WAT machine health degree automatic monitoring method, device and equipment and storage medium

Abstract

The present application belongs to the technical field of semiconductor manufacturing, and discloses a WAT machine health degree automatic monitoring method, device, equipment and storage medium. The method comprises the following steps: collecting and outputting original multi-source data sets from a plurality of heterogeneous data sources related to a WAT test machine; processing the original data and outputting layered aligned time series data sets; inputting the multi-dimensional quantitative feature vector calculated based on the current monitoring window and the historical baseline window into an artificial intelligence model after baseline standardization processing, calculating the real-time comprehensive health score, future risk probability and health trend prediction result of the machine; and generating on-demand verification decisions or maintenance warning work orders according to the comprehensive health score, risk probability and prediction result combined with threshold rules. Through the above method, the present application changes the traditional fixed period verification into an on-demand verification mode based on the real health status of the machine, significantly reduces the maintenance cost and unplanned downtime probability, and realizes real-time, quantitative and interpretable monitoring of the health degree of the WAT machine.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and in particular to a method, apparatus, equipment and storage medium for automatic monitoring of the health status of WAT (Wastewater Acquisition) machines. Background Technology

[0002] Traditional WAT testing machine health management primarily relies on a reactive maintenance model. This means that problems are only identified and addressed through manual troubleshooting by engineers using SPC control charts, and the machine's condition is restored via repair, calibration, or verification, after abnormal test data, decreased yield, or machine errors occur. Furthermore, routine machine verification typically employs a fixed-cycle preventative maintenance strategy, performed according to plan regardless of the machine's actual condition.

[0003] Existing technologies have the following drawbacks: they are highly reactive, only intervening after problems have caused test anomalies or production line shutdowns, making it impossible to detect gradual degradation in advance; they waste resources significantly, requiring fixed-cycle calibration of healthy and stable machines, consuming substantial manpower, time, and spare parts costs; their judgments are highly subjective, relying on human experience to analyze multi-source data, which can easily lead to missed judgments or excessive intervention; and they lack forward-looking prediction, failing to effectively integrate multi-source information such as equipment operating status and maintenance history for comprehensive risk assessment and degradation trend prediction.

[0004] 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

[0005] The main objective of this invention is to provide an automatic monitoring method, device, equipment, and storage medium for the health status of WAT testing machines, aiming to solve the technical problems of lagging maintenance, resource waste, subjective judgment, and lack of prediction in the existing technology of WAT testing machines.

[0006] To achieve the above objectives, the present invention provides an automatic health monitoring method for WAT (Wastewater Acquisition) machines, the method comprising the following steps: Full data is periodically collected from multiple heterogeneous data sources related to the WAT test machine, and a time-aligned original multi-source dataset is output based on the full data. The multiple heterogeneous data sources include at least a test result database, a statistical process control system, and a manufacturing execution system. The original multi-source dataset is cleaned, hierarchically processed, and aligned to output a hierarchically aligned time-series dataset. Based on the predefined current monitoring window and historical baseline window, as well as the hierarchical aligned time series dataset, a multidimensional quantized feature vector is output, wherein the multidimensional quantized feature vector is used to reflect the health status of the WAT test machine. The multidimensional quantized feature vector is baseline-normalized, and the normalized feature vector is input into an artificial intelligence model to calculate the real-time comprehensive health score and future runaway risk probability of the WAT test machine, and to predict future health trends. Based on the comprehensive health score, future risk of loss of control probability, and future health trend prediction results in the assessment result set, combined with preset threshold rules, an on-demand verification decision or maintenance early warning work order is generated. The future health trend prediction results include the health score prediction sequence of the WAT test machine within a specified future time window, the confidence interval of the prediction sequence, and the expected time point when the health score prediction sequence first reaches the preset risk threshold.

[0007] In one embodiment, the hierarchical processing involves grouping and storing data into multiple independent subsets according to at least one dimension, such as product family tree, test program version, test temperature level, and site configuration; the alignment processing involves resampling the timeline at the minute or event level to align at least two types of heterogeneous data, such as test results, statistical process control events, equipment state transitions, alarm logs, and maintenance records, to a unified timeline.

[0008] In one embodiment, the multidimensional quantization feature vector includes test distribution and stability features, statistical process control features, equipment operation and reliability features, maintenance and verification effect features, and wafer diagram spatial structure features. The test distribution and stability features include at least one or more of the following sub-features: distribution drift features, used to quantify the degree of deviation in the overall distribution position of test data between the current monitoring window and the historical baseline window; fluctuation amplification features, used to quantify the magnitude of change in the dispersion of test data within the current monitoring window relative to the historical baseline window; and tail risk features, used to quantify the proportion and magnitude of test data exceeding specification limits or control limits within the current monitoring window. The statistical process control features include at least one or more of the following sub-features: exceedance intensity features, used to quantify the cumulative frequency and magnitude of exceedance events in the SPC control chart per unit time; trend persistence features, used to quantify the duration and degree of deviation of continuous violations of SPC discrepancy rules; and recovery time features, used to quantify the number of tests or time required for test data to return to normal after an SPC event is triggered. The equipment operation and reliability characteristics include at least one or more of the following sub-features: alarm density feature, used to quantify the frequency of machine alarms per unit time; alarm clustering feature, used to quantify the co-occurrence relationship and clustering intensity of different alarm codes within the same time window; downtime variation feature, used to quantify the relative change rate of the cumulative downtime of the machine between the current monitoring window and the historical baseline window; throughput fluctuation feature, used to quantify the fluctuation amplitude of the machine's output per unit time or cycle time. The maintenance and verification effect characteristics are used to quantify the changes in the machine's health status before and after a maintenance or verification event, as well as the average interval between the recurrence of similar problems. The spatial structure features of the wafer image include at least one or more of the following sub-features: spatial pattern similarity features, used to quantify the degree of spatial structural similarity between two wafer images in terms of defect distribution, electrical parameter distribution, or yield distribution; and abnormal hotspot features, used to quantify the intensity of systematic deviation of local areas on the wafer image relative to the overall distribution.

[0009] In one embodiment, the baseline normalization process for the multidimensional quantized feature vector includes: The stable segment data is filtered according to the aligned time series dataset, wherein the stable segment data meets the conditions of no SPC triggering, no maintenance events, no alarm aggregation for a consecutive preset number of days, and the distribution of key test items is not significantly different from the reference distribution. Based on the stable segment data, an individual baseline is established for each WAT test machine, and a baseline parameter set is output. The baseline includes at least position parameters and scale parameters. Based on the baseline parameter set and the multidimensional quantized feature vector, the multidimensional feature vector is standardized using a standardization method based on robust statistics, and the standardized feature vector is output.

[0010] In one embodiment, the formula for calculating the comprehensive health score is:

[0011] in, To calculate the overall health score, To include the number of feature dimensions in the fusion, For the first The weights of each feature For the first Robust Z-scores for each feature To take the absolute value, it means that the deviation intensity does not distinguish between positive and negative directions.

[0012] In one embodiment, the prediction of future health trends includes: Historical time-series data is acquired, and a Long Short-Term Memory (LSTM) network is used to perform multi-step predictions on the comprehensive health score and the time series of key risk features extracted from the historical time-series data. The predicted sequence, confidence interval, and expected time point for reaching the risk threshold within a specified number of days are output. The multi-step prediction formula is as follows:

[0013] in, At the current time point, To predict the number of steps, For lookback length, it indicates that the model input uses past data. Data at each time step For the past A sequence of health scores for each step. For the past Other covariates / key risk feature sequences of the step, For the future Step-by-step health score prediction sequence This is a time-series mapping function implemented using a Long Short-Term Memory (LSTM) network.

[0014] In one embodiment, the conditions for generating the on-demand verification decision include: An immediate verification decision is generated when the probability of future loss of control exceeds a first risk threshold. When the future health trend prediction results show that the comprehensive health score will exceed the second risk threshold within a preset future time period, an immediate verification decision is generated. If, within a consecutive first preset number of days to a second preset number of days, the overall health score remains below the third risk threshold, the probability of future loss of control remains below the fourth risk threshold, and no abnormal features are triggered, a verification exemption suggestion is generated.

[0015] Furthermore, to achieve the above objectives, the present invention also proposes an automatic monitoring device for the health status of WAT (Wastewater Acquisition) machines. This device is applied to the automatic monitoring method for the health status of WAT machines described above, and includes: The data acquisition module is used to periodically acquire full data from multiple heterogeneous data sources related to the WAT test machine, and output a time-aligned original multi-source dataset based on the full data. The multiple heterogeneous data sources include at least a test result database, a statistical process control system, and a manufacturing execution system. The data governance module is used to perform cleaning, hierarchical processing and alignment processing on the original multi-source dataset, and output a hierarchical aligned time series dataset. The feature construction module is used to output a multidimensional quantized feature vector based on the predefined current monitoring window and historical baseline window and the hierarchical aligned time series dataset, wherein the multidimensional quantized feature vector is used to reflect the health status of the WAT test machine; The health inference module is used to perform baseline standardization processing on the multidimensional quantized feature vector, input the standardized feature vector into the artificial intelligence model to calculate the real-time comprehensive health score and future runaway risk probability of the WAT test machine, and predict future health trends. The decision-making module is used to generate on-demand verification decisions or maintenance warning work orders based on the comprehensive health score, future risk probability of loss of control, and future health trend prediction results in the evaluation result set, combined with preset threshold rules. The future health trend prediction results include the health score prediction sequence of the WAT test machine within a specified future time window, the confidence interval of the prediction sequence, and the expected time point when the health score prediction sequence first reaches the preset risk threshold.

[0016] Furthermore, to achieve the above objectives, the present invention also proposes an automatic monitoring device for the health status of WAT machines. The automatic monitoring device for the health status of WAT machines includes: a memory, a processor, and an automatic monitoring program for the health status of WAT machines stored in the memory and executable on the processor. The automatic monitoring program for the health status of WAT machines is configured to implement the steps of the automatic monitoring method for the health status of WAT machines as described above.

[0017] In addition, to achieve the above objectives, the present invention also proposes a storage medium storing a WAT ​​machine health automatic monitoring program, wherein the WAT machine health automatic monitoring program, when executed by a processor, implements the steps of the WAT machine health automatic monitoring method described above.

[0018] This invention collects and outputs raw multi-source datasets from multiple heterogeneous data sources related to WAT test equipment; processes the raw data and outputs hierarchically aligned time-series datasets; calculates multi-dimensional quantized feature vectors based on the current monitoring window and historical baseline windows, performs baseline standardization, and inputs them into an artificial intelligence model to calculate the real-time comprehensive health score, future runaway risk probability, and health trend prediction results of the test equipment; and generates on-demand verification decisions or maintenance warning work orders based on the comprehensive health score, risk probability, and prediction results combined with threshold rules. Through this method, this invention transforms traditional fixed-cycle verification into an on-demand verification mode based on the actual health status of the equipment, significantly reducing maintenance costs and unplanned downtime rates, and achieving real-time, quantitative, and interpretable monitoring of WAT equipment health. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the automatic monitoring method for the health status of WAT machines according to the present invention. Figure 2 This is a structural block diagram of the first embodiment of the WAT machine health automatic monitoring device of the present invention.

[0020] 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

[0021] 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.

[0022] This invention provides an automatic monitoring method for the health status of WAT (Wastewater Acquisition) machines, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the automatic monitoring method for the health status of WAT machines according to the present invention.

[0023] In this embodiment, the automatic health monitoring method for WAT machines includes the following steps: Step S10: Periodically collect full data from multiple heterogeneous data sources related to the WAT test machine, and output a time-aligned original multi-source dataset based on the full data.

[0024] In this embodiment, the executing entity is the WAT machine health automatic monitoring device, which has functions such as data processing, data communication and program execution. The WAT machine health automatic monitoring 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.

[0025] It should be noted that traditional WAT test machine health management mainly relies on a passive maintenance model. This means that problems are only identified after test data anomalies, yield declines, or machine errors occur. Troubleshooting is then conducted using engineers' manual experience combined with SPC control charts, and the machine's condition is restored through repair, calibration, or verification. Furthermore, routine machine calibration typically employs a fixed-cycle preventative maintenance strategy, performed according to plan regardless of the machine's actual condition. Existing technology suffers from the following drawbacks: strong lag, intervention is only possible after problems have caused test anomalies or production line downtime, failing to detect gradual degradation in advance; significant resource waste, as even healthy and stable machines still require fixed-cycle calibration, consuming substantial manpower, time, and spare parts costs; strong subjectivity in judgment, relying on manual experience to analyze multi-source data, easily leading to missed diagnoses or over-intervention; and a lack of forward-looking prediction, failing to effectively integrate multi-source information such as equipment operating status and maintenance history for comprehensive risk assessment and degradation trend prediction.

[0026] To address the aforementioned technical issues, this embodiment collects and outputs raw multi-source datasets from multiple heterogeneous data sources related to the WAT test equipment; processes the raw data and outputs a hierarchically aligned time-series dataset; calculates multi-dimensional quantized feature vectors based on the current monitoring window and historical baseline windows, performs baseline standardization, and inputs them into an artificial intelligence model to calculate the real-time comprehensive health score, future runaway risk probability, and health trend prediction results of the test equipment; and generates on-demand verification decisions or maintenance warning work orders based on the comprehensive health score, risk probability, and prediction results combined with threshold rules. Through this method, the present invention transforms traditional fixed-cycle verification into an on-demand verification mode based on the actual health status of the equipment, significantly reducing maintenance costs and unplanned downtime rates, and achieving real-time, quantitative, and interpretable monitoring of the health of WAT test equipment. Specifically, it can be implemented as follows.

[0027] In its implementation, the overall process of this solution involves the system first collecting all relevant data from a multi-source heterogeneous system to form a continuous time-series monitoring stream; then, rigorously cleaning, layering, and aligning the raw data to ensure data quality and consistency; constructing multi-dimensional quantitative features based on the clean data to transform abstract equipment status and testing behavior into calculable numerical indicators; establishing dual baselines for individuals and groups using historical stable data and standardizing the features to form a unified health evaluation standard; next, employing a multi-source fusion AI model to comprehensively score the features, calculate risk probability, and predict future trends, while providing confidence and interpretability outputs; finally, generating on-demand verification triggers, priority ranking, and proactive maintenance warnings based on the model results, thus realizing the transformation from "passive maintenance" to "predictive maintenance."

[0028] Furthermore, the overall system architecture in this embodiment includes at least the following software functional modules, which can be deployed on servers or edge computing devices and form a closed-loop data flow through data bus / database / API connections. Data Acquisition Module: Extracts data from the test result library, SPC system, MES, maintenance system, consumables system, sensors, and environmental system; supports real-time streaming and timed batch processing. Outputs the raw data stream and extracted water level. Data Governance Module: Performs cleaning, deduplication, layering, alignment, and quality scoring on the raw data, outputting a "layered aligned time sequence frame" and a quality score. Feature building module: In and The system calculates features such as drift, fluctuation, trend, alarm, downtime, maintenance effect, and spatial pattern, and outputs the unstandardized feature vector. The baseline management module filters stable segments, establishes individual and group baselines, continuously updates baseline parameters, and outputs the features required for robust standardization. Etc. Health Inference Module: Input standardized feature vector, output health score. Risk probability With explanatory information (such as Top-K feature contributions). Trend prediction module: for Perform multi-step predictions with key feature sequences, outputting the future sequence, confidence interval, and time to threshold. Confidence assessment module: outputs confidence scores based on data quality, sample size, and model uncertainty. And execute gating / degradation strategies. Decision and work order module: based on The rule base outputs "immediate verification / warning / exemption" suggestions, calculates Priority, and generates structured work order records. The storage and auditing module stores raw data, features, baselines, model versions, decision records, and evidence chains, supporting traceability and playback.

[0029] In this embodiment, the multiple heterogeneous data sources include at least a test result database, a statistical process control system (SPC), and a manufacturing execution system (MES). Specifically, the full data includes: a test result database containing the measured values, pass / fail judgments, upper / lower limit specs, wafer ID, test site, chip coordinates, lot information, timestamps, etc., for each test item. The SPC system contains key control chart indicators (such as mean, standard deviation, Cp / Cpk), out-of-limit event records, and trigger logs for trend / pattern violations of Western Electric rules. The MES contains machine status codes, alarm codes and descriptions, downtime / standby / runtime statistics, equipment utilization, throughput, cycle time, and other operational indicators. The maintenance management system records repair, maintenance, and periodic verification events, including event type, execution time, participating personnel, and comparison of test results after retesting. The consumables management system contains configuration information, replacement records, and cumulative usage counts for key consumables such as probe cards and handlers. The probe station sensors monitor physical parameters such as temperature, pressure, and vibration in real time. Environmental monitoring system (if available): Environmental variables such as cleanroom temperature and humidity, and particle size. In one embodiment, the above process is specifically implemented, for example, by using the machine ID as the global primary key to establish a continuous time-series data stream, ensuring that the data of each machine forms an independent timeline. Data extraction is organized and extracted according to a multi-dimensional structure of "machine ID - product family - recipe - time window," where the time window is divided into a near-real-time monitoring window (the most recent 24-48 hours, used for immediate early warning) and a historical stability window (the most recent 3-6 months, used for baseline construction and model training). Event anchors are established for key events (such as maintenance / calibration completion, SPC rule triggering, alarm code aggregation, and sudden yield drop), and comparative data windows of 7-14 days before and after the event are automatically extracted for subsequent evaluation of the event's impact. An incremental data retrieval mechanism is implemented: under normal circumstances, full historical data is periodically retrieved; when an abnormal signal is detected (such as SPC triggering), incremental supplementation is immediately initiated to ensure that data is updated within minutes after the occurrence of abnormal symptoms.

[0030] The output is a structured relational data table or time-series database record, containing fields such as: machine ID, timestamp, genealogy dimension key, and multi-source original indicator values. The purpose of this step is to construct a "digital profile of machine health," covering the entire chain of information including test output quality, equipment operating status, and maintenance history.

[0031] Step S20: Perform cleaning, hierarchical processing and alignment processing on the original multi-source dataset to output a hierarchically aligned time-series dataset.

[0032] In the specific implementation, layered processing involves grouping and storing data into multiple independent subsets according to at least one dimension: product family tree, test program version, test temperature level, and site configuration. Alignment processing involves resampling the timeline at the minute or event level to align at least two types of heterogeneous data from test results, statistical process control events, equipment state transitions, alarm logs, and maintenance records to a unified timeline. Specific implementations include cleaning, which involves using statistical methods to handle missing values ​​(filling in with the median of the same item within the same family tree to avoid the mean being affected by extreme values); using the 3σ rule combined with business thresholds to mark outliers (retaining original values ​​and adding anomaly flags for later feature usage); unifying the timestamp to the UTC standard time zone and correcting for clock drift; and deduplicating based on a composite key (machine ID + timestamp + test item name) to prevent duplicate log entries. Quality filtering includes removing non-representative samples through a configurable rule table, such as the first 100-200 wafers during the machine warm-up phase, data explicitly marked as engineering lot, and data from known process change transition periods, while recording the filtering ratio and reason statistics for data quality auditing. Layering involves grouping and storing data according to key dimensions such as product lineage, recipe version, test temperature level, and site configuration, forming multiple independent subsets to avoid the natural differences under different test conditions from confusing health judgments and ensure that the model learns degradation patterns under the same conditions. Alignment involves resampling the timeline at the minute or event level to align heterogeneous data such as test results, SPC events, equipment state transitions, alarm logs, and maintenance records to a unified timeline; at the same time, it extracts "vertical self-comparison" data (the mean / distribution of the machine's own historical stable period) and "horizontal group comparison" data (the baseline of other machines in the same lineage during the same period) to form a multi-source fusion structured input frame. The final output is a clean, aligned layered time series dataset with an additional data quality score (e.g., completeness rate >95%, anomaly labeling rate <5%), which is subsequently used as an important input for model confidence.

[0033] Step S30: Output a multidimensional quantized feature vector based on the predefined current monitoring window and historical baseline window and the hierarchical aligned time series dataset.

[0034] It should be noted that the multidimensional quantitative feature vector in this embodiment includes test distribution and stability features, statistical process control features, equipment operation and reliability features, maintenance and verification effect features, and wafer diagram spatial structure features. The test distribution and stability features include at least one or more of the following sub-features: distribution drift feature, used to quantify the degree of deviation of the overall distribution position of test data between the current monitoring window and the historical baseline window; fluctuation amplification feature, used to quantify the magnitude of change in the dispersion of test data within the current monitoring window relative to the historical baseline window; and tail risk feature, used to quantify the proportion and magnitude of test data exceeding specification limits or control limits within the current monitoring window. The statistical process control features include at least one or more of the following sub-features: exceedance intensity features, used to quantify the cumulative frequency and magnitude of exceedance events in the SPC control chart per unit time; trend persistence features, used to quantify the duration and degree of deviation of continuous violations of SPC discrepancy rules; and recovery time features, used to quantify the number of tests or time required for test data to return to normal after an SPC event is triggered. The equipment operation and reliability characteristics include at least one or more of the following sub-features: alarm density feature, used to quantify the frequency of machine alarms per unit time; alarm clustering feature, used to quantify the co-occurrence relationship and clustering intensity of different alarm codes within the same time window; downtime variation feature, used to quantify the relative change rate of the cumulative downtime of the machine between the current monitoring window and the historical baseline window; throughput fluctuation feature, used to quantify the fluctuation amplitude of the machine's output per unit time or cycle time. The maintenance and verification effect characteristics are used to quantify the changes in the machine's health status before and after a maintenance or verification event, as well as the average interval between the recurrence of similar problems. The spatial structure features of the wafer image include at least one or more of the following sub-features: spatial pattern similarity features, used to quantify the degree of spatial structural similarity between two wafer images in terms of defect distribution, electrical parameter distribution, or yield distribution; and abnormal hotspot features, used to quantify the intensity of systematic deviation of local areas on the wafer image relative to the overall distribution.

[0035] In one embodiment, the test distribution and stability characteristics include: (a) Distribution drift: The cumulative distribution difference between the current window and the baseline window was quantified using the Kolmogorov-Smirnov test statistic.

[0036] in, The KS statistic (dimensionless, range) The larger the value, the greater the difference between the current distribution and the baseline distribution. For all possible Take the "upper bound / maximum value" The value of the test item (the unit depends on the test item, such as V, Ω, A, μA, etc.). This is the empirical cumulative distribution function (ECDF) of the samples within the current monitoring window. This represents the ECDF of samples within the baseline window.

[0037] (b) Increased volatility: Using robust median absolute deviation (MAD) as the standard deviation estimate is more robust to long-tailed distributions (such as leakage currents);

[0038] in: For the current window Sample values ​​(units same as test items). For sample set the median of This is the absolute deviation of the median (the core definition of MAD). It is a commonly used scaling factor, so that when the data is approximately normally distributed, the MAD magnitude can be aligned to the standard deviation.

[0039] (c) Additional characteristics: tail risk ratio (exceeding 3σ ratio), skewness variation, kurtosis variation, etc.

[0040] In one embodiment, SPC and process control characteristics include: out-of-limit intensity, such as the cumulative number of out-of-limit points multiplied by the average out-of-limit magnitude ((actual value - spec) / spec); trend persistence, such as the duration and intensity of consecutive violations of Western Electric rules (e.g., 7 consecutive points on the same side); and recovery time, such as the number of test pieces or time required for the SPC to return to normal after triggering.

[0041] (3) Equipment operation and reliability characteristics include: (a) Alarm density and clustering, such as the number of alarms per unit time (per hour), the repetition rate of the same alarm code, and alarm code co-occurrence clustering (mining using association rules). (b) Rate of change in downtime:

[0042] in, This represents the relative rate of change of downtime (dimensionless). This represents the cumulative downtime within the current monitoring window. The cumulative downtime within the baseline window (or the same caliber window for historical stable periods) (units as above). It is a small constant (unit same as downtime) used to avoid Divide by zero; in engineering, values ​​such as 1e-6 × units or 1 minute are often used that "do not change the magnitude but can prevent anomalies". (c) Throughput / cycletime fluctuation: rolling window standard deviation and trend slope.

[0043] In one embodiment, the maintenance and verification effect features include (a) Changes in health indicators before and after the event (used to quantify recovery quality; large positive values ​​indicate adequate recovery, while values ​​close to zero or negative indicate "repaired but not fully recovered"):

[0044] in, The change after the event relative to before the event (units same) ), , These represent the health indicators / health scores before and after maintenance / verification.

[0045] (b) Recurrence interval: the average time for the same problem to recur.

[0046] In one embodiment, the Map and spatial structure features include: (a) Spatial pattern similarity: After converting the wafer map into vectors, the cosine similarity is calculated.

[0047] in, , This is a vector representation of the wafer map (e.g., defect counts, intensity, or one-hot structural features expanded by die / mesh). It is the dot product (inner product). , Let L be the L2 norm (length) of the vector. Result range (If the vector elements are non-negative, the common range is) The closer to 1, the more similar the spatial patterns.

[0048] (b) Anomaly strength: Local hotspot detection (e.g., Getis-OrdGi* statistics) or edge / center systematic bias. The output is a standardized feature vector for each machine and each genealogy, which is convenient for subsequent model input.

[0049] Step S40: Perform baseline standardization on the multidimensional quantized feature vector, input the standardized feature vector into the artificial intelligence model to calculate the real-time comprehensive health score and future runaway risk probability of the WAT test machine, and predict future health trends.

[0050] In this specific implementation, the baseline standardization process for the multidimensional quantized feature vector in this embodiment includes: filtering stable segment data based on the aligned time series dataset, wherein the stable segment data meets the conditions of no SPC triggering, no maintenance events, no alarm aggregation for a consecutive preset number of days, and no significant difference between the distribution of key test items and the reference distribution; establishing an individual baseline for each WAT test machine based on the stable segment data, and outputting a baseline parameter set, wherein the baseline includes at least position parameters and scale parameters; and standardizing the multidimensional feature vector based on the baseline parameter set and the multidimensional quantized feature vector using a standardization method based on robust statistics, and outputting the standardized feature vector.

[0051] In practical applications, baseline establishment includes the following processes: Automatically selecting historically stable segments, such as data segments that meet the criteria of "no SPC triggers, no maintenance events, no alarm clusters for more than 30 consecutive days, and a key item distribution KS test p-value > 0.05". Calculating individual baselines, such as the mean mu_i and robust standard deviation (MAD) for each machine across all lineages. Calculating the group baseline, such as the weighted average mu_g of all machines in the same lineage (weights can be based on runtime). The baseline is updated daily to adapt to process fine-tuning. Feature standardization: Robust Z-scores are used to resist the influence of outliers, for example... .

[0052] Furthermore, the formula for calculating the comprehensive health score is as follows:

[0053] in, To calculate the overall health score, To include the number of feature dimensions in the fusion, For the first The weights of each feature For the first Robust Z-scores for each feature The values ​​are absolute, indicating that the deviation intensity is not differentiated by positive or negative direction. H < 1.0 is normal (green); 1.0-2.0 is of concern (yellow, requires monitoring); 2.0-3.0 is at risk (orange, intervention recommended); > 3.0 is out of control (red, must be stopped immediately for verification). It can be 50–100 dimensions. It can be set to 1.5-2.0.

[0054] In one embodiment, predicting future health trends specifically includes acquiring historical time-series data, using a long short-term memory network to perform multi-step predictions on the comprehensive health score and the time series of key risk features extracted from the historical time-series data, and outputting the predicted sequence, confidence interval, and expected time point for reaching the risk threshold within a specified number of days in the future; wherein, the multi-step prediction formula is:

[0055] in, At the current time point, To predict the number of steps, For lookback length, it indicates that the model input uses past data. Data at each time step For the past A sequence of health scores for each step. For the past Other covariates / key risk feature sequences of the step, For the future Step-by-step health score prediction sequence This is a time-series mapping function implemented using a Long Short-Term Memory (LSTM) network.

[0056] Furthermore, this embodiment also involves confidence assessment:

[0057] in, The confidence level is set to 0–1. When C < 0.7, the output is automatically downgraded to "Insufficient data, manual review or supplementary monitoring is recommended" to avoid misleading decisions based on low confidence. These are the weighting coefficients. .

[0058] It should be noted that the artificial intelligence model in this embodiment adopts a multi-source fusion health inference model, specifically using an ensemble learning algorithm, such as Random Forest or Gradient Boosting Tree (XGBoost / GBDT). The full-dimensional feature vector from step 3 is input, and the outputs are a health score H, a future risk of runaway risk P (in the 0-1 range), and a SHAP value (feature importance and contribution explanation). The training process involves collecting historical data, using actual test anomalies / runaway events or forced verification events as positive sample labels, and performing supervised binary classification / regression training. Cross-validation is used to optimize hyperparameters (tree depth, learning rate, etc.).

[0059] Step S50: Based on the comprehensive health score, the probability of future loss of control risk, and the prediction results of future health trends in the assessment result set, and combined with preset threshold rules, generate an on-demand verification decision or maintenance early warning work order.

[0060] In specific implementation, the conditions for generating on-demand verification decisions include: generating an immediate verification decision when the probability of future loss of control is greater than a first risk threshold; generating an immediate verification decision when the future health trend prediction result shows that the comprehensive health score will exceed a second risk threshold within a preset future time period; and generating a verification exemption suggestion when the comprehensive health score is continuously lower than a third risk threshold, the probability of future loss of control is continuously lower than a fourth risk threshold, and no abnormal features are triggered within a consecutive first preset number of days to a second preset number of days. For example, in specific implementation, when the risk probability P>0.6 or the trend prediction shows that H will exceed 3.0 within the next 3-5 days, an "immediate verification" work order is automatically generated; when H<1.0 and P<0.1 for 90-180 consecutive days and no abnormal features are detected, a "verification exemption" suggestion is generated with an evidence summary, supporting an extension of the verification period to 1.5-2 times the original period.

[0061] For example, a forward-looking warning work order can be generated for a machine that is "deteriorating but not yet out of control" (H is in the range of 1.5-2.5 and is on the rise). The work order includes: a description of the risk type, the expected time window for the risk to arrive, evidence of the top five contributing characteristics, and recommended preventive actions (such as replacing the probe card in advance or cleaning the area).

[0062] In one embodiment, priority ranking is also involved, specifically calculating a comprehensive priority score:

[0063] Among them, Priority is the overall priority score (dimensionless), and the larger the value, the higher the priority for maintenance / scheduling. w1, w2, w3, and w4 are weighting coefficients (dimensionless) used to reflect the importance of each factor; they are generally non-negative and are often set to satisfy w1+w2+w3+w4=1. P is the basic risk score / current risk intensity (dimensionless), which can be mapped from the health score H, anomaly score, alarm level, etc. (e.g., mapped to 0–1 or 0–100). R (impact range): the number of key product families covered (unit: number), indicating how many key product families / product groups may be affected by the machine problem. This is a historical recurrence count (unit: times), representing the number of times the same / similar problem occurs repeatedly within a statistical window. The probability of insufficient recovery (range: 0–1) represents the probability assessment result that the recovery may still be insufficient or deteriorate again after maintenance / verification.

[0064] Finally, a scheduling list is output, which is used by the maintenance department for resource allocation and scheduling decisions. The system sorts the machines according to overall priority and outputs the machine number / workstation location, current health score H and risk level, risk type and key evidence summary (such as main contributing characteristics / abnormal indicators), scope of impact (number of key product families covered), expected deterioration time window and recommended latest handling time for each machine, and provides recommended maintenance actions (such as replacing probe cards, local cleaning, calibration, etc.) and estimated man-hours / required spare parts and personnel skill types. The list can be filtered and exported (Excel / CSV) by priority, production line and deadline, so that maintenance supervisors can quickly arrange manpower and downtime windows.

[0065] In this embodiment, raw multi-source datasets are collected and output from multiple heterogeneous data sources related to the WAT test equipment. The raw data is processed to output a hierarchically aligned time-series dataset. Multi-dimensional quantized feature vectors are calculated based on the current monitoring window and historical baseline windows, and after baseline standardization, they are input into an artificial intelligence model. The model calculates the real-time comprehensive health score, future runaway risk probability, and health trend prediction results for the test equipment. Based on the comprehensive health score, risk probability, and prediction results, combined with threshold rules, an on-demand verification decision or maintenance warning work order is generated. Through this method, the present invention transforms traditional fixed-cycle verification into an on-demand verification mode based on the actual health status of the equipment, significantly reducing maintenance costs and unplanned downtime rates, and achieving real-time, quantitative, and interpretable monitoring of the health of WAT test equipment.

[0066] Furthermore, this embodiment of the invention also proposes a storage medium storing a WAT ​​machine health automatic monitoring program, which, when executed by a processor, implements the steps of the WAT machine health automatic monitoring method described above.

[0067] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the WAT machine health automatic monitoring device of the present invention.

[0068] like Figure 2 As shown, the automatic health monitoring device for WAT machines proposed in this embodiment of the invention includes: The data acquisition module 10 is used to periodically acquire full data from multiple heterogeneous data sources related to the WAT test machine, and output a time-aligned original multi-source dataset based on the full data. The multiple heterogeneous data sources include at least a test result database, a statistical process control system, and a manufacturing execution system. Data governance module 20 is used to perform cleaning, hierarchical processing and alignment processing on the original multi-source dataset, and output hierarchical aligned time series dataset; The feature construction module 30 is used to output a multi-dimensional quantized feature vector based on the predefined current monitoring window and historical baseline window and the hierarchical aligned time series dataset, wherein the multi-dimensional quantized feature vector is used to reflect the health status of the WAT test machine; The health inference module 40 is used to perform baseline standardization processing on the multidimensional quantized feature vector, input the standardized feature vector into the artificial intelligence model, calculate the real-time comprehensive health score and future runaway risk probability of the WAT test machine, and predict future health trends. The decision module 50 is used to generate on-demand verification decisions or maintenance warning work orders based on the comprehensive health score, future loss of control risk probability and future health trend prediction results in the evaluation result set, combined with preset threshold rules. The future health trend prediction results include the health score prediction sequence of the WAT test machine in a future specified time window, the confidence interval of the prediction sequence, and the expected time point when the health score prediction sequence first reaches the preset risk threshold.

[0069] In this embodiment, raw multi-source datasets are collected and output from multiple heterogeneous data sources related to the WAT test equipment. The raw data is processed to output a hierarchically aligned time-series dataset. Multi-dimensional quantized feature vectors are calculated based on the current monitoring window and historical baseline windows, and after baseline standardization, they are input into an artificial intelligence model. The model calculates the real-time comprehensive health score, future runaway risk probability, and health trend prediction results for the test equipment. Based on the comprehensive health score, risk probability, and prediction results, combined with threshold rules, an on-demand verification decision or maintenance warning work order is generated. Through this method, the present invention transforms traditional fixed-cycle verification into an on-demand verification mode based on the actual health status of the equipment, significantly reducing maintenance costs and unplanned downtime rates, and achieving real-time, quantitative, and interpretable monitoring of the health of WAT test equipment.

[0070] This application embodiment also provides an automatic health monitoring device for WAT machines, 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 automatic health monitoring program for WAT machines. When the processor executes the program stored in the memory, it implements the above-mentioned automatic health monitoring method for WAT machines.

[0071] The communication bus mentioned in the aforementioned WAT machine health automatic monitoring equipment 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.

[0072] The communication interface is used for communication between the aforementioned WAT machine health automatic monitoring equipment and other devices.

[0073] 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.

[0074] 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.

[0075] 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)).

[0076] 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.

[0077] 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.

[0078] 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.

[0079] 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.

[0080] 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.

[0081] In addition, for technical details not described in detail in this embodiment, please refer to the automatic monitoring method for WAT machine health provided in any embodiment of the present invention, which will not be repeated here.

[0082] 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.

[0083] 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.

[0084] 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.

[0085] 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.

[0086] 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 automatically monitoring the health status of a WAT ​​machine, characterized in that, The automatic health monitoring method for WAT machines includes: Full data is periodically collected from multiple heterogeneous data sources related to the WAT test machine, and a time-aligned original multi-source dataset is output based on the full data. The multiple heterogeneous data sources include at least a test result database, a statistical process control system, and a manufacturing execution system. The original multi-source dataset is cleaned, layered, and aligned to output a layered aligned time-series dataset. Based on the predefined current monitoring window and historical baseline window, and the hierarchical aligned time series dataset, a multidimensional quantized feature vector is output, wherein the multidimensional quantized feature vector is used to reflect the health status of the WAT test machine. The multidimensional quantized feature vector is baseline-normalized, and the normalized feature vector is input into an artificial intelligence model to calculate the real-time comprehensive health score and future runaway risk probability of the WAT test machine, and to predict future health trends. Based on the comprehensive health score, future risk of loss of control probability, and future health trend prediction results in the assessment result set, combined with preset threshold rules, an on-demand verification decision or maintenance early warning work order is generated. The future health trend prediction results include the health score prediction sequence of the WAT test machine within a specified future time window, the confidence interval of the prediction sequence, and the expected time point when the health score prediction sequence first reaches the preset risk threshold.

2. The automatic health monitoring method for WAT machines as described in claim 1, characterized in that, The hierarchical processing involves grouping and storing data into multiple independent subsets according to at least one dimension, such as product family tree, test program version, test temperature level, and site configuration; the alignment processing involves resampling the timeline at the minute or event level to align at least two types of heterogeneous data, such as test results, statistical process control events, equipment state transitions, alarm logs, and maintenance records, to a unified timeline.

3. The automatic health monitoring method for WAT machines as described in claim 1, characterized in that, The multidimensional quantitative feature vector includes test distribution and stability features, statistical process control features, equipment operation and reliability features, maintenance and verification effect features, and wafer diagram spatial structure features. The test distribution and stability features include at least one or more of the following sub-features: distribution drift feature, used to quantify the degree of deviation of the overall distribution position of test data between the current monitoring window and the historical baseline window; fluctuation amplification feature, used to quantify the magnitude of change in the dispersion of test data within the current monitoring window relative to the historical baseline window; and tail risk feature, used to quantify the proportion and magnitude of test data exceeding specification limits or control limits within the current monitoring window. The statistical process control features include at least one or more of the following sub-features: exceedance intensity features, used to quantify the cumulative frequency and magnitude of exceedance events in the SPC control chart per unit time; trend persistence features, used to quantify the duration and degree of deviation of continuous violations of SPC discrepancy rules; and recovery time features, used to quantify the number of tests or time required for test data to return to normal after an SPC event is triggered. The equipment operation and reliability characteristics include at least one or more of the following sub-features: alarm density feature, used to quantify the frequency of machine alarms per unit time; alarm clustering feature, used to quantify the co-occurrence relationship and clustering intensity of different alarm codes within the same time window; downtime variation feature, used to quantify the relative change rate of the cumulative downtime of the machine between the current monitoring window and the historical baseline window; throughput fluctuation feature, used to quantify the fluctuation amplitude of the machine's output per unit time or cycle time. The maintenance and verification effect characteristics are used to quantify the changes in the machine's health status before and after a maintenance or verification event, as well as the average interval between the recurrence of similar problems. The spatial structure features of the wafer image include at least one or more of the following sub-features: spatial pattern similarity features, used to quantify the degree of spatial structural similarity between two wafer images in terms of defect distribution, electrical parameter distribution, or yield distribution; and abnormal hotspot features, used to quantify the intensity of systematic deviation of local areas on the wafer image relative to the overall distribution.

4. The automatic health monitoring method for WAT machines as described in claim 1, characterized in that, The baseline normalization process for the multidimensional quantized feature vector includes: The stable segment data is filtered according to the aligned time series dataset, wherein the stable segment data meets the conditions of no SPC triggering, no maintenance events, no alarm aggregation for a consecutive preset number of days, and the distribution of key test items is not significantly different from the reference distribution. Based on the stable segment data, an individual baseline is established for each WAT test machine, and a baseline parameter set is output. The baseline includes at least position parameters and scale parameters. Based on the baseline parameter set and the multidimensional quantized feature vector, the multidimensional feature vector is standardized using a standardization method based on robust statistics, and the standardized feature vector is output.

5. The automatic health monitoring method for WAT machines as described in claim 1, characterized in that, The formula for calculating the comprehensive health score is as follows: in, To calculate the overall health score, To include the number of feature dimensions to be incorporated into the fusion, For the first The weights of each feature, For the first Robust Z-scores for each feature To take the absolute value, it means that the deviation intensity does not distinguish between positive and negative directions.

6. The automatic health monitoring method for WAT machines as described in claim 1, characterized in that, The prediction of future health trends includes: Historical time-series data is acquired, and a Long Short-Term Memory (LSTM) network is used to perform multi-step predictions on the comprehensive health score and the time series of key risk features extracted from the historical time-series data. The predicted sequence, confidence interval, and expected time point for reaching the risk threshold within a specified number of days are output. The multi-step prediction formula is as follows: in, At the current time point, To predict the number of steps, For lookback length, it indicates that the model input uses past data. Data at each time step For the past A sequence of health scores for each step. For the past Other covariates / key risk feature sequences of the step, For the future Step-by-step health score prediction sequence This is a time-series mapping function implemented using a Long Short-Term Memory (LSTM) network.

7. The automatic health monitoring method for WAT machines as described in claim 1, characterized in that, The conditions for generating the on-demand verification decision include: An immediate verification decision is generated when the probability of future loss of control exceeds a first risk threshold. When the future health trend prediction results show that the comprehensive health score will exceed the second risk threshold within a preset future time period, an immediate verification decision is generated. If, within a consecutive first preset number of days to a second preset number of days, the overall health score remains below the third risk threshold, the probability of future loss of control remains below the fourth risk threshold, and no abnormal features are triggered, a verification exemption suggestion is generated.

8. An automatic health monitoring device for WAT (Wastewater Acquisition) machines, characterized in that, The WAT machine health automatic monitoring device is applied to the WAT machine health automatic monitoring method as described in any one of claims 1 to 7, and the device comprises: The data acquisition module is used to periodically acquire full data from multiple heterogeneous data sources related to the WAT test machine, and output a time-aligned original multi-source dataset based on the full data. The multiple heterogeneous data sources include at least a test result database, a statistical process control system, and a manufacturing execution system. The data governance module is used to perform cleaning, hierarchical processing and alignment processing on the original multi-source dataset, and output a hierarchical aligned time series dataset. The feature construction module is used to output a multidimensional quantized feature vector based on the predefined current monitoring window and historical baseline window and the hierarchical aligned time series dataset, wherein the multidimensional quantized feature vector is used to reflect the health status of the WAT test machine; The health inference module is used to perform baseline standardization processing on the multidimensional quantized feature vector, input the standardized feature vector into the artificial intelligence model to calculate the real-time comprehensive health score and future runaway risk probability of the WAT test machine, and predict future health trends. The decision-making module is used to generate on-demand verification decisions or maintenance warning work orders based on the comprehensive health score, future risk probability of loss of control, and future health trend prediction results in the evaluation result set, combined with preset threshold rules. The future health trend prediction results include the health score prediction sequence of the WAT test machine within a specified future time window, the confidence interval of the prediction sequence, and the expected time point when the health score prediction sequence first reaches the preset risk threshold.

9. An automatic health monitoring device for WAT (Wastewater Acquisition) machines, characterized in that, The WAT machine health automatic monitoring device includes: a memory, a processor, and a WAT ​​machine health automatic monitoring program stored in the memory and executable on the processor. The WAT machine health automatic monitoring program is configured to implement the steps of the WAT machine health automatic monitoring method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores an automatic monitoring program for the health status of WAT machines. When the automatic monitoring program for the health status of WAT machines is executed by the processor, it implements the steps of the automatic monitoring method for the health status of WAT machines as described in any one of claims 1 to 7.

Images