A silicon wafer defect intelligent detection and process tracing system and method based on multi-modal data fusion

Abstract

The application discloses a kind of silicon polishing piece defect intelligent detection and process tracing system based on multi-modal data fusion, belong to semiconductor material production technical field, including synchronous data acquisition module, AI vision detection model, defect process tracing analysis engine and process self-adaptive feedback unit;Synchronous data acquisition module collects silicon wafer image data and whole-course time sequence process data stream by industrial camera and sensor network respectively, and relies on WaferID and unified timestamp to complete multi-modal data association binding, AI vision detection model is based on silicon wafer image intelligent identification defect and exports determination result, defect process tracing analysis engine compares actual and standard process data stream, and accurately locates abnormal process data by exception detection algorithm.Process self-adaptive feedback unit summarizes the abnormal data of multiple defect silicon wafers, excavates common process deviation and generates process adjustment suggestion parameter set, and pushes to control system to realize full-automatic parameter adjustment or parameter adjustment after manual review.

Description

Technical Field

[0001] This invention relates to the field of semiconductor material manufacturing technology, and in particular to a system and method for intelligent detection and process traceability of defects in silicon polished wafers based on multimodal data fusion. Background Technology

[0002] In the production of silicon polished wafers, surface quality (such as particle contamination, scratches, dimples, and other defects) is a core quality indicator. Currently, mainstream quality inspection methods have the following problems: 1. Delayed and one-sided detection: Traditional manual sampling or stand-alone AOI (Automated Optical Inspection) equipment only performs inspections after polishing / cleaning, which is a post-inspection and cannot intercept batch defects in real time. Moreover, relying solely on optical images, it cannot directly correlate defects with the real-time process parameters that caused the defects (such as polishing pressure, rotation speed, temperature, slurry flow rate, etc.).

[0003] 2. High false positive rate and reliance on experience: For defects with weak contrast to the background, such as dimples and shallow scratches, defect judgment relies heavily on the engineer's experience. Standards are difficult to unify and quantify, resulting in high false positive and false negative rates.

[0004] 3. Difficulty in root cause analysis of defects: When defects are discovered, process engineers need to manually backtrack massive amounts of discrete process log data, which is time-consuming and labor-intensive. It is difficult to quickly and accurately locate the abnormal process range that caused the defect (i.e., "process window drift"), which leads to repeated occurrences of the problem and slow yield improvement.

[0005] 4. Insufficient system flexibility: The existing detection model is fixed, and its adaptability is poor when product specifications (such as different crystal orientations and thicknesses) or process formulations change, resulting in a long deployment cycle.

[0006] In summary, existing quality inspection solutions are no longer sufficient to meet the mass production control requirements of high-precision silicon polished wafers. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a system and method for intelligent detection and process traceability of defects in silicon polished wafers based on multimodal data fusion.

[0008] This invention provides the following technical solution: A smart defect detection and process traceability system for silicon polished wafers based on multimodal data fusion includes: Synchronous data acquisition module: Includes industrial camera and sensor network. The industrial camera triggers the acquisition of high-definition images of silicon wafers at preset workstations, and the sensor network acquires the actual process data stream of the entire silicon wafer processing in sequence. The synchronous data acquisition module binds the image data of the same silicon wafer with the actual process data stream. AI visual inspection model: Identifies silicon wafer defects based on image data of silicon wafers and outputs defect judgment results; Defect process traceability analysis engine: Pre-stores standard process data streams for silicon wafers. When the defect determination result indicates the presence of a defect, it retrieves the actual process data stream corresponding to the silicon wafer through WaferID, compares it with the standard process data stream, and identifies and locates abnormal process data through anomaly detection algorithms. Process Adaptive Feedback Unit: It summarizes abnormal process data corresponding to multiple defective silicon wafers, identifies common process deviations, and generates a set of process adjustment suggestion parameters to be pushed to the control system; the control system automatically adjusts the process parameters, or completes the process parameter adjustment after manual confirmation, forming a closed-loop control of perception-diagnosis-optimization.

[0009] A further solution also includes an edge computing server, in which the AI ​​visual inspection model, defect process tracing analysis engine, and process adaptive feedback unit are all built-in and deployed.

[0010] In a further embodiment, the edge computing server is configured with a local storage unit for storing at least one of the following: standard process data stream for silicon wafers, bound images and actual process data streams, defect determination results, abnormal process data, and process adjustment suggestion parameters; and / or, the edge computing server uploads the above data to the control system for archiving and storage.

[0011] In a further embodiment, the abnormal process data includes the process parameters that deviated, the time point of deviation, and the amount of deviation.

[0012] In a further embodiment, the defect process tracing and analysis engine is equipped with a graded alarm module. After identifying and locating abnormal process data, it classifies the abnormality level according to the deviation amount and outputs the corresponding alarm signal.

[0013] A further solution also includes a human-machine interface terminal that communicates with the defect process tracing and analysis engine; after the anomaly detection algorithm identifies and locates abnormal process data, it generates an abnormal process root cause chart and pushes it to the human-machine interface terminal for display. The human-machine interface terminal is also used to receive manual confirmation instructions and send them to the control system to complete the adjustment of process parameters.

[0014] In a further embodiment, the industrial camera includes at least one of a high-speed global shutter area scan camera, a TDI line scan camera, a short-wave infrared camera, and a 3D contour camera.

[0015] In a further embodiment, the sensor network includes at least one of a pressure sensor, a temperature sensor, a flow sensor, and a speed sensor.

[0016] Secondly, the present invention provides a method for intelligent detection and process tracing of defects in silicon polished wafers based on multimodal data fusion, comprising the following steps: S1. Synchronous Data Acquisition and Binding: Continuous time-series data acquisition is performed throughout the entire silicon wafer polishing process to obtain the complete time-series actual process data stream; when the silicon wafer is transferred to the preset station, its high-definition image is acquired, and the image data of the same silicon wafer is associated with the unified timestamp and the image data of the entire process are associated and bound together. S2. Image Defect Recognition: Perform feature extraction and defect discrimination analysis on the image data of silicon wafers, and output the defect judgment results; S3. Process Anomaly Tracing and Comparison: Pre-store standard process data stream for silicon wafers; if the defect determination result of the silicon wafer is that there is a defect, the actual process data stream of the silicon wafer is retrieved through the corresponding WaferID and compared with the standard process data stream, and the abnormal process data is identified and located through the anomaly detection algorithm. If the anomaly detection algorithm fails to identify abnormal process data, the silicon wafer is marked and its image data and the entire actual process data stream are stored in the case library for manual analysis. S4. Common Deviation Closed-Loop Optimization: Summarize abnormal process data corresponding to multiple defective silicon wafers, quantify the influence weight of each process parameter on silicon wafer defects through process contribution analysis, screen key abnormal parameters and identify common process deviations based on parameter influence weights, generate a set of process adjustment suggestion parameters based on common process deviations, and push them to the control system; the control system automatically adjusts process parameters, or completes process parameter adjustment after manual confirmation, forming a perception-diagnosis-optimization closed-loop control.

[0017] In a further embodiment, in step S3, after identifying and locating abnormal process data through an anomaly detection algorithm, a visualized root cause chart of the abnormal process is output.

[0018] The beneficial effects of this invention are: This invention deeply couples AI visual inspection with real-time process parameters of the production line to build a closed-loop system of "perception-diagnosis-optimization". It not only realizes defect detection, but also traces the cause of defects and adaptively adjusts process parameters, transforming "post-event inspection" into "in-process control and prevention".

[0019] 1. Achieve precise binding of multimodal data and high reliability of data traceability. This invention adopts a method of continuous acquisition of full-process time-series process data and fixed-point acquisition of workstation images. By combining WaferID with a unified timestamp, it completes the accurate association between single silicon wafer images and full-process process data, which solves the problem of the traditional solution where images and process data are separated and cannot be accurately matched, and provides accurate data support for process traceability.

[0020] 2. Achieve precise location of the root cause of defects in the process, overcoming the limitations of simple surface detection. Unlike traditional visual inspection, which can only identify surface defects, this invention can accurately locate abnormal process parameters, deviation time, and deviation amount by matching defective silicon wafers with time-series process data and comparing them with standard processes. This upgrades the process from "defect detection" to "process root cause tracing," reducing the burden on process engineers.

[0021] 3. Supports batch common deviation analysis to achieve adaptive optimization of process closed loop. This invention can aggregate abnormal data from multiple defective silicon wafers, uncover common process deviations in production lines, automatically generate process adjustment parameter suggestions, and complete process optimization through automatic or manual confirmation, forming a closed-loop control of detection-traceability-parameter adjustment, effectively improving production line process stability and silicon wafer yield.

[0022] 4. Tiered early warning combined with visualization improves operation and maintenance efficiency. This invention can generate graded alarms based on process deviations, and at the same time generate a visual root cause chart to intuitively display the process anomaly trend and time period, which facilitates quick manual verification and calibration, reduces the difficulty of operation and maintenance, and improves the efficiency of fault handling.

[0023] 5. Dual storage architecture and edge deployment ensure superior system real-time performance and security. The core functions are deployed on edge servers, resulting in low response latency and strong real-time performance. Combined with the dual storage mode of edge local storage and cloud archiving of the control system, it not only meets the needs of high-speed on-site retrieval but also achieves data backup and retention, ensuring data integrity and traceability.

[0024] 6. Wide adaptability and strong engineering implementation. This invention is compatible with a variety of image acquisition devices and process sensors, adapts to various production scenarios for silicon wafer polishing, has a modular architecture with strong versatility, can be quickly deployed to various production lines, and effectively reduces the cost of quality inspection and process optimization.

[0025] 7. The model has strong generalization ability, and new products and processes do not require extensive data retraining. Unlike traditional AI quality inspection methods that rely on large-scale labeled data and require retraining for scenario migration, this invention achieves traceability and optimization by comparing time-series process data and statistically analyzing common deviations, rather than relying on model-specific training features. When product specifications or polishing process formulas change, only the corresponding standard process data stream needs to be updated to complete the adaptation, without the need for extensive sample labeling and model iteration training. This significantly reduces the cost of production line iteration and upgrades, and offers high flexibility in scenario adaptation.

[0026] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] Figure 1 This is a system framework diagram of an embodiment of the present invention.

[0028] Figure 2 This is a flowchart of a method according to an embodiment of the present invention.

[0029] Figure 3 In this embodiment of the invention, the defective silicon wafer image data is identified as follows: (a) is a central depression in the silicon wafer, and (b) is that the edge is locally too thin.

[0030] Figure 4 In this embodiment of the invention, multiple graphics are stacked to visualize particle density distribution and defect annotation, and the AI ​​visual inspection model automatically identifies the distribution patterns of defects such as particle aggregation, scratches, and COP.

[0031] Figure labels: 1. Production line equipment; 2. Synchronous data acquisition module; 3. Edge computing server; 3. Multimodal fusion analysis model; 3.1. Human-computer interaction terminal; 4. Control system; 5. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0033] like Figure 1 As shown, this embodiment provides an intelligent defect detection and process traceability system for silicon polished wafers based on multimodal data fusion, including a synchronous data acquisition module, an edge computing server, and a human-computer interaction terminal. The edge computing server has a built-in multimodal fusion analysis model and a local storage unit. The multimodal fusion analysis model includes an AI visual detection model, a defect process traceability analysis engine, and a process adaptive feedback unit.

[0034] The synchronous data acquisition module includes an industrial camera and a sensor network. The industrial camera is used for high-definition image acquisition of silicon wafers. In some embodiments, the industrial camera is equipped with both a 3D contour camera and a high-speed global shutter area array camera (or can be replaced by any one or more combinations of a TDI line scan camera and a short-wave infrared camera). The industrial camera acquires high-definition images of silicon wafers after being triggered at a preset station (e.g., a wafer cassette exchange station). A high-speed global shutter area array camera acquires two-dimensional optical images of the silicon wafer surface, identifying planar defects such as scratches, particles, and dirt. A 3D contour camera simultaneously acquires three-dimensional thickness contour data of the silicon wafer, outputting a thickness mapping heatmap to characterize central depressions (such as…). Figure 3 (a) shows uneven thickness at the edges (as shown in the image). Figure 3(b) shows thickness uniformity defects; short-wave infrared cameras can assist in detecting hidden layer defects inside silicon wafers, and TDI line scan cameras are adapted to continuous conveying scenarios on production lines to achieve high-speed uninterrupted imaging. Figure 3 The color chart represents the thickness distribution of the silicon wafer, with red → blue representing thick → thin, and the unit is micrometers (μm).

[0035] The sensor network collects time-series process data streams throughout the silicon wafer processing process. The sensor network includes various sensors on the production line equipment. In some embodiments, it includes at least one of pressure sensors, temperature sensors, flow sensors, and speed sensors. The pressure sensor collects real-time pressure data from targets such as the polishing head and the support suction cup; the temperature sensor monitors the temperature of the polishing slurry, polishing table, and silicon wafer; the flow sensor collects the flow rate of targets such as the polishing slurry and deionized water spray; and the speed sensor acquires the rotational speed of targets such as the polishing head and spindle. All sensors synchronously output continuous time-series parameters, forming an actual process data stream for process traceability and comparison. Image data of the same silicon wafer captured by industrial cameras and the actual process data stream of the entire silicon wafer processing process are bound together through a unified timestamp and WaferID to complete multimodal data fusion, supporting subsequent defect identification and process traceability; The AI ​​visual detection model: such as Figure 4 As shown, the AI ​​visual inspection model identifies silicon wafer defects based on image data of the silicon wafer and outputs the defect judgment result. Figure 4 (The right side shows the judgment result). The particle analyzer (SP1) can scan the surface of each silicon wafer and generate a corresponding particle distribution map. The judgment method is as follows: obtain the particle distribution map corresponding to the silicon wafer cut from the same single crystal silicon rod, superimpose multiple maps from a top-down perspective, and determine whether there are characteristic patterns corresponding to single crystal defects; if characteristic patterns are detected, they are further compared with the standard single crystal defect map to determine the specific type of single crystal defect.

[0036] The aforementioned AI visual inspection model adopts an existing mature semiconductor defect visual inspection model. This embodiment only utilizes this model to complete the silicon wafer defect identification and defect judgment result output, without involving any improvement to the structure, algorithm, or training method of the AI ​​model itself. For example, the existing models that can be selected include: YOLOv8m, YOLOv10m, YOLOv11m; BDSD-Net: a real-time anomaly detector designed specifically for semiconductor wafer visual inspection, which adopts a lightweight MVHNet backbone network and combines the advantages of CNN and Transformer; SEMI-CenterNet is a CenterNet variant customized for semiconductor defect detection, which achieves efficient defect localization and classification on SEM images.

[0037] The defect process tracing and analysis engine: The engine's database pre-stores standard process data streams corresponding to silicon polished wafers of various specifications. The standard process data streams include standard time-series curves and allowable fluctuation threshold ranges for various process parameters such as pressure, temperature, flow rate, and rotation speed throughout the entire processing flow. When the AI ​​visual inspection model outputs a defective result for the silicon wafer, the defect process traceability analysis engine retrieves the entire actual process data stream bound to the unique WaferID of the silicon wafer from the database. It then compares the actual process data stream with the pre-stored standard process data stream of the same specification time-stamp by time. The engine has a built-in anomaly detection algorithm (using existing mature algorithms) that automatically identifies abnormal conditions such as parameter deviation, sudden drift, range exceeding limits, and time offset. It also accurately locates the process time point, sensor location, and abnormal parameter type corresponding to the abnormal data and outputs a visual anomaly root cause chart to the human-machine interaction terminal.

[0038] If the anomaly detection algorithm fails to identify any abnormal process data exceeding the standard fluctuation range after traversing all process parameters, it determines that the defect cannot be automatically traced through conventional process parameters. The engine automatically adds a mark to the silicon wafer, packages the high-definition image of the silicon wafer and the entire actual process data stream into the defect case library, and pushes it to the human-computer interaction terminal or a dedicated manual analysis workstation. Process engineers then conduct offline review and manual tracing. The review conclusions can be reverse-entered into the case library to expand the sample. The abnormal process data traced manually can be processed together with the data traced automatically by the process adaptive feedback unit.

[0039] The process adaptive feedback unit continuously summarizes all abnormal process data corresponding to multiple defective silicon wafers output by the defect process tracing analysis engine. In some embodiments, the abnormal process data covers all-dimensional time-series abnormal parameters such as polishing pressure, stage temperature, polishing fluid flow rate, and spindle speed. The unit has a built-in process contribution analysis model to quantify the influence weight of each type of process parameter on various defects of silicon wafers, sorts and filters key abnormal parameters based on weight, and explores common process deviations that recur in multiple batches of defect samples. The process adaptive feedback unit automatically generates a standardized set of process adjustment suggestion parameters based on the identified common process deviations and pushes it to the control system of the production line polishing equipment in real time. In some embodiments, the set of process adjustment suggestion parameters includes the correction range, adjustment interval, and step-by-step adjustment strategy for each parameter. The control system provides at least one of the following parameter activation modes: the first is a fully automatic mode, in which the control system directly loads the set of process adjustment suggestion parameters and updates the operating process parameters of the processing equipment in real time; the second is a manual review mode, in which the control system pushes the set of process adjustment suggestion parameters to the human-machine interface terminal, and the process personnel review and confirm it through the human-machine interface terminal before issuing the execution parameter modification to the control system.

[0040] By linking image and process parameter acquisition, visual defect detection, source analysis engine diagnosis of process anomalies, and feedback unit optimization of process parameters, a fully automated closed-loop control system of perception-diagnosis-optimization is constructed to continuously reduce the probability of recurrence of similar defects.

[0041] In this embodiment, the defect process tracing and analysis engine has a built-in hierarchical alarm module for quantitative evaluation and hierarchical alerts of identified and located process anomaly data. In some embodiments, after the anomaly detection algorithm completes the identification of abnormal process data, the location of abnormal time periods and abnormal parameters, the hierarchical alarm module quantifies and classifies the process anomalies based on multi-dimensional indicators such as the deviation of actual process parameters from standard process thresholds, duration, and degree of mutation, dividing the anomalies into different severity levels. The process contribution analysis algorithm in this embodiment can use existing mature algorithms, such as using Minitab to conduct main effect and interaction effect analysis to determine the contribution degree of each process parameter.

[0042] Among them, minor anomalies that deviate slightly from the standard range, fluctuate briefly, and do not affect the polishing quality of the silicon wafer are judged as low-level anomalies and a regular warning alarm is output; operating condition fluctuations that have large parameter deviations, long durations, and are likely to induce minor batch defects are judged as medium-level anomalies and a warning alarm is output and process attention information is pushed; process anomalies that deviate severely from the process threshold, have drastic changes, directly cause structural defects in the silicon wafer, and are very likely to cause batch defects are judged as high-level anomalies and a high-order alarm signal is output.

[0043] The graded alarm module outputs alarm signals and prompts corresponding to the matched anomaly level, realizing differentiated early warning reminders for process anomalies. This allows on-site personnel to take targeted intervention measures based on the severity of the anomalies, prioritizing the handling of high-risk process deviations. At the same time, it provides quantitative basis for subsequent analysis of common process deviations and parameter optimization, further improving the precision of process diagnosis and on-site adaptability of the entire system.

[0044] The local storage unit is used to store at least one of the following: standard process data stream for silicon wafers, high-definition images of silicon wafers bound to WaferID and a unified timestamp, actual process data stream throughout the entire process, defect judgment results output by the AI ​​visual inspection model, abnormal process data located by the defect process tracing analysis engine, a case library containing difficult defect samples, and process adjustment suggestion parameters generated by the process adaptive feedback unit. The local storage unit locally carries the large-capacity multimodal data required for real-time defect identification and process tracing, ensuring low-latency retrieval for online analysis.

[0045] Meanwhile, the edge computing server has data upload and forwarding functions, which can upload the various types of detection data and process data stored in the local storage unit to the production line control system (MES system or equipment automation system) in batches or on demand through the industrial communication link. The control system completes the long-term archiving and preservation of all data. This forms a two-level storage architecture that combines high-speed local storage at the edge with centralized archiving by the control system. This not only meets the speed requirements of real-time computing at the workstation, but also realizes the unified collection and traceable management of silicon wafer production data throughout its entire life cycle.

[0046] The defect process tracing and analysis engine is equipped with a graded alarm module. After identifying and locating abnormal process data, it classifies the abnormality level according to the deviation amount and outputs the corresponding alarm signal.

[0047] This embodiment also provides a method for intelligent detection and process tracing of defects in silicon polished wafers based on multimodal data fusion. This method is executed by the aforementioned intelligent detection and process tracing system for silicon polished wafers based on multimodal data fusion. It relies on the coordinated operation of industrial camera multimodal image acquisition, multi-type sensor condition monitoring, a defect process tracing analysis engine, a hierarchical alarm module, a process adaptive feedback unit, and an edge storage archiving architecture to achieve accurate detection of defects in silicon polished wafers, process anomaly tracing, hierarchical early warning, and process closed-loop optimization. The method includes the following steps: S1. Synchronous Data Acquisition and Binding: A multi-type sensor network deployed on the polishing production line continuously acquires time-series data throughout the entire silicon wafer polishing process. This includes pressure data from targets such as the polishing head and chuck; temperature data from targets such as the polishing slurry, polishing table, and silicon wafer; flow rate data from targets such as the polishing slurry and deionized water spray; and rotational speed data from targets such as the polishing disc and polishing spindle. These core operating parameters are integrated to form a complete time-series actual process data stream. Simultaneously, when the silicon wafer moves to a preset image acquisition station, high-definition image data of the silicon wafer is acquired using at least one industrial camera selected from high-speed global shutter area array cameras, TDI line scan cameras, short-wave infrared cameras, and 3D contour cameras. This method uses a unique WaferID as the association subject, combined with a unified high-precision timestamp, to accurately associate and bind the image data of the same silicon wafer with its entire actual process data stream. This generates a silicon wafer-specific multimodal dataset, which is stored in the local storage unit built into the edge computing server, providing a complete data foundation for subsequent defect identification and process traceability.

[0048] S2. Image Defect Recognition: Retrieves the high-definition image data of the silicon wafer that has been bound in the local storage unit. Through the preset defect recognition model, the image is subjected to deep feature extraction, morphology analysis and defect discrimination analysis to accurately identify various macroscopic and microscopic defects such as scratches, dirt, particles, uneven edge thickness, and central depression on the silicon wafer surface. Finally, it outputs accurate silicon wafer defect judgment results, which clearly indicate whether there are defects in the current silicon wafer and the corresponding defect type, defect location and defect degree.

[0049] S3. Process Anomaly Tracing and Comparison: The edge computing server pre-stores standard process data streams corresponding to silicon wafers of different specifications locally. Generally, the standard process data stream contains standard time-series curves and allowable fluctuation threshold ranges for each process parameter, serving as the benchmark for process anomaly judgment. If the defect judgment result of the silicon wafer indicates the presence of a defect, the defect process tracing and analysis engine accurately retrieves the entire actual process data stream bound to the defective silicon wafer through the corresponding WaferID, and accurately compares the actual process time-series curve with the standard process data stream of the same specification time-stamp and parameter dimension. Through the built-in time-series anomaly detection algorithm, it intelligently identifies abnormal operating conditions such as parameter mutations, range exceeding limits, and long-term deviations, and accurately locates abnormal process parameters, the time period of the anomaly, and the corresponding sensor monitoring points. After completing the identification and location of abnormal process data, the system automatically generates and outputs a visualized abnormal process root cause chart, which is sent to the human-computer interaction terminal to intuitively display the process deviation trend, deviation magnitude, and the correlation between the defect. If the anomaly detection algorithm fails to identify any abnormal process data exceeding the standard fluctuation range after traversing all process parameters, the system marks the defective silicon wafer, packages its image data and the entire actual process data stream, and stores it in the defect case library. This data is then transferred to process engineers for manual review and analysis, continuously enriching the database of difficult defect samples. Simultaneously, the built-in graded alarm module in this step classifies the identified process anomalies according to the magnitude, duration, and severity of deviation of process parameters from the standard threshold, outputting corresponding alarm signals and warning prompts to achieve differentiated graded early warning.

[0050] S4. Common Deviation Closed-Loop Optimization: The process adaptive feedback unit continuously summarizes all abnormal process data corresponding to multiple defective silicon wafers, integrates batch defective process sample datasets, and quantifies the influence weight of various process parameters such as pressure, temperature, flow rate, and rotation speed on different types of silicon wafer defects through the process contribution analysis algorithm. It accurately screens out key abnormal parameters with high correlation and batch mines common process deviation problems that recur in multiple batches of production. Based on the identified common process deviations and parameter influence weights, an adaptive parameter set of process adjustment suggestions is generated to match the operating conditions of the production line equipment. This set includes core information such as parameter correction range, adjustment range, and step-by-step adjustment strategy, and is pushed to the production line control system in real time. The control system supports a dual-mode parameter adjustment mechanism. It can directly load the parameter set in fully automatic mode and automatically update the polishing equipment's operating process parameters. Alternatively, it can use a manual review mode, where process adjustment suggestions are pushed to the human-machine interface terminal. After manual review and confirmation by process personnel, the process parameters are issued and adjusted. This ultimately forms a closed-loop control system of "image perception - defect diagnosis - process optimization" that continuously corrects process deviations and reduces the probability of recurrence of similar defects.

[0051] This invention can achieve adaptive iterative optimization of process parameters, forming a sustainable and self-updating production closed-loop control system, and significantly reducing the recurrence rate of similar defects.

[0052] The system and method provided in this embodiment can improve the silicon wafer defect identification rate by more than 20% and reduce the manual judgment time by more than 80% in actual production, effectively preventing particulate defects from leaving the production line.

[0053] The above description is only a specific embodiment of the present invention, but the structural features of the present invention are not limited thereto. Any changes or modifications made by those skilled in the art within the scope of the present invention are covered by the patent scope of the present invention.

Claims

1. A smart defect detection and process traceability system for silicon polished wafers based on multimodal data fusion, characterized in that, include: Synchronous data acquisition module: includes an industrial camera and a sensor network. The industrial camera triggers the acquisition of high-definition images of silicon wafers at a preset workstation, and the sensor network acquires the actual process data stream of the entire silicon wafer processing process in sequence. The synchronous data acquisition module binds image data and actual process data streams of the same silicon wafer using a unified timestamp and WaferID; AI visual inspection model: Identifies silicon wafer defects based on image data of silicon wafers and outputs defect judgment results; Defect process traceability analysis engine: Pre-stores standard process data streams for silicon wafers. When the defect determination result indicates the presence of a defect, it retrieves the actual process data stream corresponding to the silicon wafer through WaferID, compares it with the standard process data stream, and identifies and locates abnormal process data through anomaly detection algorithms. Process Adaptive Feedback Unit: It summarizes abnormal process data corresponding to multiple defective silicon wafers, identifies common process deviations, and generates a set of process adjustment suggestion parameters to be pushed to the control system; the control system automatically adjusts the process parameters, or completes the process parameter adjustment after manual confirmation, forming a closed-loop control of perception-diagnosis-optimization.

2. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 1, characterized in that, It also includes an edge computing server, in which the AI ​​visual inspection model, defect process tracing analysis engine, and process adaptive feedback unit are all built-in and deployed.

3. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 2, characterized in that, The edge computing server is configured with a local storage unit for storing at least one of the following: standard process data stream for silicon wafers, bound images and actual process data streams, defect judgment results, abnormal process data, and process adjustment suggestion parameters; and / or, the edge computing server uploads the above data to the control system for archiving and storage.

4. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 1, characterized in that, The abnormal process data includes the process parameters that deviated, the time point of deviation, and the amount of deviation.

5. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 4, characterized in that, The defect process tracing and analysis engine is equipped with a graded alarm module. After identifying and locating abnormal process data, it classifies the abnormality level according to the deviation amount and outputs the corresponding alarm signal.

6. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 1, characterized in that, It also includes a human-machine interface terminal that communicates with the defect process tracing and analysis engine; after the anomaly detection algorithm identifies and locates abnormal process data, it generates an abnormal process root cause chart and pushes it to the human-machine interface terminal for display. The human-machine interface terminal is also used to receive manual confirmation instructions and send them to the control system to complete the adjustment of process parameters.

7. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 1, characterized in that, The industrial camera includes at least one of a high-speed global shutter area scan camera, a TDI line scan camera, a short-wave infrared camera, and a 3D contour camera.

8. The intelligent detection and process traceability system for silicon polished wafer defects based on multimodal data fusion according to claim 1, characterized in that, The sensor network includes at least one of a pressure sensor, a temperature sensor, a flow sensor, and a speed sensor.

9. A method for intelligent defect detection and process traceability of silicon polished wafers based on multimodal data fusion, characterized in that, Includes the following steps: S1. Synchronous Data Acquisition and Binding: Continuous time-series data acquisition is performed throughout the entire silicon wafer polishing process to obtain the complete time-series actual process data stream; when the silicon wafer is transferred to the preset station, its high-definition image is acquired, and the image data of the same silicon wafer is associated with the unified timestamp and the image data of the entire process are associated and bound together. S2. Image Defect Recognition: Perform feature extraction and defect discrimination analysis on the image data of silicon wafers, and output the defect judgment results; S3. Process Anomaly Tracing and Comparison: Pre-store standard process data stream for silicon wafers; if the defect determination result of the silicon wafer is that there is a defect, the actual process data stream of the silicon wafer is retrieved through the corresponding WaferID and compared with the standard process data stream, and the abnormal process data is identified and located through the anomaly detection algorithm. If the anomaly detection algorithm fails to identify abnormal process data, the silicon wafer is marked and its image data and the entire actual process data stream are stored in the case library for manual analysis. S4. Common Deviation Closed-Loop Optimization: Summarize abnormal process data corresponding to multiple defective silicon wafers, quantify the influence weight of each process parameter on silicon wafer defects through process contribution analysis, screen key abnormal parameters and identify common process deviations based on parameter influence weights, generate a set of process adjustment suggestion parameters based on common process deviations, and push them to the control system; the control system automatically adjusts process parameters, or completes process parameter adjustment after manual confirmation, forming a perception-diagnosis-optimization closed-loop control.

10. The intelligent detection and process traceability method for defects in silicon polished wafers based on multimodal data fusion according to claim 9, characterized in that, In step S3, after identifying and locating abnormal process data through an anomaly detection algorithm, a visualized root cause chart of the abnormal process is output.

Images