AI deep learning-based semiconductor chip test management and control system
The semiconductor chip test management system based on AI deep learning uses XGBoost and Bi-LSTM algorithms for multi-dimensional data analysis, which solves the problems of low accuracy in yield control and low efficiency in fault root cause location in semiconductor chip testing, and realizes automated spare parts management and rapid fault response.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TONGFU CHAOWEI (SUZHOU) MICROELECTRONICS CO LTD
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-24
AI Technical Summary
In the current semiconductor chip testing process, the accuracy of yield control is low, the efficiency of fault root cause location is low, the decision-making on spare parts replacement is lagging and the accuracy is low, there is a lack of a multi-dimensional data integrated processing framework, and AI deep learning algorithms are not deeply integrated with yield control, fault diagnosis, and spare parts management.
A semiconductor chip test management system based on AI deep learning is adopted, including a data acquisition and preprocessing module, a real-time monitoring and early warning module, a fault root cause analysis and decision-making module, and a spare parts recommendation and execution module. The system uses XGBoost and Bi-LSTM algorithms to perform multi-dimensional data analysis and fault matching, and builds a full-dimensional yield difference analysis model and a fault-spare parts intelligent matching model.
It enables accurate identification of yield differences under multi-dimensional segmentation, quickly and objectively locates the root cause of failure, automatically generates spare parts replacement plans, shortens maintenance response time, improves spare parts replacement accuracy, and reduces unplanned downtime.
Smart Images
Figure CN122451331A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor manufacturing and intelligent control technology, and in particular to a semiconductor chip testing and control system based on AI deep learning. Background Technology
[0002] In the current semiconductor chip testing process, including Final Test (FT, the final product testing stage after chip packaging) and System Level Test (SLT, a comprehensive testing stage simulating actual chip application scenarios), yield control is the most critical node. Yield control relies heavily on traditional manual statistical analysis and a basic threshold alarm system. Its core logic is: a unified yield pass threshold is pre-set; when test data falls below this threshold, an alarm is triggered. During troubleshooting, engineers manually compare test failure information, equipment operation logs, and historical spare parts replacement records, combining this with their personal experience to determine the root cause of the failure and develop a spare parts replacement plan.
[0003] Existing yield control solutions have at least the following problems: 1. Low accuracy in identifying yield differences: Traditional methods use a single threshold standard and do not consider multi-dimensional subdivided data such as product model number (OPN, Original Part Number), production batch number (Lot), machine number, test site (Site) number, test program version, and test code. They cannot distinguish between reasonable yield fluctuations and abnormal differences in different scenarios, resulting in unclear identification of inefficient links. 2. Inefficient root cause location of failures: It relies on engineers' personal experience to compare failure information with spare parts records, lacks a standardized analysis system, is greatly affected by subjective factors, and the analysis process is time-consuming, which cannot meet the real-time requirements of large-scale production. 3. Delayed and inaccurate decision-making for spare parts replacement: Spare parts replacement requires manual verification of fault information and inventory status one by one, which has a long response cycle and is prone to incorrect or missed replacements, resulting in extended unplanned downtime.
[0004] After research, the applicant believes that the root cause of the above problems is that the existing technology lacks an integrated processing framework for multi-dimensional data of semiconductor chip testing, and fails to deeply integrate AI deep learning algorithms with the entire process of yield control, fault diagnosis, and spare parts management, resulting in a disconnect between data processing, analysis and decision-making, and execution. Summary of the Invention
[0005] This invention provides a semiconductor chip test and control system based on AI deep learning, addressing the aforementioned problems in existing yield control solutions. The technical solution is as follows: Embodiments of the present invention provide a semiconductor chip test control system based on AI deep learning, comprising: The data acquisition and preprocessing module is used to acquire multi-dimensional test data from the test equipment and preprocess the multi-dimensional test data to generate a standardized dataset. The multi-dimensional test data includes at least: product model number (OPN), production batch number (Lot), machine number, test site number (Site number), test program version, test code, test results, test failure information, failure scenario information, and spare parts replacement records. The real-time monitoring and early warning module is used to input the first real-time dataset preprocessed by the data acquisition and preprocessing module into the trained full-dimensional yield difference analysis model, obtain the real-time weighted average yield output by the full-dimensional yield difference analysis model, and determine whether the difference between the real-time weighted average yield and the baseline yield exceeds the yield difference threshold. If it does, an early warning message is generated. The first real-time dataset includes at least OPN, Lot, machine number, site number, test program version, test code, and test result. The full-dimensional yield difference analysis model is obtained by hierarchically grouping and training the historical OPN, historical Lot, historical machine number, historical site number, historical test program version, historical test code, and historical test result in the standardized dataset as training samples according to a preset dimension combination. The fault root cause analysis and decision-making module is used to parse the Fail information and fault scenario information in the fault data when a fault is detected, and input them into a trained fault-spare part intelligent matching model to obtain the fault root cause, corresponding confidence level, and corresponding spare part output by the fault-spare part intelligent matching model; wherein the fault-spare part intelligent matching model is trained using historical Fail information, historical fault scenario information, and historical spare part replacement records in the standardized dataset as training samples and using a Bi-LSTM model. The spare parts recommendation and execution module is used to confirm the availability of the corresponding spare parts based on the corresponding spare parts output by the fault-spare parts intelligent matching model, and generate a standardized replacement plan.
[0006] Optionally, it also includes: a model training and optimization module; The model training and optimization module is used to take the historical OPN, historical Lot, historical machine ID, historical site ID, historical test program version, historical test code, and historical test results in the standardized dataset as training samples, and to stratify and group them according to the dimensions of OPN-Lot-machine ID-Site ID-test program version-Test code to obtain multiple groups. Based on the historical average yield of each group, the weighting coefficients of the test data of each dimension are optimized by the gradient boosting tree (XGBoost) algorithm to finally obtain the trained full-dimensional yield difference analysis model. Furthermore, it is used to use historical Fail information, historical fault scenario information, and historical spare parts replacement records as training samples, and employs the Bi-LSTM algorithm to establish a mapping relationship between fault features and spare parts, thereby constructing an initial fault-spare parts intelligent matching model. The initial fault-spare parts intelligent matching model is trained until the accuracy of the output result of the initial fault-spare parts intelligent matching model reaches a preset value, at which point training stops, and the trained fault-spare parts intelligent matching model is obtained.
[0007] Optionally, the model training and optimization module is further configured to automatically trigger iterative training of the model when preset conditions are met; the preset conditions include: a fixed period, and / or, when the amount of new data exceeds a preset number of records.
[0008] Optionally, the model training and optimization module is specifically used to automatically trigger iterative training of the fault-spare part intelligent matching model whenever a preset number of valid fault-maintenance records are added.
[0009] Optionally, the data acquisition and preprocessing module includes: The data interface submodule is used to collect multi-dimensional test data of the test equipment in real time. The multi-dimensional test data includes OPN, Lot, machine number, site number, test program version, test code, test results, Fail information, fault scenario information, and spare parts replacement records. The preprocessing submodule is used to clean, transform, and standardize the multi-dimensional test data collected by the data interface submodule to generate a standardized dataset.
[0010] Optionally, the data interface submodule integrates an OPC UA protocol conversion unit and a data verification unit, supporting protocol adaptation with different models of FT / SLT test equipment.
[0011] Optionally, the AI-based deep learning-based semiconductor chip test management system integrates a local AI server; The local AI server is connected to the data acquisition and preprocessing module, the real-time monitoring and early warning module, the fault root cause analysis and decision-making module, the spare parts recommendation and execution module, and the model training and optimization module via a high-speed data bus. The local AI server is used to store the standardized dataset and supports millions of data points per second for parallel training of deep learning models.
[0012] Optionally, it also includes: a KPI visualization and analysis module; The KPI visualization and analysis module is used to realize the real-time calculation and chart display of key indicators; wherein the key indicators include at least one of the following: yield rate, failure rate, spare parts replacement accuracy rate, and mean time to repair (MTTR).
[0013] Optionally, the real-time monitoring and early warning module is also used to adjust the yield difference threshold based on the received feedback.
[0014] Optionally, the model training and optimization module is also used to optimize the model feature input and train the full-dimensional yield difference analysis model or the fault-spare part intelligent matching model based on the new input training samples.
[0015] The beneficial effects of the above-mentioned technical solution of the present invention are: In the semiconductor chip test management system based on AI deep learning provided in this embodiment of the invention, the real-time monitoring and early warning module is used to input the first real-time dataset preprocessed by the data acquisition and preprocessing module into the trained full-dimensional yield difference analysis model, obtain the real-time weighted average yield output by the full-dimensional yield difference analysis model, and determine whether the difference between the real-time weighted average yield and the benchmark yield exceeds the yield difference threshold. If it exceeds the threshold, an early warning message is generated. It can be seen that this embodiment of the invention achieves accurate identification of yield differences under multi-dimensional segmentation, and accurately locates OPN / Lot / Machine ID / Site ID / Test Program Version / Test. Inefficient processes involving combinations of dimensions such as code; the fault root cause analysis and decision-making module is used to parse the Fail information and fault scenario information in the fault data when a fault is detected, and input them into the trained fault-spare part intelligent matching model to obtain the fault root cause and corresponding confidence level output by the fault-spare part intelligent matching model, as well as the corresponding spare part. It can be seen that the embodiment of the present invention constructs a standardized fault root cause analysis system, and automatically associates Fail information with spare part replacement records through an AI model to achieve rapid and objective location of the fault root cause; the spare part recommendation and execution module is used to confirm the availability of the corresponding spare part based on the corresponding spare part output by the fault-spare part intelligent matching model, and generate a standardized replacement plan. It can be seen that the embodiment of the present invention can automatically generate a spare part replacement plan when a fault occurs, shorten the maintenance response time, and improve the accuracy of spare part replacement. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the structure of a semiconductor chip test control system based on AI deep learning, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the data acquisition and preprocessing module in an embodiment of the present invention. Detailed Implementation
[0017] To make the technical problems, technical solutions, and advantages of this invention clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments. In the following description, specific details such as particular configurations and components are provided merely to aid in a comprehensive understanding of the embodiments of this invention. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this invention. Furthermore, for clarity and brevity, descriptions of known functions and structures have been omitted.
[0018] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.
[0019] In various embodiments of the present invention, it should be understood that the sequence number of each process described below does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0020] like Figure 1 As shown, this embodiment of the invention provides a semiconductor chip test and control system based on AI deep learning, including a data acquisition and preprocessing module 1, a real-time monitoring and early warning module 2, a fault root cause analysis and decision-making module 3, and a spare parts recommendation and execution module 4. Specifically: The data acquisition and preprocessing module 1 is used to acquire multi-dimensional test data from the test equipment and preprocess the multi-dimensional test data to generate a standardized dataset. The multi-dimensional test data includes at least: product model number OPN, production batch number Lot, machine number, test site number, test program version, test code, test results, test failure information, failure scenario information, and spare parts replacement records.
[0021] In this embodiment of the invention, the data acquisition and preprocessing module 1 may specifically include a data interface submodule 11 and a preprocessing submodule 12. The data interface submodule 11 is used to acquire multi-dimensional test data from the test equipment in real time. This multi-dimensional test data includes OPN, Lot, machine number, Site number, test program version, Test code, test results, Fail information, fault scenario information, and spare parts replacement records. The preprocessing submodule 12 is used to clean, transform, and standardize the multi-dimensional test data acquired by the data interface submodule 11 to generate a standardized dataset.
[0022] In this embodiment of the invention, the data interface submodule 11 can integrate an OPC UA protocol conversion unit and a data verification unit, supporting protocol adaptation with different models of FT / SLT test equipment. In practical applications, the data interface submodule 11 develops an adapter program based on the OPC UA protocol, supporting communication with Advantest V93000 (FT equipment) and Teradyne J750 (SLT equipment). The data acquisition frequency is set to 1 time / second, and the data verification rule is to automatically fill in missing values when the percentage is ≤0.5% and trigger an alarm when it is >0.5%. In this way, the data interface submodule 11 can collect multi-dimensional test data such as OPN, Lot, machine number, Site number, test program version, Test code, test results (pass / fail), Fail information (such as Fail code), fault scenario information, spare parts replacement records, and equipment operating parameters in real time through the OPC UA protocol.
[0023] The preprocessing submodule 12 cleans (removes missing values and outliers), transforms (unifies data format), and standardizes (normalizes numerical data) the multi-dimensional test data collected by the data interface submodule 11 to generate a standardized dataset. Preferably, the generated standardized dataset is stored on a local AI server. In practical applications, the Python Pandas library can be used for data cleaning, the Z-Score method can be used to remove outliers other than 3σ, character data (such as OPN and Test code) can be labeled and encoded, and numerical data (such as test duration and voltage parameters) can be normalized (mapped to the [0,1] interval).
[0024] The real-time monitoring and early warning module 2 is used to input the first real-time dataset preprocessed by the data acquisition and preprocessing module 1 into the trained full-dimensional yield difference analysis model, obtain the real-time weighted average yield output by the full-dimensional yield difference analysis model, and determine whether the difference between the real-time weighted average yield and the baseline yield exceeds the yield difference threshold. If it does, an early warning message is generated. The first real-time dataset includes OPN, Lot, machine number, site number, test program version, test code, and test results. The full-dimensional yield difference analysis model is obtained by hierarchically grouping and training the historical OPN, historical Lot, historical machine number, historical site number, historical test program version, historical test code, and historical test results in the standardized dataset as training samples according to a preset dimension combination.
[0025] This invention utilizes historical OPN, historical Lot, historical machine ID, historical site ID, historical test program version, historical test code, and historical test results from a standardized dataset as training samples to train a full-dimensional yield difference analysis model. This model can calculate a real-time weighted average yield based on the input first real-time dataset, including OPN, Lot, machine ID, site ID, test program version, test code, and test results, according to a preset dimension combination. The real-time monitoring and early warning module 2, based on the real-time weighted average yield output by the full-dimensional yield difference analysis model, determines whether the difference between the real-time weighted average yield and the baseline yield exceeds a yield difference threshold. If it does, an early warning message is generated.
[0026] The yield variation threshold, for example, is 5%, which can be flexibly defined according to actual needs. Warning information can include the dimension of variation, abnormal values, and benchmark reference values.
[0027] In a preferred embodiment of the present invention, the real-time monitoring and early warning module can also be used to adjust the yield difference threshold based on the received feedback. For example, the system adjusts the yield difference threshold based on feedback collected from maintenance personnel (e.g., receiving 120 optimization suggestions) (e.g., some high-precision testing programs adjust this threshold to 3%).
[0028] The fault root cause analysis and decision module 3 is used to parse the Fail information and fault scenario information in the fault data when a fault is detected, and input them into the trained fault-spare part intelligent matching model to obtain the fault root cause and corresponding confidence level, as well as the corresponding spare part output by the fault-spare part intelligent matching model; wherein the fault-spare part intelligent matching model is trained using historical Fail information, historical fault scenario information and historical spare part replacement records in the standardized dataset as training samples and using a Bi-LSTM model.
[0029] In this embodiment of the invention, the fault root cause analysis and decision-making module 3 can identify the occurrence of a fault by receiving an early warning message generated by the real-time monitoring and early warning module 2. The fault data can come from Test history, Testdash, etc.
[0030] This invention utilizes historical Fail information, historical fault scenario information, and historical spare parts replacement records from a standardized dataset, and trains a fault-spare parts intelligent matching model using a Bi-LSTM algorithm. This fault-spare parts intelligent matching model can output the root cause of the fault, the corresponding confidence level, and the corresponding spare parts based on the input Fail information and fault scenario information.
[0031] For example, the fault-spare parts intelligent matching model outputs the top 3 fault root causes and their corresponding confidence levels (≥80%), as well as the corresponding spare parts. Preferably, it can also provide maintenance priority suggestions based on equipment operating parameters.
[0032] The spare parts recommendation and execution module 4 is used to confirm the availability of the corresponding spare parts based on the corresponding spare parts output by the fault-spare parts intelligent matching model 3, and generate a standardized replacement plan.
[0033] In practical applications, the spare parts recommendation and execution module 4 can determine the available quantity and storage location of the corresponding spare parts based on the corresponding spare parts output by the fault-spare parts intelligent matching model 3 by querying the inventory management system. It then generates a standardized replacement plan that includes the spare parts model, quantity, replacement steps, tool list, and safety precautions, which can be pushed to maintenance personnel via the system interface and mobile app. After the maintenance personnel complete the execution, they can report the results back to the system, at which point a valid fault-repair record is stored in the system.
[0034] Furthermore, embodiments of the present invention may also include a model training and optimization module for implementing model training and optimization. Specifically, in embodiments of the present invention, the model training and optimization module is used to train a full-dimensional yield difference analysis model and a fault-spare part intelligent matching model.
[0035] The full-dimensional yield difference analysis model: The model training and optimization module uses historical OPN, historical Lot, historical machine number, historical site number, historical test program version, historical test code, and historical test results from the standardized dataset as training samples. It performs hierarchical grouping according to the combination of OPN-Lot-machine number-site number-test program version-test code to obtain multiple groups. Using the historical average yield of each group as the benchmark, the weighting coefficients of the test data in each dimension are optimized by the gradient boosting tree XGBoost algorithm, and finally the trained full-dimensional yield difference analysis model is obtained.
[0036] For example, select 12 months of historical test data from a standardized dataset, including 5 types of OPNs, 1000 Lots, 20 machines, 80 sites, 15 test procedures, and 50 test codes, as training samples. Group these samples hierarchically according to the dimensions of OPN-Lot-Machine-Site-Test Procedure-Test Code, resulting in 12,000 groups. Use the historical average yield of each group (e.g., 95 points) as the baseline yield. Optimize the weights of each dimension of the test data using the XGBoost algorithm. For example, machine weight 0.3, site weight 0.25, test procedure weight 0.2, test code weight 0.15, and OPN / Lot weight 0.1. Set the training iterations to 1000 and the learning rate to 0.01. Train the full-dimensional yield difference analysis model in this way until training is complete and a well-trained full-dimensional yield difference analysis model is obtained.
[0037] Fault-Spare Part Intelligent Matching Model: The model training and optimization module uses historical Fail information, historical fault scenario information, and historical spare part replacement records from the standardized dataset as training samples. It employs the Bi-LSTM algorithm to establish a mapping relationship between fault features and spare parts, thereby constructing an initial fault-spare part intelligent matching model. The initial fault-spare part intelligent matching model is trained until the accuracy of its output reaches a preset value, at which point training stops, and the trained fault-spare part intelligent matching model is obtained.
[0038] For example, select 24 months of historical fault data from a standardized dataset, including 80 types of Fail codes, 30 common fault scenarios, and 50 types of spare parts replacement records, totaling 15,000 records, and divide them into training and validation sets in a 7:3 ratio; use a Bi-LSTM model, set the hidden layer dimension to 256, the dropout rate to 0.2, the optimizer to Adam, and train for 500 iterations, stopping training when the validation set accuracy reaches 92%.
[0039] After obtaining the trained full-dimensional yield difference analysis model and fault-spare part intelligent matching model, further accuracy verification of the models is required.
[0040] The model training and optimization module can select real-time test data as the test set, such as selecting 3 months of real-time test data as the test set to test the two models.
[0041] In practice, the accuracy of the full-dimensional yield difference analysis model reached 96.2%, and the site-level accuracy reached 91.5%; the recommendation accuracy of the fault-spare part intelligent matching model reached 93%, meeting the preset target.
[0042] More preferably, the model training and optimization module can also be used to automatically trigger iterative training of the model when preset conditions are met; the preset conditions include: a fixed period, and / or, when the amount of newly added data exceeds a preset number of records. The fixed period is, for example, six months or one year, and the amount of newly added data exceeding the preset number of records is, for example, when the amount of newly added data is ≥5000 records. That is, in this embodiment of the invention, iterative training of the full-dimensional yield difference analysis model and the fault-spare part intelligent matching model is automatically triggered every six months or one year, or every time the amount of newly added data is ≥5000 records.
[0043] As a preferred embodiment of the present invention, the model training and optimization module is specifically used to automatically trigger iterative training of the fault-spare part intelligent matching model whenever a preset number (e.g., 5000) of valid fault-repair records are added.
[0044] In actual system application, it can be set that every 100 new valid fault-maintenance records will trigger the model training and optimization module to perform iterative training on the fault-spare part intelligent matching model, so as to ensure the accuracy of the fault-spare part intelligent matching model.
[0045] Furthermore, the model training and optimization module can also be used to optimize model feature inputs, training a full-dimensional yield difference analysis model or a fault-spare part intelligent matching model based on new input training samples. Optimizing model feature inputs includes, for example, adding new equipment operating temperature parameters.
[0046] As a further preferred embodiment of the present invention, the semiconductor chip test management system based on AI deep learning provided in the present invention may further include a KPI visualization and analysis module.
[0047] The KPI visualization and analysis module is used to realize the real-time calculation and chart display of key indicators; wherein the key indicators include at least one of the following: yield rate, failure rate, spare parts replacement accuracy rate, and mean time to repair (MTTR).
[0048] The charts and graphs displayed include bar charts, line charts, and heatmaps.
[0049] In practical applications, the KPI visualization and analysis module 6 monitors indicators such as yield improvement, failure rate, MTTR, and spare parts replacement accuracy in real time. When a certain indicator fails to meet expectations for a consecutive month (such as spare parts recommendation accuracy < 88%), the module analyzes the reasons and notifies the model training and optimization module 5 to start optimizing model parameters.
[0050] Based on the foregoing embodiments, the semiconductor chip test management system based on AI deep learning provided in this embodiment of the invention integrates a local AI server; The local AI server is connected to the data acquisition and preprocessing module 1, the real-time monitoring and early warning module 2, the fault root cause analysis and decision-making module 3, the spare parts recommendation and execution module 4, the model training and optimization module, and the KPI visualization and analysis module via a high-speed data bus.
[0051] The local AI server is used to store the standardized dataset and supports millions of data points per second for parallel training of deep learning models.
[0052] In actual deployment, the local AI server is configured with 8 NVIDIA A100 GPUs, 2TB DDR5 memory, and 10PB distributed storage. It is equipped with the Ubuntu 22.04 operating system and the TensorFlow 2.10 deep learning framework, and features 8 high-performance GPU computing units, 2TB memory, and 10PB distributed storage devices, supporting millions of data processing per second and parallel training of deep learning models.
[0053] The semiconductor chip test and control system based on AI deep learning provided in this invention has at least the following advantages compared to existing technologies: 1. Comprehensive and refined management: Breaking through the limitations of traditional single thresholds, it achieves accurate analysis under multi-dimensional combinations, adapting to the complex scenario requirements of semiconductor chip testing; 2. Deep integration of algorithms and business: The XGBoost and Bi-LSTM algorithms are used to address the technical pain points of yield analysis and fault matching, and the model has strong adaptability. 3. Full-process automated closed loop: Achieve full-process automation from "data collection - analysis and early warning - root cause location - spare parts recommendation - execution feedback", reduce manual intervention and improve management efficiency; 4. Continuous optimization capability: The model supports iterative training based on new data, and the experience base is dynamically updated, which can adapt to changes in scenarios such as production process upgrades and equipment updates; 5. Significantly reduced costs: By reducing unplanned downtime, lowering the rate of incorrect spare parts replacement, and optimizing inventory structure, the annual maintenance cost of a single production line can be reduced by 10%-15%.
[0054] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A semiconductor chip test control system based on AI deep learning, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-dimensional test data from the test equipment and preprocess the multi-dimensional test data to generate a standardized dataset. The multi-dimensional test data includes at least: product model number (OPN), production batch number (Lot), machine number, test site number, test program version, test code, test results, test failure information, failure scenario information, and spare parts replacement records. The real-time monitoring and early warning module is used to input the first real-time dataset preprocessed by the data acquisition and preprocessing module into the trained full-dimensional yield difference analysis model, obtain the real-time weighted average yield output by the full-dimensional yield difference analysis model, and determine whether the difference between the real-time weighted average yield and the baseline yield exceeds the yield difference threshold. If it does, an early warning message is generated. The first real-time dataset includes OPN, Lot, machine number, site number, test program version, test code, and test results. The full-dimensional yield difference analysis model is obtained by hierarchically grouping and training the historical OPN, historical Lot, historical machine number, historical site number, historical test program version, historical test code, and historical test results in the standardized dataset as training samples according to a preset dimension combination. The fault root cause analysis and decision-making module is used to parse the Fail information and fault scenario information in the fault data when a fault is detected, and input them into a trained fault-spare part intelligent matching model to obtain the fault root cause, corresponding confidence level, and corresponding spare part output by the fault-spare part intelligent matching model; wherein the fault-spare part intelligent matching model is trained using historical Fail information, historical fault scenario information, and historical spare part replacement records in the standardized dataset as training samples and using a Bi-LSTM model. The spare parts recommendation and execution module is used to confirm the availability of the corresponding spare parts based on the corresponding spare parts output by the fault-spare parts intelligent matching model, and generate a standardized replacement plan.
2. The semiconductor chip test control system based on AI deep learning according to claim 1, characterized in that, It also includes: a model training and optimization module; The model training and optimization module is used to take the historical OPN, historical Lot, historical machine ID, historical site ID, historical test program version, historical test code, and historical test results in the standardized dataset as training samples, and to stratify and group them according to the dimensions of OPN-Lot-machine ID-Site ID-test program version-Test code to obtain multiple groups. Based on the historical average yield of each group, the weighting coefficients of the test data of each dimension are optimized by the gradient boosting tree (XGBoost) algorithm to finally obtain the trained full-dimensional yield difference analysis model. Furthermore, it is used to use historical Fail information, historical fault scenario information, and historical spare parts replacement records as training samples, and employs the Bi-LSTM algorithm to establish a mapping relationship between fault features and spare parts, thereby constructing an initial fault-spare parts intelligent matching model. The initial fault-spare parts intelligent matching model is trained until the accuracy of the output result of the initial fault-spare parts intelligent matching model reaches a preset value, at which point training stops, and the trained fault-spare parts intelligent matching model is obtained.
3. The semiconductor chip test control system based on AI deep learning according to claim 2, characterized in that, The model training and optimization module is also used to automatically trigger iterative training of the model when preset conditions are met; the preset conditions include: a fixed period, and / or, when the amount of new data exceeds a preset number of records.
4. The semiconductor chip test control system based on AI deep learning according to claim 3, characterized in that, The model training and optimization module is specifically used to automatically trigger iterative training of the fault-spare part intelligent matching model whenever a preset number of valid fault-maintenance records are added.
5. The semiconductor chip test control system based on AI deep learning according to claim 1, characterized in that, The data acquisition and preprocessing module includes: The data interface submodule is used to collect multi-dimensional test data of the test equipment in real time. The multi-dimensional test data includes OPN, Lot, machine number, site number, test program version, test code, test results, Fail information, fault scenario information, and spare parts replacement records. The preprocessing submodule is used to clean, transform, and standardize the multi-dimensional test data collected by the data interface submodule to generate a standardized dataset.
6. The semiconductor chip test control system based on AI deep learning according to claim 5, characterized in that, The data interface submodule integrates an OPC UA protocol conversion unit and a data verification unit, supporting protocol adaptation with different models of FT / SLT test equipment.
7. The semiconductor chip test control system based on AI deep learning according to claim 2, characterized in that, The semiconductor chip test and control system based on AI deep learning integrates a local AI server; The local AI server is connected to the data acquisition and preprocessing module, the real-time monitoring and early warning module, the fault root cause analysis and decision-making module, the spare parts recommendation and execution module, and the model training and optimization module via a high-speed data bus. The local AI server is used to store the standardized dataset and supports millions of data points per second for parallel training of deep learning models.
8. The semiconductor chip test control system based on AI deep learning according to claim 1, characterized in that, Also includes: KPI visualization and analysis module; The KPI visualization and analysis module is used to realize the real-time calculation and chart display of key indicators; wherein the key indicators include at least one of the following: yield rate, failure rate, spare parts replacement accuracy rate, and mean time to repair (MTTR).
9. The semiconductor chip test control system based on AI deep learning according to claim 1, characterized in that, The real-time monitoring and early warning module is also used to adjust the yield difference threshold based on the received feedback.
10. The semiconductor chip test control system based on AI deep learning according to claim 2, characterized in that, The model training and optimization module is also used to optimize the model feature input and train the full-dimensional yield difference analysis model or the fault-spare part intelligent matching model based on the new input training samples.