Intelligent traceability method, system and device for wat parameter anomaly and storage medium

Abstract

This invention belongs to the field of semiconductor manufacturing technology and discloses an intelligent cause-tracing method, system, device, and storage medium for WAT parameter anomalies. The method includes: collecting WAT test-related metadata from multiple systems; performing unified identification and time-series alignment; evaluating data reliability based on test link status; detecting anomalies and generating abnormal events according to reliability constraints; aggregating multi-source evidence to construct an event-level heterogeneous correlation graph; and using a pre-trained graph neural network model for graph reasoning to output root cause candidates, confidence levels, and traceable reasoning paths. Through these methods, automatic differentiation between test contamination and process anomalies is achieved, improving the accuracy, interpretability, and processing efficiency of cause tracing.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and in particular to a method, system, device, and storage medium for intelligent cause tracing of WAT parameter anomalies. Background Technology

[0002] Wafer Acceptance Test (WAT) is typically performed at critical points in the wafer manufacturing process. It involves electrical measurements on pre-designed test structures on the wafer to obtain key parameters reflecting device and process status, and to generate statistical distributions at the lot, wafer, and die / pad / area levels. These electrical parameters may include, but are not limited to, threshold voltage, leakage current, breakdown voltage, chip resistance, contact resistance, capacitance, and interconnect resistance, and are used to evaluate process stability, device consistency, and process deviation trends.

[0003] In actual production, WAT data is often combined with upper and lower limit specifications (USL / LSL), statistical process control (SPC) rules, and historical baseline models to identify anomalies such as exceeding limits, drift, increased variance, spatial clustering, and cross-parameter co-occurrence. Once an anomaly occurs, it is necessary to complete the qualitative analysis and root cause localization of the anomaly within a short period of time in order to take measures such as isolating batches, retesting and verifying, checking and calibrating equipment, maintaining probe cards, or locking down process steps for investigation, so as to avoid the anomaly from spreading and causing a decrease in yield, increased rework, or batch risk.

[0004] However, WAT anomalies do not necessarily correspond to manufacturing process anomalies. Because the measurement chain involves testing equipment, probe stations, probe cards, and environmental factors such as temperature and humidity, issues like poor contact, tip contamination / wear, instrument drift, and environmental deviations can trigger "measurement contamination anomalies." These anomalies are similar in data presentation to "manufacturing anomalies" such as actual process deviations, making them easily confused. Meanwhile, WAT results, equipment logs / EAP, FDC / measurement, yield / inspection data are scattered across different systems, and timestamp calibers, object identifiers, and sampling granularity are inconsistent, making it difficult to quickly establish a unified time-series correlation and evidence chain after anomalies occur. Therefore, relying solely on threshold alarms or manual cross-system troubleshooting often fails to promptly and accurately distinguish the source of anomalies and locate the root cause, resulting in long investigation cycles, high misjudgment rates, and inconsistent handling.

[0005] Based on this, existing factories typically adopt the following technical approach:

[0006] (1) Parameter threshold / statistical process control (SPC) judgment: set upper and lower limits or control chart rules (such as mean drift, 3σ out of bounds) for a single WAT parameter and issue an alarm;

[0007] (2) Manual cross-checking: Maintenance / testing / process engineers query logs and data in multiple systems such as the testing system, equipment engineering automation system (EAP), fault detection and classification system (FDC), metering system, and yield system, and establish correlations based on experience;

[0008] (3) Local rule base: Set rules to trigger a small number of known problems (such as probe card lifespan expiration, machine alarm) and output troubleshooting suggestions.

[0009] The aforementioned prior art has at least the following shortcomings:

[0010] (1) The multi-source data is stored in a scattered manner and lacks a unified time alignment method. The WAT results are difficult to automatically correlate with the machine status, probe station environment, probe card health indicators, upstream and downstream processes and measurement data on the same time axis, resulting in a long positioning cycle.

[0011] (2) Test system problems and manufacturing process defects are easily confused. Relying solely on WAT results or a single system log makes it difficult to distinguish between "measurement contamination" and "actual process anomalies," which can easily lead to misjudgment and incorrect handling.

[0012] (3) The rule base has limited coverage, insufficient generalization ability when facing new products, new processes or new types of faults, and high rule maintenance costs;

[0013] (4) The root cause conclusions lack quantitative credibility and traceable reasoning links, making it difficult to support auditing and continuous improvement.

[0014] Therefore, a technical solution is needed that can perform multi-source data fusion, credibility assessment, and intelligent cause tracing for WAT parameter anomalies in order to improve the efficiency of anomaly identification and location.

[0015] 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

[0016] The main objective of this invention is to provide an intelligent cause-tracing method, system, device, and storage medium for WAT parameter anomalies, aiming to solve the technical problems in the prior art, such as WAT anomaly localization relying on manual methods, data being scattered and difficult to align, test contamination being easily confused with process anomalies, and the lack of an interpretable cause-tracing link.

[0017] To achieve the above objectives, the present invention provides an intelligent method for tracing the causes of WAT parameter anomalies, the method comprising the following steps:

[0018] Metadata is collected from multiple independent systems associated with WAT testing. The metadata includes at least one or more of the following: WAT parameter test results, test equipment status data, probe station status and environmental data, probe card health index data, and Fault Detection and Classification (FDC) data, measurement data, and yield data during the manufacturing process.

[0019] The metadata is uniformly identified and time-series aligned to generate a set of associated data within a unified time window;

[0020] The associated data set is subjected to data quality checks, and a credibility score is output for each WAT record based on the test link status information, wherein the test link status information includes at least the test machine status, probe card health indicators and environmental data;

[0021] Based on the credibility score, the WAT parameters are subjected to constrained anomaly detection. When an anomaly is detected, an anomaly event is generated, wherein the anomaly event includes anomaly parameters, location information, anomaly type, and anomaly intensity characteristics.

[0022] Centered on the abnormal event, relevant data within a preset time window and spatial window are extracted from the associated data set as multi-source evidence. Based on the multi-source evidence and the characteristics of the abnormal event, an event-level heterogeneous association graph is constructed. The nodes of the event-level heterogeneous association graph include at least anomaly phenomenon nodes representing the abnormal event, state nodes representing the test link state, and multiple root cause candidate nodes. The nodes are connected by edges based on temporal adjacency, attribution relationship, or engineering prior knowledge.

[0023] The event-level heterogeneous correlation graph is input into a pre-trained fault diagnosis model, which performs graph reasoning calculations and outputs one or more root cause candidates, the confidence level of each root cause candidate, and the reasoning path from the abnormal phenomenon node to the root cause candidate node to complete the cause tracing.

[0024] In one embodiment, the step of performing a data quality check on the associated data set and outputting a credibility score for each WAT record based on the test link status information includes:

[0025] Identify potential data contamination records in the associated data set caused by poor test link status. The identification criteria for potential data contamination include probe contact resistance exceeding limits, abnormal machine calibration status, environmental parameters exceeding limits, or anomaly detection results of multidimensional features based on isolated forest or single-class support vector machine models.

[0026] Based on the identification results of potential data contamination and the severity of associated alarms, a credibility score is output for each WAT record.

[0027] In one embodiment, the constraint anomaly detection is to raise the anomaly determination threshold of the corresponding WAT parameter record or mark the WAT parameter record as pending retesting when the confidence score is lower than a preset threshold.

[0028] In one embodiment, the step of extracting relevant data within a preset time window and spatial window as multi-source evidence from the associated data set, centered on the abnormal event, includes:

[0029] Based on the event identifier of the abnormal event, determine the associated time window and associated spatial range for retrieving evidence;

[0030] Pull multi-source evidence data related to the abnormal event within the associated time window and associated spatial range from the associated data set. The multi-source evidence data includes at least WAT measurement sequence data, test machine alarm logs, probe card health timing indicators, environmental monitoring data, FDC statistical characteristics of associated process steps, and associated measurement or yield trend data.

[0031] In one embodiment, constructing an event-level heterogeneous correlation graph based on the multi-source evidence and the characteristics of the anomalous event includes:

[0032] The abnormal events are subjected to structured analysis to generate an abnormal profile, which includes at least the determination of the abnormal type, the quantitative value of the abnormal intensity, the spatial morphology description of the abnormality, and the results of the retest sensitivity analysis.

[0033] Based on the anomaly profile, the multi-source evidence data, and the features of the anomaly event, the event-level heterogeneous correlation graph is instantiated. The nodes of the heterogeneous correlation graph further include: feature nodes representing specific features in the anomaly profile, tool status nodes representing specific process equipment, and monitoring nodes representing measurement or yield deviation. The edges of the heterogeneous correlation graph also include statistical correlation edges generated based on the event co-occurrence relationship or statistical correlation in the multi-source evidence data.

[0034] In one embodiment, the fault diagnosis model is a model built based on a graph neural network (GNN). The pre-training process involves training the GNN model using historical labeled samples. Each historical labeled sample includes an event-level heterogeneous association graph constructed based on historical abnormal events, and a final root cause label confirmed by engineering. The goal of the training is to enable the model to learn to aggregate the information of abnormal phenomenon nodes and associated state nodes in the graph to root cause candidate nodes through a message passing mechanism, and output a root cause prediction probability distribution consistent with the final root cause label.

[0035] In one embodiment, the method further includes:

[0036] The root cause candidates, the confidence levels, and the reasoning paths are presented in the form of reports, and user confirmation or correction information for the root causes is received.

[0037] The abnormal event, the corresponding event-level heterogeneous association graph, and the final root cause confirmed by the user are stored as a new labeled sample in the training sample library.

[0038] The fault diagnosis model is incrementally trained or its parameters are calibrated based on the newly labeled samples.

[0039] Furthermore, to achieve the above objectives, this invention also proposes an intelligent cause-finding system for WAT parameter anomalies, which is applied to the intelligent cause-finding method for WAT parameter anomalies described above. The system includes:

[0040] The acquisition module is used to acquire metadata from multiple independent systems associated with WAT testing. The metadata includes at least one or more of the following: WAT parameter test results, test equipment status data, probe station status and environment data, probe card health index data, and fault detection and classification (FDC) data, measurement data, and yield data during the manufacturing process.

[0041] The processing module is used to uniformly identify and align the metadata in time sequence, and generate a set of related data within a unified time window;

[0042] The evaluation module is used to perform data quality checks on the associated data set and output a credibility score for each WAT record based on the test link status information, wherein the test link status information includes at least the test machine status, probe card health indicators and environmental data.

[0043] The detection module is used to perform constrained anomaly detection on the WAT parameters based on the credibility score. When an anomaly is detected, an anomaly event is generated, wherein the anomaly event includes anomaly parameters, location information, anomaly type, and anomaly intensity characteristics.

[0044] The construction module is used to extract relevant data within a preset time window and spatial window from the associated data set as multi-source evidence, with the abnormal event as the center, and to construct an event-level heterogeneous association graph based on the multi-source evidence and the characteristics of the abnormal event. The nodes of the event-level heterogeneous association graph include at least anomaly phenomenon nodes representing the abnormal event, state nodes representing the test link state, and multiple root cause candidate nodes. The nodes are connected by edges based on temporal adjacency, attribution relationship, or engineering prior knowledge.

[0045] The diagnostic module is used to input the event-level heterogeneous correlation graph into a pre-trained fault diagnosis model, which performs graph reasoning calculations and outputs one or more root cause candidates, the confidence level of each root cause candidate, and the reasoning path from the abnormal phenomenon node to the root cause candidate node to complete the cause tracing.

[0046] Furthermore, to achieve the above objectives, the present invention also proposes a WAT ​​parameter anomaly intelligent tracing device, which includes: a memory, a processor, and a WAT ​​parameter anomaly intelligent tracing program stored in the memory and executable on the processor. The WAT parameter anomaly intelligent tracing program is configured to implement the steps of the WAT parameter anomaly intelligent tracing method described above.

[0047] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a WAT ​​parameter anomaly intelligent tracing program, wherein when the WAT parameter anomaly intelligent tracing program is executed by a processor, it implements the steps of the WAT parameter anomaly intelligent tracing method described above.

[0048] This invention collects WAT test-related metadata from multiple systems; performs unified identification and time-series alignment; assesses data credibility based on test link status; performs anomaly detection and generates abnormal events according to credibility constraints; aggregates multi-source evidence to construct an event-level heterogeneous association graph; and uses a pre-trained graph neural network model for graph reasoning, outputting root cause candidates, confidence levels, and traceable reasoning paths. Through these methods, it achieves automatic differentiation between test contamination and process anomalies, improving the accuracy, interpretability, and processing efficiency of cause tracing. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the first embodiment of the intelligent cause-tracing method for WAT parameter anomalies of the present invention.

[0050] Figure 2 This is a schematic diagram of the overall process of the intelligent cause-tracing method for WAT parameter anomalies in the present invention.

[0051] Figure 3 This is a schematic diagram illustrating the classification and role of multi-source data in the cause-tracing process of the WAT parameter anomaly intelligent cause-tracing method of the present invention;

[0052] Figure 4 This is a schematic diagram of the data processing of the unified identifier and time-series association module in the WAT parameter anomaly intelligent tracing method of the present invention;

[0053] Figure 5 This is a schematic diagram of the data cleaning and credibility assessment process in the WAT parameter anomaly intelligent tracing method of the present invention;

[0054] Figure 6 This is a schematic diagram of the template structure of the traceable inference link in the WAT parameter anomaly intelligent tracing method of the present invention;

[0055] Figure 7 This is a schematic diagram of the WAT parameter anomaly detection and event construction process in the WAT parameter anomaly intelligent tracing method of the present invention;

[0056] Figure 8 This is a flowchart illustrating the evidence aggregation, graph instantiation, and cause inference processes in the WAT parameter anomaly intelligent cause tracing method of the present invention.

[0057] Figure 9 This is a schematic diagram of the result presentation, feedback loop and system update mechanism in the intelligent cause-tracing method for WAT parameter anomalies of the present invention;

[0058] Figure 10 This is a structural block diagram of the first embodiment of the WAT parameter anomaly intelligent cause tracing system of the present invention.

[0059] 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

[0060] 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.

[0061] This invention provides an intelligent cause-tracing method for WAT parameter anomalies, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the intelligent cause-tracing method for WAT parameter anomalies according to the present invention.

[0062] In this embodiment, the intelligent cause-finding method for WAT parameter anomalies includes the following steps:

[0063] Step S10: Collect metadata from multiple independent systems associated with the WAT test.

[0064] In this embodiment, the executing entity is a WAT ​​parameter anomaly intelligent tracing device. This WAT parameter anomaly intelligent tracing device has functions such as data processing, data communication, and program execution. The WAT parameter anomaly intelligent tracing 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 application.

[0065] It should be noted that the aforementioned existing technologies have at least the following shortcomings: Multi-source data is stored in a scattered manner and lacks a unified time-series alignment method; WAT results are difficult to automatically correlate with machine status, probe station environment, probe card health indicators, upstream and downstream processes, and measurement data on the same timeline, resulting in long positioning cycles; test system problems and manufacturing process defects are easily confused. Relying solely on WAT results or a single system log makes it difficult to distinguish between "measurement contamination" and "real process anomalies," easily leading to misjudgments and incorrect handling; the rule base has limited coverage, insufficient generalization ability when facing new products, new processes, or new types of faults, and high rule maintenance costs; root cause conclusions lack quantitative credibility and traceable reasoning links, making it difficult to support auditing and continuous improvement.

[0066] To address the aforementioned technical issues, this embodiment collects WAT test-related metadata from multiple systems; performs unified identification and time-series alignment; assesses data reliability based on test link status; performs anomaly detection and generates abnormal events according to reliability constraints; aggregates multi-source evidence to construct an event-level heterogeneous association graph; and utilizes a pre-trained graph neural network model for graph reasoning, outputting root cause candidates, confidence levels, and traceable reasoning paths. Through these methods, automatic differentiation between test contamination and process anomalies is achieved, improving the accuracy, interpretability, and processing efficiency of cause tracing. Specifically, this can be implemented as follows.

[0067] In this embodiment, for ease of understanding, the processing of "intelligent cause tracing of WAT parameter anomalies" is abstracted into several consecutive process steps, which correspond one-to-one with the system functional modules (M1-M7). For example... Figure 2 As shown, the overall process begins with multi-source data entering the system, proceeds through unified identification and time-series correlation, data cleaning and credibility assessment, anomaly detection, and causal reasoning, and finally outputs an auditable report and enters the feedback loop. Figure 3 As shown, different data categories play different roles in the process. Among them, WAT results are a necessary input for anomaly detection, machine / probe / environment data are mainly used for reliability assessment and test contamination identification, and FDC / measurement / yield data are mainly used to support the inference of root causes on the process side.

[0068] Furthermore, the method may include the following steps: Step S101: Collect and access multi-source metadata (corresponding to...) Figure 2 Phase 2 Figure 3 ), to obtain the data stream to be stored; Step S102: Perform unified identification and time sequence alignment on the data (corresponding to Figure 2 Phase 3), generating associated data within a unified time window; Step S103: Performing cleaning, quality labeling, and credibility scoring on the associated data (corresponding to...) Figure 2 Phase 4 Figure 5), output credibility score and quality label; Step S104: Perform WAT parameter anomaly detection and anomaly event construction under credibility constraints (corresponding to Figure 2 Phase 5 Figure 7 ), to obtain the abnormal event Event; Step S105: Aggregate multi-source evidence around the abnormal event and instantiate the event-level association graph (corresponding to Figure 2 Phase 6 Figure 4 , Figure 6 , Figure 8 Step S106: Perform abductive reasoning on the event-level association graph and generate a traceable reasoning link (corresponding to...) to form the inference input; Step S106: Perform abductive reasoning on the event-level association graph and generate a traceable inference link (corresponding to...) Figure 8 ), output the Top-K root causes and their confidence scores; Step S107: Generate a root cause report and receive manual feedback, perform sample write-back and model / knowledge base update (corresponding to Figure 2 Phase 7).

[0069] In its implementation, the data acquisition and access module (M1) collects data from WAT testing-related equipment and information systems, and assigns a "source system timestamp" and a "collection timestamp" to each record for timing correction and alignment in subsequent steps. The metadata includes at least one or more of the following: WAT parameter test results, test equipment status data, probe station status and environmental data, probe card health index data, and fault detection and classification (FDC) data, measurement data, and yield data from the manufacturing process. For example, WAT parameter test results include: parameter name, test structure / Pad coordinates, measured value, specification upper and lower limits, test site, test sequence, sampling time, etc.; test equipment status includes: equipment number, hardware self-test results, SMU / instrument calibration status, key component temperature, power / communication status, alarm and event logs, etc.; probe station status and environment includes: temperature, humidity, Chuck temperature control, cleanliness-related indicators, applied stress conditions, etc.; probe card health includes: probe card number, number of uses / mileage, tip contact resistance, leakage current, contact uniformity, maintenance records, etc.; key process step FDC data includes: equipment parameter statistics and original curve index; measurement / yield / defect data includes: key measurement indicators, yield trends, defect statistics, etc.; engineering-related priors include: a correlation table between WAT parameters and process steps / equipment / measurement indicators, used to constrain the subsequent cause-finding search space.

[0070] It should be noted that the access methods can include, but are not limited to: device-side subscription to data collection via SECS / GEM or EDA (Interface A); system-side synchronization via REST API, message queue, or database (JDBC / ODBC); or batch import via offline files (CSV / Parquet, etc.). To improve cross-system consistency, clock synchronization (NTP / PTP) can be performed on the data collection nodes, and the metadata required for time drift correction can be retained.

[0071] Step S20: Perform unified identification and time-series alignment on the metadata to generate a set of associated data within a unified time window.

[0072] It should be noted that the unified identification and timing association module (M2) performs object identification unification and timing alignment on the above-mentioned accessed data, so that data from different systems can form a related data set around the same batch / wafer / equipment / process step on the same time axis.

[0073] In the specific implementation, unified identification and time-series alignment are implemented, such as object identifier normalization: mapping batch numbers, wafer numbers, test structure IDs, equipment numbers, process step IDs, etc., from different systems to a unified ID. Timestamp correction: calculating and correcting drift based on the source system time and acquisition time. Time window association: selecting a preset time window centered on the WAT test occurrence time, associating machine logs, probe environment, probe card health, FDC statistics, etc., with the same event. Granularity alignment: when data granularity differs (e.g., probe card to pin granularity, FDC to second-level, WAT to point-level), aggregation / interpolation / nearest neighbor matching methods are used to form a unified representation usable for inference. The system achieves cross-table association through a unified primary key, which must include at least: Lot_ID, Wafer_ID, (optional) Die_ID, Test_Structure_ID, Parameter_ID, Equipment_ID, Process_Step_ID, and Timestamp.

[0074] Step S30: Perform a data quality check on the associated data set and output a credibility score for each WAT record based on the test link status information.

[0075] In its specific implementation, the quality inspection and credibility score output process includes identifying potential data contamination records in the associated data set due to poor test link status. The identification criteria for potential data contamination include probe contact resistance exceeding limits, abnormal equipment calibration status, environmental parameters exceeding limits, or anomaly detection results of multi-dimensional features based on isolated forest or single-class support vector machine models. The credibility score is output for each WAT record based on the identification results of potential data contamination and the severity of associated alarms. The test link status information includes at least the test equipment status, probe card health indicators, and environmental data.

[0076] In its implementation, the above process includes: integrity checks: removal or completion of missing fields, duplicate records, and invalid coordinates / structure IDs. Range and physical consistency rules: records with measured values ​​exceeding the instrument's range or contact resistance / leakage current exceeding thresholds are marked as low confidence. Robust statistical filtering: IQR or Hampel filtering is used to identify spike noise. Model-assisted identification: Isolation Forest or One-Class SVM is used to identify measurement contamination samples based on multi-dimensional features such as Value, Contact_R, Humidity, and Cal_Status. Confidence synthesis: The score is calculated by combining machine health, probe health, environmental out-of-bounds conditions, and alarm severity, and a quality label is output (e.g., suspected poor contact / suspected expired calibration / suspected environmental out-of-bounds conditions). For example, if a WAT ​​point is abnormal, and the probe contact resistance increases while the machine experiences a contact-related alarm, the score of that record is reduced, and "test link abnormality" is given priority as a root cause candidate in subsequent steps.

[0077] Step S40: Perform constraint anomaly detection on the WAT parameters based on the credibility score. When an anomaly is detected, generate an anomaly event.

[0078] It should be noted that, in this embodiment, the constraint anomaly detection is to raise the anomaly judgment threshold of the corresponding WAT parameter record or mark the WAT parameter record as pending retesting when the confidence score is lower than the preset threshold. The anomaly event includes anomaly parameters, location information, anomaly type and anomaly intensity characteristics.

[0079] In the specific implementation, a baseline model is established for each WAT parameter under a given product / structure / site, and anomaly scores are calculated: (a) Statistical baseline: a threshold is constructed using robust mean / median and MAD; (b) Temporal baseline: drift, abrupt changes, and periodic anomalies are identified using sliding window, EWMA, or predictive models; (c) Spatial morphology: point coordinates are clustered, and spatial clustering and hotspot areas are output; (d) Co-occurrence features: co-occurrence patterns of multiple parameters simultaneously exhibiting anomalies in the same batch are extracted. When the anomaly score exceeds the threshold, an anomaly event is generated and recorded: anomaly parameter set, batch / wafer / structure, anomaly type, severity, spatial clustering features, trigger time window, etc.

[0080] Step S50: Centered on the abnormal event, extract relevant data within the preset time window and spatial window from the associated data set as multi-source evidence, and construct an event-level heterogeneous association graph based on the multi-source evidence and the characteristics of the abnormal event.

[0081] In a specific implementation, based on the event identifier of the abnormal event, a related time window and related spatial range for retrieving evidence are determined; multi-source evidence data related to the abnormal event within the related time window and related spatial range are pulled from the related data set, wherein the multi-source evidence data includes at least WAT measurement sequence data, test machine alarm logs, probe card health timing indicators, environmental monitoring data, FDC statistical characteristics of related process steps, and related measurement or yield trend data.

[0082] It should be noted that when the anomaly detection module outputs an anomaly event, the system first standardizes the description of the Event to form a unified event identifier and analysis scope: the event identifier includes at least: Lot_ID, Wafer_ID, Die_ID / Pad coordinates, Test_Structure_ID, Parameter_ID, Test_Station_ID, Equipment_ID, and Event_TS; the analysis scope includes a time window and a spatial window: the time window is used to limit the scope before and after evidence retrieval, and the spatial window is used to limit the set of points involved in the anomaly (e.g., an anomaly cluster area, the same set of pins, or the same Chuck area).

[0083] Furthermore, based on event identifiers and analysis scope, the system pulls evidence data related to this anomaly from multiple source systems and performs unified identifier mapping and unified timestamp alignment, including at least: WAT-side evidence: the distribution of this parameter in this batch / recent batches, measurement values ​​at each site, measurement order, retest records and differences; Test equipment evidence: alarm codes, event logs, self-test results, SMU / measuring unit calibration status, and temperature / power / communication anomalies of key components; Probe-side evidence: probe card number and health indicators (contact resistance, leakage current, number of uses, maintenance time), probe station status, and Chuck temperature control; Environmental evidence: temperature / humidity exceedances, duration of exceedances, and cleanliness-related indicators (if any); Manufacturing process evidence: process steps associated with this parameter, FDC key parameter statistics / curve indexes of the corresponding equipment, and process route change information; Measurement / yield evidence: key measurement indicator offsets, defect / yield trends and synchronous changes. The aforementioned evidence data was processed through a unified identifier and time-series association module: field mapping, primary key association, time drift correction, and granular alignment (forming statistical features for high-frequency data within a Window).

[0084] Furthermore, the step of constructing an event-level heterogeneous correlation graph based on the multi-source evidence and the characteristics of the anomalous events includes: performing structured analysis on the anomalous events to generate an anomaly profile, wherein the anomaly profile includes at least a determination of the anomaly type, a quantitative value of the anomaly intensity, a spatial morphological description of the anomaly, and the results of a retest sensitivity analysis. The event-level heterogeneous correlation graph is instantiated based on the anomaly profile, the multi-source evidence data, and the characteristics of the anomalous events. The nodes of the heterogeneous correlation graph further include: feature nodes representing specific features in the anomaly profile, tool status nodes representing specific process equipment, and monitoring nodes representing measurement or yield deviations. The edges of the heterogeneous correlation graph also include statistical correlation edges generated based on the event co-occurrence relationships or statistical correlations in the multi-source evidence data.

[0085] It should be noted that: 1) Anomaly type determination includes: out-of-bounds anomalies (exceeding USL / LSL), drift anomalies (continuous mean shift / trend change), abrupt change anomalies (step change in a short period of time), variance anomalies (significantly increased fluctuations), spatial clustering anomalies (local hotspots / strip anomalies), cross-parameter co-occurrence anomalies (synchronous anomalies of multiple parameters in the same structure / same site), etc. 2) Anomaly intensity quantification includes: out-of-bounds distance (Value-USL or LSL-Value), standardized anomaly score (z-score / MAD score), drift slope, abrupt change magnitude, variance increase, and clustering indicators (such as hotspot ratio / cluster area), etc. 3) Anomaly Pattern: Descriptions of patterns such as consistent offset across the entire wafer, anomalies only in individual pads / dies, concentrated anomalies at the same site, consistent anomalies across sites, and regional clustering; 4) Anomaly Impact Scope: Number of wafers affected, percentage of sites, site range, structural range, and whether it spreads across batches; 5) Retest Sensitivity (if retesting exists): Whether retesting recovers, the extent of recovery, and differences in retesting conditions, used to determine the tendency of "measurement contamination". The above Anom_Profile serves as part of the subsequent association graph node attributes and model input features, and is included in the report as an auditable basis for anomaly analysis (corresponding to...). Figure 6 The "Abnormal Profile / Features" output item.

[0086] In one embodiment, an event-level heterogeneous association graph G_event=(V,E) is constructed based on the Context, where: Node V includes at least: Symptom node (carrying Anom_Profile), Tester_State node (calibration / self-test / alarm, etc.), Probe_State node (contact resistance, usage count, maintenance, etc.), Env_State node (temperature and humidity exceeding limits, etc.), Process_Step and Tool_State nodes (FDC key parameter offset, etc.), Metrology_Yield node (measurement offset, yield trend, etc.), and RootCause node (test anomaly type, probe anomaly type, environmental anomaly type, process anomaly type, measurement anomaly type, etc.). Edge E includes at least: temporal adjacency edge (event association within the same window), attribution edge (attribution relationship of the same device / same probe card / same process step), and prior association edge (parameter). The graph structure includes engineering prior relationships for process steps / equipment types / measurement items, and statistically related edges (historical co-occurrence / relatedness reaching a threshold). Edge weights and node weights are jointly determined by Weight(e) and engineering priors / historical statistics, ensuring that the graph structure contains both engineering interpretable constraints and retains data-driven relationships.

[0087] Step S60: Input the event-level heterogeneous association graph into the pre-trained fault diagnosis model, and the fault diagnosis model performs graph reasoning calculations to output one or more root cause candidates, the confidence level of each root cause candidate, and the reasoning path from the abnormal phenomenon node to the root cause candidate node, so as to complete the cause tracing.

[0088] In the specific implementation, the G_event of the current event is input into the WAT parameter anomaly fault diagnosis model based on graph neural network to obtain the root cause candidate set {RC_i} and its confidence level Conf_i. To improve engineering handling efficiency, the system can classify and output root causes according to their categories: measurement link related root causes (test machine / pin card / environment, etc.) and their confidence levels; manufacturing process related root causes (process steps / equipment parameter drift, etc.) and their confidence levels; measurement / data link related root causes (measurement equipment / Recipe, etc.) and their confidence levels. The system outputs the Top-K root causes in descending order of Conf_i and provides the location object (equipment number / pin card number / process step / alarm code / measurement item, etc.) and suggested action priorities. To ensure traceability and interpretability, the system extracts the most contributing path* from G_event for Top-K root causes, forming a structured inference chain: Symptom (Anom_Profile) → State (alarm / calibration / contact resistance / humidity exceedance / FDC offset, etc.) → Root Cause → Object. Simultaneously, it outputs an Evidence List: each piece of evidence includes at least the source system, record ID / timestamp, key fields, change / exceedance amount, Weight(e), and contribution ranking. This chain and the evidence list together constitute an auditable basis for inference.

[0089] It should be noted that the fault diagnosis model is a model built on graph neural network (GNN). The pre-training process involves training the graph neural network model using historical labeled samples. Each historical labeled sample includes an event-level heterogeneous association graph constructed based on historical abnormal events, and a final root cause label confirmed by engineering. The goal of the training is to enable the model to learn to aggregate the information of abnormal phenomenon nodes and associated state nodes in the graph to root cause candidate nodes through a message passing mechanism, and output a root cause prediction probability distribution consistent with the final root cause label. Specific implementation examples include: 1) Training sample construction: Extracting samples from historical anomaly events, each sample containing G_event, Anom_Profile, evidence set, and engineering-confirmed root cause label RootCause_Label and location object Object_Label; 2) Feature encoding: Encoding the anomaly profile and evidence features into node / edge features, such as: anomaly intensity, drift slope, clustering degree, alarm severity, calibration status, contact resistance trend, humidity violation duration, FDC offset, measurement / yield offset, and confidence score; 3) GNN message passing and representation learning: Performing multiple rounds of message passing on the heterogeneous graph to aggregate information from phenomenon nodes and evidence nodes for root cause candidate nodes; 4) Root cause prediction and ranking: Outputting the probability / score distribution of each root cause candidate for Top-K ranking; 5) Training objectives and constraints: Jointly optimizing the root cause classification loss (such as cross-entropy) and Top-K ranking related loss, and introducing interpretable constraints to make high-contribution paths conform as much as possible. Figure 4 The template structure shown; 6) Model versioning and threshold setting: Record the model version, feature version, and graph construction rule version, and set trigger thresholds for different risk levels (e.g., if the confidence level is lower than the threshold, manual review is required).

[0090] Furthermore, in this embodiment, the root cause candidates, the confidence level, and the inference path are presented in the form of a report, and the user's confirmation or correction information on the root cause is received; the abnormal event, the corresponding event-level heterogeneous association graph, and the final root cause confirmed by the user are stored as a new labeled sample in the training sample library; the fault diagnosis model is incrementally trained or its parameters are calibrated based on the new labeled sample.

[0091] For ease of understanding, this embodiment describes a specific implementation. In this implementation, the diagnostic results output from the above steps are transformed into an executable presentation and report, and the engineer's confirmed conclusions are written back to form a closed loop, thereby continuously optimizing the WAT parameter anomaly fault diagnosis model and fault mode library based on graph neural networks. This includes the following processes, which can be referred to in detail. Figure 9 As shown: Results presentation and processing suggestions generated (corresponding) Figure 9S107-1 in the system): The system presents at least the following information in the form of alarm cards, lists, or dashboards: 1) Anomaly overview: anomaly parameters, batch / wafer / structure / site / coordinate range, trigger time, Anom_Profile (type / intensity / morphology / range / retest sensitivity); 2) Root cause Top-K and confidence: displaying root cause name / category, Conf, location object (device number / pin card number / process step, etc.), and highlighting the root cause with the highest confidence; 3) Evidence summary: providing several key pieces of evidence for each root cause (alarm code, calibration expiration, contact resistance trend, humidity exceeding limits, FDC offset, measurement / yield offset, etc.) and the contribution of the evidence; 4) Visualization of inference chain: by Figure 4 The template displays "Phenomenon → Status → Root Cause → Object," supporting expanded viewing of original evidence records and curve indexes; 5) Suggested Actions and Priorities: Executable suggestions are generated based on the root cause category. For example: For measurement link root causes: suggestions include retesting and verification, cleaning / replacing the SIM card, calibrating / self-testing the equipment, and environmental verification; For manufacturing process root causes: suggestions include batch locking, locking associated process steps / equipment, pulling FDC curves, and initiating equipment / process investigation; For measurement link root causes: suggestions include verifying measurement equipment / Recipe and retesting measurement indicators. The system can assign priorities and verification paths to suggested actions (e.g., "Retest first—if recovery occurs, the measurement contamination tendency is higher").

[0092] Furthermore, the automatic generation of cause-finding reports and audit records (corresponding) Figure 9 S107-2: The system automatically generates a structured causal report, which should include at least: basic event information (event ID, anomaly level, time window / spatial window); a list of input data (data source, data range, data version / table version); anomaly analysis results (Anom_Profile and key statistics); root cause Top-K, confidence level Conf, location objects and suggested actions; inference path* and evidence list_List (including evidence weight (e), change, source record ID / index); audit information (model version, feature version, graph construction rule version, threshold configuration, generation time). The report should support tracing back to the original records and curve indexes, meeting the requirements for retrospective analysis and quality auditing.

[0093] Furthermore, manual verification, rewriting of labels, and closed-loop sample sedimentation (corresponding to...) Figure 9(S107-3 in the document): The system provides an engineer confirmation entry point, where engineers can confirm or correct Top-K root causes and fill in the action and verification results. The system writes this result back as a structured label, used as a closed-loop sample: Event_ID→Final_RootCause→Final_Object→Action_Taken→Verification_Result (whether it has recovered after retesting / calibration / needle card replacement / process troubleshooting)→Remark. Among them, Verification_Result can serve as a strong verification signal to improve the credibility of sample labels and the effect of subsequent model updates.

[0094] Furthermore, model and knowledge base updates (corresponding to) Figure 9 S107-4): The system updates the WAT parameter anomaly fault diagnosis model and knowledge base based on the written-back samples: confirmed events are added to the training sample library, the sample balance is maintained by root cause category and strong / weak labels are marked; when the preset triggering conditions are met (such as sample size threshold, periodic triggering or new process version online), the diagnostic model is incrementally trained or probabilistically calibrated to form a new model version and retain the historical version for traceability; the edge weights and rule weights in the fault mode library / knowledge graph are updated synchronously (for example, if the association between a certain piece of evidence and a certain root cause is repeatedly disproven, the weight is reduced); the closed-loop effect indicators (Top-K hit rate, average location time, false positive rate, change in the number of retests, etc.) are recorded for continuous evaluation and optimization.

[0095] Furthermore, the output includes: executable project results, cause-finding reports, write-back tag data, model / knowledge base update records, and performance metrics.

[0096] This embodiment collects WAT test-related metadata from multiple systems; performs unified identification and time-series alignment; assesses data credibility based on test link status; performs anomaly detection and generates abnormal events according to credibility constraints; aggregates multi-source evidence to construct an event-level heterogeneous association graph; and uses a pre-trained graph neural network model for graph reasoning, outputting root cause candidates, confidence levels, and traceable reasoning paths. Through these methods, automatic differentiation between test contamination and process anomalies is achieved, improving the accuracy, interpretability, and processing efficiency of cause tracing.

[0097] Furthermore, this embodiment of the invention also proposes a storage medium storing a WAT ​​parameter anomaly intelligent tracing program, which, when executed by a processor, implements the steps of the WAT parameter anomaly intelligent tracing method described above.

[0098] Reference Figure 10 , Figure 10 This is a structural block diagram of the first embodiment of the WAT parameter anomaly intelligent cause tracing system of the present invention.

[0099] like Figure 10 As shown, the intelligent cause-tracing system for WAT parameter anomalies proposed in this embodiment of the invention includes:

[0100] The acquisition module 10 is used to acquire metadata from multiple independent systems associated with WAT testing. The metadata includes at least one or more of the following: WAT parameter test results, test equipment status data, probe station status and environment data, probe card health index data, and fault detection and classification (FDC) data, measurement data, and yield data during the manufacturing process.

[0101] Processing module 20 is used to uniformly identify and time-align the metadata, and generate a set of associated data within a unified time window;

[0102] Evaluation module 30 is used to perform data quality checks on the associated data set and output a credibility score for each WAT record based on the test link status information, wherein the test link status information includes at least the test machine status, probe card health indicators and environmental data.

[0103] The detection module 40 is used to perform constrained anomaly detection on the WAT parameters based on the credibility score. When an anomaly is detected, an anomaly event is generated, wherein the anomaly event includes anomaly parameters, location information, anomaly type, and anomaly intensity characteristics.

[0104] The construction module 50 is used to extract relevant data within a preset time window and spatial window from the associated data set as multi-source evidence, with the abnormal event as the center, and to construct an event-level heterogeneous association graph based on the multi-source evidence and the characteristics of the abnormal event. The nodes of the event-level heterogeneous association graph include at least an abnormal phenomenon node representing the abnormal event, a status node representing the test link state, and multiple root cause candidate nodes. The nodes are connected by edges based on temporal adjacency, attribution relationship, or engineering prior knowledge.

[0105] The diagnostic module 60 is used to input the event-level heterogeneous correlation graph into a pre-trained fault diagnosis model, and the fault diagnosis model performs graph reasoning calculations to output one or more root cause candidates, the confidence level of each root cause candidate, and the reasoning path from the abnormal phenomenon node to the root cause candidate node, so as to complete the cause tracing.

[0106] This embodiment collects WAT test-related metadata from multiple systems; performs unified identification and time-series alignment; assesses data credibility based on test link status; performs anomaly detection and generates abnormal events according to credibility constraints; aggregates multi-source evidence to construct an event-level heterogeneous association graph; and uses a pre-trained graph neural network model for graph reasoning, outputting root cause candidates, confidence levels, and traceable reasoning paths. Through these methods, automatic differentiation between test contamination and process anomalies is achieved, improving the accuracy, interpretability, and processing efficiency of cause tracing.

[0107] This application embodiment also provides a WAT ​​parameter anomaly intelligent tracing 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 parameter anomaly intelligent tracing program. When the processor executes the program stored in the memory, it implements the above-mentioned WAT parameter anomaly intelligent tracing method.

[0108] The communication bus mentioned in the aforementioned intelligent cause-finding device for abnormal WAT parameters 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.

[0109] The communication interface is used for communication between the aforementioned WAT parameter anomaly intelligent cause tracing device and other devices.

[0110] 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.

[0111] 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.

[0112] 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)).

[0113] 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.

[0114] 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.

[0115] 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.

[0116] 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.

[0117] 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.

[0118] In addition, for technical details not described in detail in this embodiment, please refer to the intelligent cause-finding method for abnormal WAT parameters provided in any embodiment of the present invention, which will not be repeated here.

[0119] 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.

[0120] 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.

[0121] 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.

[0122] 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.

[0123] 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. An intelligent method of tracing the cause of WAT parameter abnormality, characterized in that, The intelligent cause-finding method for WAT parameter anomalies includes: Metadata is collected from multiple independent systems associated with WAT testing. The metadata includes at least one or more of the following: WAT parameter test results, test equipment status data, probe station status and environmental data, probe card health index data, and Fault Detection and Classification (FDC) data, measurement data, and yield data during the manufacturing process. The metadata is uniformly identified and time-series aligned to generate a set of associated data within a unified time window; The associated data set is subjected to data quality checks, and a credibility score is output for each WAT record based on the test link status information, wherein the test link status information includes at least the test machine status, probe card health indicators and environmental data; Based on the credibility score, the WAT parameters are subjected to constrained anomaly detection. When an anomaly is detected, an anomaly event is generated, wherein the anomaly event includes anomaly parameters, location information, anomaly type, and anomaly intensity characteristics. Centered on the abnormal event, relevant data within a preset time window and spatial window are extracted from the associated data set as multi-source evidence. Based on the multi-source evidence and the characteristics of the abnormal event, an event-level heterogeneous association graph is constructed. The nodes of the event-level heterogeneous association graph include at least anomaly phenomenon nodes representing the abnormal event, state nodes representing the test link state, and multiple root cause candidate nodes. The nodes are connected by edges based on temporal adjacency, attribution relationship, or engineering prior knowledge. The event-level heterogeneous correlation graph is input into a pre-trained fault diagnosis model, which performs graph reasoning calculations and outputs one or more root cause candidates, the confidence level of each root cause candidate, and the reasoning path from the abnormal phenomenon node to the root cause candidate node to complete the cause tracing.

2. The intelligent cause-tracing method for WAT parameter anomalies as described in claim 1, characterized in that, The step of performing data quality checks on the associated data set and outputting a credibility score for each WAT record based on the test link status information includes: Identify potential data contamination records in the associated data set caused by poor test link status. The identification criteria for potential data contamination include probe contact resistance exceeding limits, abnormal machine calibration status, environmental parameters exceeding limits, or anomaly detection results of multidimensional features based on isolated forest or single-class support vector machine models. Based on the identification results of potential data contamination and the severity of associated alarms, a credibility score is output for each WAT record.

3. The intelligent cause-tracing method for WAT parameter anomalies as described in claim 1, characterized in that, The constraint anomaly detection involves raising the anomaly judgment threshold of the corresponding WAT parameter record or marking the WAT parameter record as pending retesting when the credibility score is lower than a preset threshold.

4. The intelligent cause-tracing method for WAT parameter anomalies as described in claim 1, characterized in that, The step of extracting relevant data within a preset time window and spatial window from the associated data set, centered on the abnormal event, as multi-source evidence includes: Based on the event identifier of the abnormal event, determine the associated time window and associated spatial range for retrieving evidence; Pull multi-source evidence data related to the abnormal event within the associated time window and associated spatial range from the associated data set. The multi-source evidence data includes at least WAT measurement sequence data, test machine alarm logs, probe card health timing indicators, environmental monitoring data, FDC statistical characteristics of associated process steps, and associated measurement or yield trend data.

5. The intelligent cause-tracing method for WAT parameter anomalies as described in claim 4, characterized in that, The construction of an event-level heterogeneous correlation graph based on the multi-source evidence and the characteristics of the anomalous events includes: The abnormal events are subjected to structured analysis to generate an abnormal profile, which includes at least the determination of the abnormal type, the quantitative value of the abnormal intensity, the spatial morphology description of the abnormality, and the results of the retest sensitivity analysis. Based on the anomaly profile, the multi-source evidence data, and the features of the anomaly event, the event-level heterogeneous correlation graph is instantiated. The nodes of the heterogeneous correlation graph further include: feature nodes representing specific features in the anomaly profile, tool status nodes representing specific process equipment, and monitoring nodes representing measurement or yield deviation. The edges of the heterogeneous correlation graph also include statistical correlation edges generated based on the event co-occurrence relationship or statistical correlation in the multi-source evidence data.

6. The intelligent cause-tracing method for WAT parameter anomalies as described in claim 1, characterized in that, The fault diagnosis model is a model built based on a graph neural network (GNN). The pre-training process involves training the graph neural network model using historical labeled samples. Each historical labeled sample includes an event-level heterogeneous association graph constructed based on historical abnormal events, as well as a final root cause label confirmed by engineering. The goal of the training is to enable the model to learn to aggregate information from anomalous nodes and associated state nodes in the graph to root cause candidate nodes through a message passing mechanism, and output a root cause prediction probability distribution that is consistent with the final root cause label.

7. The intelligent cause-finding method for WAT parameter anomalies as described in any one of claims 1 to 6, characterized in that, The method further includes: The root cause candidates, the confidence levels, and the reasoning paths are presented in the form of reports, and user confirmation or correction information for the root causes is received. The abnormal event, the corresponding event-level heterogeneous association graph, and the final root cause confirmed by the user are stored as a new labeled sample in the training sample library. The fault diagnosis model is incrementally trained or its parameters are calibrated based on the newly labeled samples.

8. A smart cause-tracing system for abnormal WAT parameters, characterized in that, The WAT parameter anomaly intelligent tracing system is applied to the WAT parameter anomaly intelligent tracing method as described in any one of claims 1 to 7, the system comprising: The acquisition module is used to acquire metadata from multiple independent systems associated with WAT testing. The metadata includes at least one or more of the following: WAT parameter test results, test equipment status data, probe station status and environment data, probe card health index data, and fault detection and classification (FDC) data, measurement data, and yield data during the manufacturing process. The processing module is used to uniformly identify and align the metadata in time sequence, and generate a set of related data within a unified time window; The evaluation module is used to perform data quality checks on the associated data set and output a credibility score for each WAT record based on the test link status information, wherein the test link status information includes at least the test machine status, probe card health indicators and environmental data. The detection module is used to perform constrained anomaly detection on the WAT parameters based on the credibility score. When an anomaly is detected, an anomaly event is generated, wherein the anomaly event includes anomaly parameters, location information, anomaly type, and anomaly intensity characteristics. The construction module is used to extract relevant data within a preset time window and spatial window from the associated data set as multi-source evidence, with the abnormal event as the center, and to construct an event-level heterogeneous association graph based on the multi-source evidence and the characteristics of the abnormal event. The nodes of the event-level heterogeneous association graph include at least anomaly phenomenon nodes representing the abnormal event, state nodes representing the test link state, and multiple root cause candidate nodes. The nodes are connected by edges based on temporal adjacency, attribution relationship, or engineering prior knowledge. The diagnostic module is used to input the event-level heterogeneous correlation graph into a pre-trained fault diagnosis model, which performs graph reasoning calculations and outputs one or more root cause candidates, the confidence level of each root cause candidate, and the reasoning path from the abnormal phenomenon node to the root cause candidate node to complete the cause tracing.

9. A smart cause-tracing device for abnormal WAT parameters, characterized in that, The WAT parameter anomaly intelligent tracing device includes: a memory, a processor, and a WAT ​​parameter anomaly intelligent tracing program stored in the memory and executable on the processor, wherein the WAT parameter anomaly intelligent tracing program is configured to implement the steps of the WAT parameter anomaly intelligent tracing method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a WAT ​​parameter anomaly intelligent tracing program, which, when executed by a processor, implements the steps of the WAT parameter anomaly intelligent tracing method as described in any one of claims 1 to 7.

Images