is calculated by using the difference model and the benchmark condition vector corresponding to the benchmark identifier through counterfactual reasoning, and is used as the systematic offset to be compensated. The corrected measurement value is generated and output based on the systematic offset and the confidence weight determined based on the data quality score to complete real-time compensation.

[0008] In one embodiment, the change point detection algorithm employs the Page-Hinkley algorithm, and the formula for updating and determining the cumulative deviation value is as follows:

[0009] in, The cumulative deviation value up to time t. This is the cumulative deviation value from the previous moment. The current observation value at time t, The cumulative average up to time t. The minimum deviation threshold, As the trigger threshold, The minimum cumulative deviation is used to detect negative drift. When the minimum cumulative deviation min(PHt) is lower than the trigger threshold... When λ, it is determined that a change point occurs at time t.

[0010] In one embodiment, the counterfactual reasoning formula corresponding to the systematic offset is:

[0011] in, This is the predicted output of the LightGBM model. The intrinsic feature vector of the die being tested. This is the current actual test condition vector. This is the baseline condition vector.

[0012] In one embodiment, the method further includes: The difference model is decomposed using SHAP TreeExplainer to generate attribution values ​​for the contribution of each test condition feature to the systematic offset.

[0013] In one embodiment, the formula for the corrected measurement value is: Correction formula:

[0014] in, These are the original measured values. For the calculated systematic offset, For confidence level weights.

[0015] In one embodiment, the method further includes: When a new test condition with insufficient historical data is identified, cluster centers similar to the condition are used as surrogate conditions for prediction, and the corresponding confidence weights are reduced. Once the sample size under the new test conditions reaches a preset threshold, incremental training is enabled to update the difference model and restore the confidence weights.

[0016] In one embodiment, the method further includes: Continuously monitor the prediction error, data distribution drift index, and correction benefit index of the difference model; When the prediction error, data distribution drift index, and correction benefit index exceed the preset threshold, the compensation benchmark identifier is automatically switched or a historical model version is loaded, and the switching event is recorded.

[0017] Furthermore, to achieve the above objectives, the present invention also proposes a real-time compensation system for semiconductor WAT test data, which is applied to the real-time compensation method for semiconductor WAT test data described above. The system includes: The data acquisition module is used to collect raw measurement values ​​and associated multidimensional test condition data of the wafer under test (WAT) in real time from the test equipment, equipment logs and sensor network to form a time-series data stream. The time-series data stream includes timestamps, measurement values ​​and condition vectors. The multidimensional test conditions include at least probe card identifier, equipment identifier, power distribution panel identifier, test program version, probe station set temperature and actual temperature. The data preprocessing module is used to preprocess the time-series data stream and output a clean dataset corresponding to each data record and a data quality score that characterizes the credibility of the record. The data preprocessing includes timestamp correction, missing value imputation, encoding standardization and outlier detection. The change point detection module is used to automatically identify the boundaries where test conditions change explicitly or implicitly based on the measurement value sequence of key test items in the cleaning dataset, using a change point detection algorithm, dividing the data into different condition segments, and assigning a unique segment identifier to each segment. The condition recognition module is used to determine the stability score of each historical condition segment based on the clean dataset and the segment identifier, according to a preset stability index, select the segment with the highest score as the compensation benchmark, assign a unique benchmark identifier, and freeze the condition distribution summary and time range corresponding to the segment with the highest score. The benchmark definition module is used to take the inherent characteristics of the tube core in the cleaning dataset and the multidimensional test condition data as input features, and the original measurement value as the prediction target. It is trained using a gradient boosting decision tree model to obtain a difference model for predicting the expected measurement value under any test condition. The difference calculation module is used to calculate the difference between the model prediction value under the current actual test conditions and the benchmark conditions for the current die to be calibrated, using the difference model and the benchmark condition vector corresponding to the benchmark identifier, through counterfactual reasoning, as the systematic offset to be compensated. The real-time correction module is used to generate and output corrected measurement values ​​based on the systematic offset and the confidence weight determined based on the data quality score, so as to complete real-time compensation.

[0018] Furthermore, to achieve the above objectives, the present invention also proposes a real-time compensation device for semiconductor WAT test data. The real-time compensation device for semiconductor WAT test data includes: a memory, a processor, and a real-time compensation program for semiconductor WAT test data stored in the memory and executable on the processor. The real-time compensation program for semiconductor WAT test data is configured to implement the steps of the real-time compensation method for semiconductor WAT test data as described above.

[0019] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a real-time compensation program for semiconductor WAT test data, wherein the real-time compensation program for semiconductor WAT test data, when executed by a processor, implements the steps of the real-time compensation method for semiconductor WAT test data as described above.

[0020] This invention acquires raw WAT measurements and associated multidimensional test condition data in real time, forming a time-series data stream; after preprocessing, it outputs a clean dataset and data quality scores; based on stability indicators, it selects the optimal historical segment as the compensation benchmark; using the inherent characteristics of the die and test conditions as input, it trains a gradient boosting tree model to obtain a difference model; it calculates the prediction difference between the current and benchmark conditions as a systematic offset through counterfactual reasoning; and it generates and outputs corrected measurements by combining confidence weights. This invention achieves automatic, accurate, and real-time compensation and interpretable attribution for systematic offsets caused by multiple test conditions and their interactions, significantly improving test data consistency and reducing the cost of manual intervention. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the first embodiment of the semiconductor WAT test data real-time compensation method of the present invention; Figure 2 This is a structural block diagram of the first embodiment of the semiconductor WAT test data real-time compensation system of the present invention.

[0022] 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

[0023] 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.

[0024] This invention provides a method for real-time compensation of semiconductor WAT test data, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of a semiconductor WAT test data real-time compensation method according to the present invention.

[0025] In this embodiment, the real-time compensation method for semiconductor WAT test data includes the following steps: Step S10: Collect raw measurement values ​​of the wafer under test (WAT) and associated multidimensional test condition data in real time from the test equipment, equipment logs, and sensor network to form a time-series data stream.

[0026] In this embodiment, the executing entity is a semiconductor WAT test data real-time compensation device. This semiconductor WAT test data real-time compensation device has functions such as data processing, data communication, and program execution. The semiconductor WAT test data real-time compensation device can be a computer terminal device or other network device, or other devices with similar functions. This embodiment does not limit the scope of the device.

[0027] It should be noted that traditional compensation methods primarily rely on manual processing by engineers: calculating the average offset of each test item under different combinations of conditions through historical data statistics, then manually creating a compensation table in Excel or a script, and adjusting the measured values ​​or spec limits item by item. Recent practices of this method include some factories introducing simple scripts to automate single-condition statistics, but still require manual specification of condition boundaries and baselines. The main drawbacks of existing technologies include: extremely low efficiency and significant time consumption: each change in conditions (such as replacing probe cards, adjusting temperature settings, or upgrading distribution panels) requires re-screening historical data, manually calculating differences, and verifying them. A single complete calibration often takes several hours to several days, making it difficult to meet the demands of high-volume production. Insufficient accuracy: manual statistics cannot fully capture the nonlinear interaction effects between conditions (such as the amplification effect of probe card contact resistance when the temperature rises), easily leading to incomplete or over-compensation, with residual offsets still reaching several percentage points. Poor consistency and weak traceability: different engineers have inconsistent choices of reference segments and statistical methods (such as window size and filtering rules), resulting in increased deviations in the judgment standards for the same product at different time periods or on different production lines, increasing the risk of misjudgment and batch release. Poor adaptability and difficulty in expansion: when facing newly introduced conditions (such as new probe cards or new temperature ranges), there is a lack of sufficient historical data, making it impossible to quickly establish reliable compensation amounts. It is necessary to rely on experience for estimation or temporarily not compensate, leading to decreased stability during the transition period. Lack of interpretability: traditional methods only output the average offset, which cannot explain which conditions or interaction combinations cause the differences, making it difficult to trace the source of problems and optimize the process.

[0028] To address the aforementioned technical issues, this embodiment acquires raw WAT measurements and associated multi-dimensional test condition data in real time, forming a time-series data stream. After preprocessing, a clean dataset and data quality score are output. The optimal historical segment is selected as the compensation benchmark based on stability indicators. A gradient boosting tree model is trained using the inherent characteristics of the die and test conditions as input to obtain a difference model. The difference between the prediction under the current and benchmark conditions is calculated using counterfactual reasoning as a systematic offset. Corrected measurements are generated and output by combining confidence weights. This invention achieves automatic, accurate, and real-time compensation and interpretable attribution for systematic offsets caused by multiple test conditions and their interactions, significantly improving test data consistency and reducing manual intervention costs. Specifically, it can be implemented as follows.

[0029] In practical implementation, the method proposed in this embodiment can be deployed on the central data server, edge computing nodes of test equipment, or cloud platform of semiconductor manufacturing enterprises. It achieves centralized real-time compensation of WAT test data from multiple machines and under multiple conditions through software modules. The technical solution realizes the complete compensation process through the following eight sequentially executed, closed-loop steps. The core lies in transforming the conditional differences from "manual statistics" to a fully automated chain of "automatic detection—machine learning modeling—counterfactual difference calculation—real-time correction—online adaptation."

[0030] In this embodiment, the system functional modules are as follows: Data acquisition module: aggregates data from the test result library, device logs, sensor and version library and generates condition vectors; Data preprocessing module: performs time correction, missing data handling, coding standardization, outlier labeling and quality scoring; Condition identification module: generates Segment_ID based on change point detection and anomaly detection and locates condition changes; Baseline management module: selects Baseline_ID for historical segment scoring, supports freezing, switching, rollback and auditing; Difference modeling module: trains regression model f(\u00b7) and manages model versions; Difference calculation and interpretation module: calculates Δ and outputs feature contribution (such as SHAP); Correction output module: generates y_corr and correction benefit index according to configuration; Online monitoring and rollback module: monitors error / drift / change point frequency and triggers deweighting, pause, rollback or retraining.

[0031] Furthermore, the overall process involves the system continuously aggregating test results and full context conditions to form a high-dimensional time-series data stream; ensuring data quality through rigorous preprocessing; automatically identifying conditional change points and latent anomalies using statistical learning; automatically selecting the most stable historical period as a unified benchmark; explicitly learning conditional main effects and higher-order interactions using a gradient boosting tree model; calculating accurate differences based on counterfactual inference and providing SHAP-interpretable attribution; applying differences in real time to correct measured values ​​and quantifying correction benefits; and finally, achieving incremental model updates and anomaly rollback through online monitoring of multiple indicators to ensure long-term mass production stability and controllable risks.

[0032] In this embodiment, the time-series data stream includes timestamps, measured values, and condition vectors. Multi-dimensional test conditions include at least probe card identifiers, machine identifiers, distribution panel identifiers, test program versions, probe station set temperature, and actual temperature. Specifically, the system establishes a high-throughput data acquisition module, employing a combination of Kafka streaming and timed batch ETL to support second-level real-time performance and historical integrity. Data is continuously aggregated from the following sources: Test result database: item measured values, pass / fail, spec limits, wafer / lot / die coordinates, timestamps; Equipment logs and sensors: real-time probe station temperature sequence (sampled every minute), probe card ID and cumulative number of used pins, distribution panel configuration ID, machine ID, program version; Auxiliary systems: ambient temperature and humidity, contact resistance monitoring (if applicable), and other potential influencing factors. Specifically, a structured record of "measured value—condition vector—timestamp" is established using the waferID as the primary key, ensuring that each die-level data entry is associated with a complete condition snapshot. A high-dimensional time-series dataset is output, supporting subsequent minute-level online processing.

[0033] Step S20: Perform data preprocessing on the time-series data stream, and output the clean dataset corresponding to each data record and the data quality score that characterizes the credibility of the record.

[0034] In this embodiment, data preprocessing includes timestamp correction, missing value imputation, code standardization, and outlier detection. Specifically, multi-layer cleaning and standardization are performed; timestamps are unified to UTC and machine clock deviations are corrected; continuous conditions (such as temperature) are standardized in units, and categorical conditions (such as probe card IDs) are mapped to unified codes; missing values ​​are handled: continuous values ​​are filled with the rolling median, and categorical values ​​are filled with the mode or forward imputation; outliers are marked using an improved 3σ (quantile robust version) + business rules (such as physical limits), retaining the original values ​​but adding weight flags (0.1-1.0); deduplication and consistency checks are performed to ensure die-level uniqueness. A clean dataset is output, with a quality score appended to each record.

[0035] Step S30: Based on the measurement value sequence of key test items in the cleaning dataset, the change point detection algorithm is used to automatically identify the boundaries where the test conditions change explicitly or implicitly, divide the data into different condition segments, and assign a unique segment identifier to each segment.

[0036] In its implementation, the change point detection algorithm adopts the Page-Hinkley algorithm, and the formula for updating and determining the cumulative deviation value is as follows:

[0037] in, The cumulative deviation value up to time t. This is the cumulative deviation value from the previous moment. The current observation value at time t, The cumulative average up to time t. The minimum deviation threshold, As the trigger threshold, The minimum cumulative deviation is used to detect negative drift. When the minimum cumulative deviation min(PHt) is lower than the trigger threshold... When λ is used, a change point is determined to occur at time t. Here, the parameter δ = 0.5σ, and λ is dynamically adjusted. Simultaneously, IsolationForest (contamination = 0.05) is run within the segment to detect latent anomalies. The output is an automatically labeled conditional segment sequence, supporting dual recognition of explicit switching (e.g., card replacement) and latent drift (e.g., progressive contamination).

[0038] Step S40: Based on the clean dataset and the fragment identifier, determine the stability score of each historical condition fragment according to the preset stability index, select the fragment with the highest score as the compensation benchmark, assign a unique benchmark identifier, and freeze the condition distribution summary and time range corresponding to the fragment with the highest score.

[0039] It should be noted that the historical segments are evaluated using a multi-metric comprehensive score: lowest MAD fluctuation, most stable KS distribution, anomaly rate <3%, and sample size >5000 dies. The highest-scoring segment is prioritized as the benchmark, and stability is verified periodically (weekly). If drift occurs, a new benchmark is automatically switched and a change log is recorded. Once a benchmark segment is selected, a Baseline_ID is generated and frozen. The frozen content includes: segment time range, conditional distribution summary (mean / quantiles / class frequency), screening rules, statistical indicators (MAD / KS / anomaly rate / sample size), and configuration hash. The preferred benchmark switching trigger condition is: KS < 0.05 or PSI > 0.1 for k consecutive batches (e.g., k=3) with a continuously increasing residual error. Before switching, the system compares the correction gains of the old / new benchmarks in parallel in shadow mode; the switch is only formally initiated after the gains stabilize and become positive. If the gains are negative after switching or a high-risk alarm is triggered, the system automatically rolls back to the previous Baseline_ID and records the rollback event.

[0040] Step S50: Using the inherent characteristics of the tube core in the cleaning dataset and the multidimensional test condition data as input features, and the original measurement value as the prediction target, a gradient boosting decision tree model is trained to obtain a difference model for predicting the expected measurement value under any test condition.

[0041] In the specific implementation, LightGBM regression is adopted, and the input features are divided into: the features of the tested object: die coordinates (x,y), lot process label, wafer center / edge flag; and the condition features: temperature (numerical value), probe card / distribution panel / machine (one-hot or target encoding to prevent high cardinality).

[0042] The objective function is measurement regression, supporting L1 / L2 regularization. Tree splitting naturally captures interactions (e.g., amplified probe card A offset when temperature > 30℃). Training uses a hierarchical dataset (by product family tree), with cross-validation optimizing depth / learning rate. Warm-start incremental training is supported when new conditions arise.

[0043] When a new ProbeCard_ID / PowerBoard_ID / Program_Version or a new temperature range is detected and the training sample size is insufficient, the system enters a cold start strategy: it preferentially uses cluster centers with similar conditions as surrogate conditions (e.g., clustering by machine family / probe card model / temperature range) and reduces w_cov for conservative correction; incremental training is then enabled and w is restored to normal after the cumulative sample size of the new condition reaches n_min. This strategy ensures that overcorrection will not occur during periods of data sparsity.

[0044] Step S60: For the current die to be calibrated, using the difference model and the benchmark condition vector corresponding to the benchmark identifier, calculate the difference between the model prediction value under the current actual test conditions and the benchmark conditions through counterfactual reasoning, as the systematic offset to be compensated.

[0045] In this embodiment, the formula for the corrected measurement value is: Correction formula:

[0046] in, These are the original measured values. For the calculated systematic offset, For confidence level weights.

[0047] Furthermore, the explainable attribution is calculated using SHAP TreeExplainer, and the SHAP value decomposition is performed simultaneously:

[0048] in: This represents the contribution (positive or negative) of the i-th conditional feature (such as temperature or probe card ID) to the current predicted offset. The system outputs a dependency plot showing higher-order interaction effects (such as the superimposed offset of "temperature × probe card"), as well as the Top 5 dominant conditions, used to guide subsequent process adjustments and equipment maintenance. The output dependency plot shows the superimposed effect of "temperature × probe card", as well as the top 5 dominant conditions, used to guide subsequent process adjustments. Step S70: Generate and output the corrected measurement value based on the systematic offset and the confidence weight determined based on the data quality score to complete real-time compensation.

[0049] In practical implementation, the formula for the corrected measurement value is: Correction formula:

[0050] in, These are the original measured values. For the calculated systematic offset, For confidence level weights.

[0051] Furthermore, in this embodiment, when a new test condition with insufficient historical data is identified, a cluster center similar to the condition is used as a proxy condition for prediction, and the corresponding confidence weight is reduced; after the sample size of the new test condition reaches a preset threshold, incremental training is enabled to update the difference model and restore the confidence weight.

[0052] In one embodiment, the prediction error, data distribution drift index, and correction benefit index of the difference model are continuously monitored; when the prediction error, data distribution drift index, and correction benefit index exceed preset thresholds, the compensation benchmark identifier is automatically switched or a historical model version is loaded, and the switching event is recorded.

[0053] Specifically, the system includes at least three operating states: normal correction, deweighted correction, and paused correction, which only outputs suggested systematic offsets. The preferred state transition rules are: deweighted correction is entered when PSI > 0.1 or KS < 0.05; paused correction is entered when the correction gain is continuously negative, or MAE exceeds the threshold and the frequency of change points suddenly increases; and normal correction is returned after manual review or retraining verification is completed under paused correction.

[0054] Each status record includes trigger indicators, threshold versions, model versions, and a list of affected batches for auditing and traceability. Multi-dimensional monitoring thresholds are configurable: residual error (MAE) > threshold (e.g., 0.5% spec width); distribution drift (PSI) > 0.1 or KS < 0.05; sudden increase in change point frequency (> 2 times normal); abnormal segment ratio > 10%; negative correction gain (error increases after correction). Handling process when correction fails or is abnormal: (1) Immediately trigger multi-level alarm: system log recording + real-time pop-up notification to test engineer + email / SMS push to equipment maintenance manager; (2) Automatically pause the automatic correction function of the current model and switch to "only output original value + difference suggestion" mode to prevent the error correction from spreading to production judgment; (3) Data labeling: mark all measurement records in this period as "low confidence - correction abnormal" and isolate and store them for subsequent manual review or model retraining; (4) Manual intervention: the system generates a detailed abnormal report (including Top SHAP contribution, monitoring index curve, and list of affected wafers), and engineers can confirm rollback or forced update after manual verification; (5) Model rollback: automatically load the most recently verified stable model version (based on the lowest historical RMSE) and trigger incremental training to absorb the latest labeled data; (6) Post-event audit: record the complete link log of all abnormal events to support tracing the root cause (such as insufficient learning when a new probe card is introduced). Alternatives: LightGBM can be replaced with XGBoost or CatBoost (better for class feature processing); change point detection can be improved with the CUSUM algorithm; SHAP can supplement LIME for single-sample local interpretation.

[0055] Three specific embodiments are further illustrated. Embodiment 1 is a probe card replacement scenario: taking test item Item_A of product P1 as an example. A certain lot is being tested on machine T1. After the probe card is changed from PC_01 to PC_02, the mean of the original data y_raw shifts upward. In step 3, the system detects the change point and generates a new Segment_ID; in step 4, it selects a historically stable segment as the Baseline_ID; in step 6, it calculates... The SHAP output shows that "ProbeCard_ID" is the top contributor; step 7 outputs y_corr and sets w to 1 (sufficient sample coverage). After correction, compared with the golden lot, the mean difference decreases, KS consistency improves, and the proportion of tail excess risk (die) decreases. The entire process requires no manual table creation, and Baseline_ID, Model_Version, Δ, w, and decision mode are all audited and logged.

[0056] Example 2 illustrates temperature drift and latent anomalies: Under slight temperature control instability, Temp_Actual shows a slow upward trend without explicit card / board replacement. The system detects progressive drift using Page-Hinkley and uses Isolation Forest to mark latent anomaly segments; in S1 deweighted state, it outputs conservative correction (w decreases), and SHAP indicates that the "temperature" contribution is continuously increasing, triggering a maintenance suggestion to "check temperature control system calibration". After maintenance, the drift index recovers, and the system returns to normal S0 calibration.

[0057] Example 3 illustrates a cold start under new conditions: When a new PowerBoard_ID=PB_NEW is introduced, historical samples are insufficient. Following the cold start strategy, the system first employs a clustering proxy with similar configurations and reduces w_cov, outputting "Recommend Δ: Do not force application"; once the sample size reaches n_min, incremental training is enabled, w recovers, and normal correction begins. This process avoids miscalibration during the period of sparsity under new conditions.

[0058] In this embodiment, raw WAT measurements and associated multi-dimensional test condition data are acquired in real time to form a time-series data stream. After preprocessing, a clean dataset and data quality score are output. The optimal historical segment is selected as the compensation benchmark based on stability indicators. A gradient boosting tree model is trained using the inherent characteristics of the die and test conditions as input to obtain a difference model. The difference between the prediction under the current and benchmark conditions is calculated using counterfactual reasoning as a systematic offset. Corrected measurements are generated and output by combining confidence weights. This invention achieves automatic, accurate, and real-time compensation and interpretable attribution for systematic offsets caused by multiple test conditions and their interactions, significantly improving test data consistency and reducing manual intervention costs.

[0059] Furthermore, this embodiment of the invention also proposes a storage medium storing a real-time compensation program for semiconductor WAT test data. When the real-time compensation program for semiconductor WAT test data is executed by a processor, it implements the steps of the real-time compensation method for semiconductor WAT test data as described above.

[0060] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the semiconductor WAT test data real-time compensation system of the present invention.

[0061] like Figure 2 As shown, the real-time compensation system for semiconductor WAT test data proposed in this embodiment of the invention includes: The data acquisition module 10 is used to acquire the raw measurement values ​​of the wafer under test (WAT) and the associated multi-dimensional test condition data in real time from the test equipment, equipment logs and sensor network to form a time-series data stream. The time-series data stream includes timestamps, measurement values ​​and condition vectors. The multi-dimensional test conditions include at least probe card identifier, equipment identifier, power distribution panel identifier, test program version, probe station set temperature and actual temperature. Data preprocessing module 20 is used to preprocess the time-series data stream and output a clean dataset corresponding to each data record and a data quality score that characterizes the credibility of the record. The data preprocessing includes timestamp correction, missing value imputation, encoding standardization and outlier detection. The change point detection module 30 is used to automatically identify the boundaries where the test conditions change explicitly or implicitly based on the measurement value sequence of key test items in the cleaning dataset, using a change point detection algorithm, dividing the data into different condition segments, and assigning a unique segment identifier to each segment. The condition recognition module 40 is used to determine the stability score of each historical condition segment based on the clean dataset and the segment identifier, according to a preset stability index, select the segment with the highest score as the compensation benchmark, assign a unique benchmark identifier, and freeze the condition distribution summary and time range corresponding to the segment with the highest score. The benchmark definition module 50 is used to take the inherent characteristics of the tube core in the cleaning dataset and the multidimensional test condition data as input features, and the original measurement value as the prediction target, and train it using a gradient boosting decision tree model to obtain a difference model for predicting the expected measurement value under any test condition. The difference calculation module 60 is used to calculate the difference between the model prediction value under the current actual test conditions and the benchmark conditions under the current benchmark conditions for the current die to be calibrated, using the difference model and the benchmark condition vector corresponding to the benchmark identifier, as the systematic offset to be compensated. The real-time correction module 70 is used to generate and output the corrected measurement value based on the systematic offset and the confidence weight determined based on the data quality score, so as to complete the real-time compensation.

[0062] In this embodiment, raw WAT measurements and associated multi-dimensional test condition data are acquired in real time to form a time-series data stream. After preprocessing, a clean dataset and data quality score are output. The optimal historical segment is selected as the compensation benchmark based on stability indicators. A gradient boosting tree model is trained using the inherent characteristics of the die and test conditions as input to obtain a difference model. The difference between the prediction under the current and benchmark conditions is calculated using counterfactual reasoning as a systematic offset. Corrected measurements are generated and output by combining confidence weights. This invention achieves automatic, accurate, and real-time compensation and interpretable attribution for systematic offsets caused by multiple test conditions and their interactions, significantly improving test data consistency and reducing manual intervention costs.

[0063] This application embodiment also provides a semiconductor WAT test data real-time compensation 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 a semiconductor WAT test data real-time compensation program. When the processor executes the program stored in the memory, it implements the above-mentioned semiconductor WAT test data real-time compensation method.

[0064] The communication bus mentioned in the aforementioned semiconductor WAT test data real-time compensation 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.

[0065] The communication interface is used for communication between the aforementioned semiconductor WAT test data real-time compensation device and other devices.

[0066] 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.

[0067] 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.

[0068] 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)).

[0069] 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.

[0070] 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.

[0071] 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.

[0072] 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.

[0073] 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.

[0074] In addition, for technical details not described in detail in this embodiment, please refer to the real-time compensation method for semiconductor WAT test data provided in any embodiment of the present invention, which will not be repeated here.

[0075] 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.

[0076] 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.

[0077] 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.

[0078] 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.

[0079] 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 real-time compensation of semiconductor WAT test data, characterized in that, The real-time compensation method for semiconductor WAT test data includes: The raw measurement values ​​of the wafer under test (WAT) and the associated multidimensional test condition data are collected in real time from the test equipment, equipment logs and sensor network to form a time-series data stream. The time-series data stream includes timestamps, measurement values ​​and condition vectors. The multidimensional test conditions include at least probe card identifier, equipment identifier, power distribution panel identifier, test program version, probe station set temperature and actual temperature. The time-series data stream is preprocessed to output a clean dataset corresponding to each data record and a data quality score that characterizes the credibility of the record. The data preprocessing includes timestamp correction, missing value imputation, encoding standardization, and outlier detection. Based on the measurement sequence of key test items in the cleaning dataset, a change point detection algorithm is used to automatically identify the boundaries where test conditions change explicitly or implicitly, divide the data into different condition segments, and assign a unique segment identifier to each segment. Based on the clean dataset and the fragment identifier, the stability score of each historical condition fragment is determined according to the preset stability index. The fragment with the highest score is selected as the compensation benchmark, a unique benchmark identifier is assigned, and the condition distribution summary and time range corresponding to the fragment with the highest score are frozen. Using the inherent characteristics of the tube core in the cleaning dataset and the multidimensional test condition data as input features, and the original measurement value as the prediction target, a gradient boosting decision tree model is trained to obtain a difference model for predicting the expected measurement value under any test condition. For the current die to be calibrated, the difference between the model prediction under the current actual test conditions and the benchmark condition under the benchmark condition is calculated by using the difference model and the benchmark condition vector corresponding to the benchmark identifier through counterfactual reasoning, and is used as the systematic offset to be compensated. The corrected measurement value is generated and output based on the systematic offset and the confidence weight determined based on the data quality score to complete real-time compensation; The counterfactual reasoning formula corresponding to the systematic offset is: in, This is the predicted output of the LightGBM model. The intrinsic feature vector of the tested die, This is the current actual test condition vector. The baseline condition vector; The formula for the corrected measurement value is: Correction formula: in, These are the original measured values. For the calculated systematic offset, For confidence level weights.

2. The real-time compensation method for semiconductor WAT test data as described in claim 1, characterized in that, The change point detection algorithm uses the Page-Hinkley algorithm, and the formula for updating and determining the cumulative deviation value is as follows: in, The cumulative deviation value up to time t. This is the cumulative deviation value from the previous moment. The current observation value at time t, The cumulative average up to time t. The minimum deviation threshold, As the trigger threshold, The minimum cumulative deviation is used to detect negative drift. When the minimum cumulative deviation min(PHt) is lower than the trigger threshold... When λ, it is determined that a change point occurs at time t.

3. The real-time compensation method for semiconductor WAT test data as described in claim 1, characterized in that, The method further includes: The difference model is decomposed using SHAP TreeExplainer to generate attribution values ​​for the contribution of each test condition feature to the systematic offset.

4. The real-time compensation method for semiconductor WAT test data as described in claim 1, characterized in that, The method further includes: When a new test condition with insufficient historical data is identified, cluster centers similar to the condition are used as surrogate conditions for prediction, and the corresponding confidence weights are reduced. Once the sample size under the new test conditions reaches a preset threshold, incremental training is enabled to update the difference model and restore the confidence weights.

5. The real-time compensation method for semiconductor WAT test data as described in claim 1, characterized in that, The method further includes: Continuously monitor the prediction error, data distribution drift index, and correction benefit index of the difference model; When the prediction error, data distribution drift index, and correction benefit index exceed the preset threshold, the compensation benchmark identifier is automatically switched or a historical model version is loaded, and the switching event is recorded.

6. A real-time compensation system for semiconductor WAT test data, characterized in that, The semiconductor WAT test data real-time compensation system is applied to the semiconductor WAT test data real-time compensation method as described in any one of claims 1 to 5, the system comprising: The data acquisition module is used to collect raw measurement values ​​and associated multidimensional test condition data of the wafer under test (WAT) in real time from the test equipment, equipment logs and sensor network to form a time-series data stream. The time-series data stream includes timestamps, measurement values ​​and condition vectors. The multidimensional test conditions include at least probe card identifier, equipment identifier, power distribution panel identifier, test program version, probe station set temperature and actual temperature. The data preprocessing module is used to preprocess the time-series data stream and output a clean dataset corresponding to each data record and a data quality score that characterizes the credibility of the record. The data preprocessing includes timestamp correction, missing value imputation, encoding standardization and outlier detection. The change point detection module is used to automatically identify the boundaries where test conditions change explicitly or implicitly based on the measurement value sequence of key test items in the cleaning dataset, using a change point detection algorithm, dividing the data into different condition segments, and assigning a unique segment identifier to each segment. The condition recognition module is used to determine the stability score of each historical condition segment based on the clean dataset and the segment identifier, according to a preset stability index, select the segment with the highest score as the compensation benchmark, assign a unique benchmark identifier, and freeze the condition distribution summary and time range corresponding to the segment with the highest score. The benchmark definition module is used to take the inherent characteristics of the tube core in the cleaning dataset and the multidimensional test condition data as input features, and the original measurement value as the prediction target. It is trained using a gradient boosting decision tree model to obtain a difference model for predicting the expected measurement value under any test condition. The difference calculation module is used to calculate the difference between the model prediction value under the current actual test conditions and the benchmark conditions for the current die to be calibrated, using the difference model and the benchmark condition vector corresponding to the benchmark identifier, through counterfactual reasoning, as the systematic offset to be compensated. The real-time correction module is used to generate and output corrected measurement values ​​based on the systematic offset and the confidence weight determined based on the data quality score, so as to complete real-time compensation.

7. A real-time compensation device for semiconductor WAT test data, characterized in that, The semiconductor WAT test data real-time compensation device includes: a memory, a processor, and a semiconductor WAT test data real-time compensation program stored in the memory and executable on the processor, wherein the semiconductor WAT test data real-time compensation program is configured to implement the steps of the semiconductor WAT test data real-time compensation method as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The storage medium stores a real-time compensation program for semiconductor WAT test data, which, when executed by a processor, implements the steps of the real-time compensation method for semiconductor WAT test data as described in any one of claims 1 to 5.

Images