Wafer isolation paper shape position detection method and system

Abstract

The present application relates to the technical field of semiconductor packaging equipment and intelligent manufacturing, and particularly relates to a wafer isolation paper shape position detection method and system. The method comprises: identifying whether the top layer material of a wafer carrying container is isolation paper through a reflection sensor; synchronously collecting distance data and environmental parameters; dynamically generating a composite feature set representing the shape position state of the isolation paper; performing double judgment on the composite feature set; and outputting the final judgment and control instructions about whether the shape position of the isolation paper is abnormal. The present application combines dynamic environment self-adaptation with the pattern recognition ability of machine learning, constructs a progressive detection architecture of "coarse screening-fine judgment", significantly improves the detection accuracy and the environmental adaptability of the system, and effectively solves the problem of high misjudgment rate of the traditional fixed threshold detection method under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of semiconductor packaging equipment and intelligent manufacturing technology, and in particular to a real-time, high-precision, adaptive detection method and system for centering and flatness of release paper during wafer packaging. Background Technology

[0002] In the back-end processes of modern semiconductor manufacturing, to protect high-value wafers from scratches, friction, and electrostatic damage during storage and transportation, a layer of high-cleanliness interleaf paper (or spacer paper) is typically placed between each wafer. On automated packaging production lines, robotic arms precisely alternately place the interleaf paper and wafers into carrier containers. The core challenge in this process is that the thin and flexible interleaf paper is highly susceptible to defects such as edge warping, overall arching, unilateral folding, or center misalignment due to static electricity, air pressure, angular deviations, or minor unevenness of the carrier itself during absorption and placement. If interleaf paper with such defects is used directly to carry wafers, it can easily lead to uneven stress on the wafers during placement, causing stress concentration, which can result in microcracks at the edges or even wafer breakage and scrapping, causing significant economic losses.

[0003] Currently, most automated detection solutions for this problem in the industry use sensor arrays combined with preset fixed thresholds for judgment. The typical technical approach is as follows: First, reflective or through-beam sensors are used to distinguish whether the top layer is a wafer or a release liner; if it is identified as release liner, distance sensors (such as laser or ultrasonic sensors) arranged circumferentially measure the distance from multiple detection points on the edge of the release liner to the sensor reference plane; then, the measured values ​​are compared with preset acceptable distance ranges or the differences between the sensors. If the difference exceeds the preset fixed threshold, the shape and position of the release liner is determined to be abnormal and an alarm is triggered.

[0004] However, this fixed threshold-based detection method has many insurmountable limitations in industrial production practice: (1) Poor robustness and susceptible to process fluctuations: Small changes in the suction and release force and motion trajectory of the robotic arm, as well as the accuracy drift caused by long-term operation of the equipment, will cause the placement position and posture of qualified release paper to fluctuate normally within a small range. The fixed threshold cannot adapt to the dynamic changes within this "reasonable" range, and often generates false alarms (judging normal as abnormal), resulting in unnecessary shutdown of the production line for verification and reducing the overall equipment efficiency (OEE). (2) Unable to effectively compensate for the influence of environmental factors: Changes in temperature and humidity in the workshop will cause systematic drift in the calibration zero point and response sensitivity of the sensor (especially the optical distance sensor). The fixed threshold cannot be dynamically adjusted according to environmental changes, causing the detection standard to deviate in different seasons or even at different times of the same day, resulting in insufficient stability. (3) Weak ability to identify complex morphological defects: The fixed threshold is usually based on simple numerical comparisons (such as the difference between the maximum and minimum values ​​or the deviation from the benchmark value), which makes it difficult to capture slight, complex or local morphological defects. For example, a single sheet of release paper may have slight warping on one side, slight wrinkles in some areas, or slight asymmetrical collapse. These defects may not cause the maximum or minimum value of the four measuring points to exceed the standard significantly, but they are enough to threaten wafer safety. Traditional methods are very likely to miss such defects. (4) Lack of adaptive capability and difficulty in dealing with material batch differences: Release papers from different suppliers or different batches have slight tolerances in their inherent properties such as thickness, stiffness, and surface reflectivity. These differences will affect the reference reading of the sensor, making the fixed threshold originally calibrated for the batch of materials no longer applicable to the new batch of materials, thus causing batch misjudgment, requiring manual shutdown for recalibration, increasing maintenance costs and reducing production continuity.

[0005] Therefore, developing an intelligent detection method and system that can overcome the above-mentioned defects and possess high precision, strong adaptability, and strong robustness has become an urgent need to improve the automation level of semiconductor packaging sections and ensure product quality and production line stability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for detecting the shape and position of wafer isolation paper.

[0007] In a first aspect, the present invention provides a method for detecting the shape and position of wafer release liner, comprising: at a wafer packaging station, identifying whether the top layer material of the wafer carrier container is release liner through a reflective sensor; if it is identified as release liner, driving multiple sets of distance sensors arranged directly above the edge of the wafer carrier container to synchronously collect distance data from each point on the edge of the release liner to the distance sensors, and acquiring environmental parameters in real time; dynamically generating a composite feature set characterizing the shape and position state of the release liner based on the collected distance data and environmental parameters; performing a dual judgment on the composite feature set, the dual judgment comprising: performing a preliminary screening of the distance data using a dynamic threshold dynamically generated based on historical data; inputting the composite feature set corresponding to the data that passed the preliminary screening into a pre-trained machine learning classification model for morphological classification judgment; and outputting a final judgment on whether the shape and position of the release liner is abnormal based on the result of the dual judgment, and outputting control instructions to control the subsequent wafer placement process.

[0008] Optionally, the environmental parameters include at least real-time ambient temperature and real-time ambient humidity; the dynamically generated dynamic threshold is calculated by a sliding window based on maintained historical distance data, and environmental drift is compensated by combining environmental parameters during the calculation.

[0009] Optionally, the preliminary screening of distance data using a dynamic threshold dynamically generated based on historical data includes: calculating the range, variance, or standard deviation of the distance data; comparing the range, variance, or standard deviation of the distance data with the dynamic threshold calculated using a sliding window of historical distance data; if the range exceeds the threshold range, it is determined to be morphologically abnormal and subsequent judgment is terminated; if the range does not exceed the threshold range, the composite feature set corresponding to the preliminary screened data is input into a pre-trained machine learning classification model for further morphological classification judgment.

[0010] Optionally, the composite feature set includes at least one or more of the following features: statistical features extracted from the distance data, including the mean, the difference between the maximum and minimum values, the variance or standard deviation; the dominant frequency component and / or the proportion of high-frequency energy extracted after performing time-frequency domain transformation on the distance data; environmental compensation factors related to environmental parameters; skewness values ​​reflecting the asymmetry of the distance data distribution and / or kurtosis values ​​reflecting the sharpness.

[0011] Optionally, the pre-trained machine learning classification model uses a support vector machine as a classifier and combines it with the isolated forest algorithm to perform secondary anomaly detection on results judged as normal by the classifier.

[0012] Optionally, it also includes an online learning and updating mechanism for the model: during the detection process, new sample data with real result labels are continuously collected; when the accumulated new sample data reaches a preset number, an incremental learning process is triggered to update the parameters of the pre-trained machine learning classification model to adapt to process fluctuations and material batch changes.

[0013] Optionally, if the final judgment on the shape and position of the isolation paper is abnormal as output by the result of the dual judgment, the abnormality type information is also output, which includes at least one of overall warping, single-sided warping, folding, and center offset.

[0014] Optionally, the multiple sets of ranging sensors are uniformly and symmetrically arranged in a rectangular or equilateral polygonal manner directly above the edge of the wafer carrier container, and the dimension of the distance data is consistent with the number of ranging sensors.

[0015] Optionally, the output control instructions for the subsequent wafer placement process include: after outputting a final determination that the shape and position of the isolation paper are abnormal, sending a pause command to the robot control system and triggering an audible and visual alarm; and after outputting a final determination that the shape and position of the isolation paper are normal, sending a command to the robot control system to allow wafer placement.

[0016] A second aspect of the present invention provides a wafer isolator paper shape and position detection system for implementing the aforementioned wafer isolator paper shape and position detection method, comprising: a sensor unit, including a reflective sensor for identifying whether the top layer material of the wafer carrier container is isolator paper, multiple sets of distance measuring sensors for collecting distance data and arranged directly above the edge of the wafer carrier container, and an environmental parameter acquisition module for collecting environmental parameters; a data processing and judgment unit, for dynamically generating a composite feature set characterizing the shape and position state of the isolator paper based on the collected distance data and environmental parameters; performing dual judgment on the composite feature set, the dual judgment including: performing preliminary screening of the distance data using a dynamic threshold dynamically generated based on historical data; inputting the composite feature set corresponding to the data that passed the preliminary screening into a pre-trained machine learning classification model for morphological classification judgment; and outputting a final judgment on whether the shape and position of the isolator paper is abnormal based on the result of the dual judgment; and a control execution unit, for sending corresponding control commands to an external robotic arm control system based on the final judgment output by the data processing and judgment unit.

[0017] The present invention provides a wafer isolation paper shape and position detection method and system, which has the following significant beneficial effects.

[0018] 1. Significantly Improved Detection Accuracy and Reliability: By combining "preliminary screening of distance data using dynamically generated thresholds based on historical data" with "pre-trained machine learning classification models," a dual, progressive, and highly reliable decision-making framework is formed. This framework can not only quickly filter out significant anomalies but also accurately identify complex shapes with blurred boundaries.

[0019] 2. Significantly Enhanced Complex Defect Detection Capability: Traditional methods rely on simple threshold judgments, which have limited ability to identify defects such as local wrinkles, unilateral slight warping, and complex deformations. This invention constructs a composite feature set including statistical features, time-frequency domain features, and distribution morphology features, and utilizes the powerful pattern recognition capabilities of machine learning classification models such as support vector machines to capture the subtle signals of these complex defects, thereby increasing the detection rate of complex defects from approximately 72% to over 94%.

[0020] 3. Possesses adaptive and anti-interference capabilities: The threshold is generated based on real-time historical data and environmental parameter compensation, enabling it to automatically adapt to fluctuations in robot arm precision and long-term drift caused by equipment aging, as well as instantaneous interference caused by changes in workshop temperature and humidity. It eliminates the need for frequent manual adjustments to preset parameters, greatly improving the system's adaptability and stability to different working conditions.

[0021] 4. Achieved continuous intelligent optimization: The system's built-in online incremental learning mechanism enables it to continuously absorb new data generated during the production process (including data from new materials and new working conditions) and regularly update the machine learning classification model to ensure the optimal detection effect in the long term.

[0022] 5. Improved overall production line efficiency: Due to a significant reduction in the false alarm rate, especially the false positive rate, unplanned downtime caused by erroneous alarms from the detection system has been greatly reduced. Practical applications show that related downtime can be reduced by more than 68%. At the same time, the reduction in the false negative rate fundamentally reduces the risk of wafer breakage due to defects in the release liner, ensuring product quality and production efficiency. Attached Figure Description

[0023] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0024] Figure 1 This is the overall flowchart of the present invention.

[0025] Figure 2 This is a system framework diagram of the present invention. Detailed Implementation

[0026] The embodiments of this application will now be described in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. Furthermore, the following embodiments and features can be combined with each other unless otherwise specified. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0027] Example 1

[0028] Please see Figure 1 The present invention provides a method for detecting the shape and position of wafer isolation paper, which specifically includes the following steps.

[0029] S1: At the wafer packaging station, a reflective sensor is used to identify whether the top layer material of the wafer carrier container is release paper.

[0030] In one embodiment of this application, when performing the inspection task, the robotic arm first moves the wafer carrier container containing the target object (which may be a wafer or release liner) to a fixed packaging station. A reflection sensor located at the end of the robotic arm or a fixed point on the station is activated, emitting a beam of light of a specific wavelength towards the target surface and detecting the intensity of the reflected light. Since the wafer (usually a polished silicon surface) and the release liner (polymer or fiber material) have significantly different reflectivities in the visible or near-infrared bands (typically differing by more than three times), the reflection sensor makes a logical judgment based on this difference in reflected light intensity. If the reflection intensity is higher than a preset material differentiation threshold, the system determines that the top layer material is the wafer, eliminating the need for subsequent shape and position detection and proceeding to a waiting or wafer processing flow; if the reflection intensity is lower than the preset material differentiation threshold, the top layer is determined to be the release liner to be inspected, and subsequent shape and position measurement and analysis processes are initiated. This preliminary identification step ensures that the inspection mechanism is activated only when necessary, improving the overall system efficiency and avoiding unnecessary ranging operations on the wafer.

[0031] S2: If it is identified as isolation paper, it drives multiple sets of ranging sensors arranged directly above the edge of the wafer carrier container to synchronously collect distance data from each point on the edge of the isolation paper to the ranging sensor, and acquire environmental parameters in real time.

[0032] In one embodiment of this application, after confirming that it is release paper, the data processing and judgment unit immediately issues a synchronization trigger command to drive a set of ranging sensors (at least two, preferably four or more) pre-installed above the edge of the wafer carrier container opening to simultaneously perform measurements. These ranging sensors are uniformly and symmetrically arranged according to a preset, regular array geometry, typically a square (corresponding to four ranging sensors, located near the four corners of the outer frame above the wafer carrier container opening), a rectangle, or a regular polygon (such as a hexagon, corresponding to six ranging sensors). The geometric symmetry of the ranging sensors ensures the objectivity and consistency of the measurements. Each ranging sensor measures the vertical or approximately vertical distance from its mounting reference surface to the corresponding measurement point on the release paper below (usually in the edge area of ​​the release paper), obtaining a set of synchronized distance values ​​D, denoted as D=[d1, d2, ..., dM], where dM represents the distance value of the Mth ranging sensor. It should be noted that while constructing the distance value set D, the data processing and judgment unit unifies all distance values ​​to a unit (such as millimeters mm).

[0033] Simultaneously, an environmental parameter acquisition module (integrating at least a high-precision temperature sensor and a relative humidity sensor) located within the same environmental control area as the ranging sensor group begins to collect current environmental parameters in real time. Depending on the specific implementation environment, parameters such as air pressure can be expanded to be collected. These environmental parameters, including at least real-time ambient temperature T (unit: °C) and real-time relative humidity H (unit: %RH), will serve as an important component of dynamic adaptation, used to compensate for the impact of environmental changes on the ranging sensor measurements.

[0034] For example, in a certain working scenario, the specific scenario settings include (1) Wafer carrier container: standard 300mm wafer transfer box (FOUP), with a rectangular cavity inside, the side length is 320mm×320mm. (2) Isolation paper: circular polymer isolation paper, with a diameter of 300mm, a through hole with a diameter of 50mm in the center, uniform edge thickness, and low static polypropylene material. (3) Arrangement of ranging sensors: 4 KEYENCEIL-100 laser ranging sensors are installed directly above the FOUP opening, arranged in a square at the four corners, the horizontal distance between the center point of the laser ranging sensor and the inner wall edge of the FOUP is 10mm, and the optical axis of the sensor is vertically downward. (4) Target measurement point: the laser spot of each laser ranging sensor is precisely aligned with the four symmetrical points on the outer edge of the isolation paper (i.e., the four corner points of the square are projected onto the edge of the isolation paper). (5) Environmental parameters: real-time ambient temperature T=25.3℃, real-time relative humidity H=48% (standard working conditions). When the release paper is accurately and smoothly placed inside the FOUP by the robotic arm, without any offset, warping, or flatness, the distance values ​​measured by the four laser rangefinders are as follows (unit: millimeters, mm): D = [12.41, 12.39, 12.42, 12.40], where 12.41 refers to the distance measured by the upper left laser rangefinder, 12.39 refers to the distance measured by the upper right laser rangefinder, 12.42 refers to the distance measured by the lower right laser rangefinder, and 12.40 refers to the distance measured by the lower left laser rangefinder. This set of distance values ​​D represents the ideal shape and position of the release paper and is the core sample for training the "normal" category in the system. The system uses this set of distance values ​​D as the "qualification benchmark" for subsequent dynamic threshold calculation and positive sample training of the machine learning classification model.

[0035] S3: Based on the collected distance data and environmental parameters, dynamically generate a composite feature set representing the shape and position state of the isolation paper.

[0036] In one embodiment of this application, based on the collected distance value set D and environmental parameters (T, H), a composite feature set X that characterizes the current shape and position of the isolation paper and integrates multi-dimensional information is dynamically generated. The composite feature set X may contain any combination of one or more of the following feature classes, the selection of which depends on the specific performance requirements: (1) Statistical features: basic statistics extracted from the original distance data, such as: the average value of the distance data (reflecting the overall height), the difference between the maximum and minimum values ​​(i.e., the range, reflecting the overall fluctuation range of the distance data), variance or standard deviation (reflecting the dispersion of the distance data). (2) Time-frequency domain features: features extracted after performing time-frequency domain transformation (such as fast Fourier transform, wavelet transform) on the distance data collected in a certain spatial order (such as a spatial sequence formed by the arrangement order of the ranging sensors) or time sequence. For example, extracting the dominant frequency component of the sequence, which may be related to the vibration mode when the robot puts down the isolation paper, environmental mechanical noise, or a specific periodic warping mode; extracting the high-frequency energy ratio, which can be used to capture rapid, local, non-periodic small warping or wrinkles. It should be noted that the high-frequency energy ratio is a key indicator for measuring the proportion of high-frequency components in the frequency domain of a signal. It is often used to identify abrupt changes, spikes, or local anomalies (such as folds and wrinkles) in a signal. (3) Environmental correlation features: Environmental compensation factors calculated based on environmental parameters (T, H). These environmental compensation factors can be used as part of a composite feature set, enabling machine learning classification models to learn the differences in morphological representation under different environmental conditions, thereby achieving environmental adaptive judgment. (4) Distribution morphology features: Features obtained by analyzing the probability distribution morphology of distance data, such as skewness values ​​reflecting the asymmetry of data distribution and kurtosis values ​​reflecting the sharpness of distribution. Warping or offset often makes the distribution of distance measurement data exhibit asymmetric or "sharp" characteristics.

[0037] It should be noted that the measurement accuracy of laser rangefinders (such as KEYENCEIL-100, SICKLMS511, etc.) is significantly affected by ambient temperature and humidity, mainly manifested as: (1) Temperature drift: Thermal expansion and contraction of internal optical components (laser diode, receiver, lens) of the laser rangefinder causes optical path shift, and the output distance value undergoes a systematic shift with temperature change (typical value: ±0.01~0.05mm / ℃). (2) Humidity effect: Changes in the density of water molecules in the air affect the laser refractive index, especially in high humidity environments (>70%RH), the optical path undergoes a slight change (typical value: ±0.005mm / %RH). This invention establishes a quantitative mapping model between environmental parameters and sensor output drift, dynamically calculates environmental compensation factors, and uses them to correct the original distance value, eliminate environmental interference, and improve detection stability.

[0038] The dynamic calculation steps of the environmental compensation factor include: (1) Collecting real-time environmental parameters: Real-time environmental temperature T and real-time environmental humidity H are collected by a high-precision temperature and humidity sensor (such as SensirionSHT35) installed near the detection station. (2) Obtaining calibration reference parameters: The system pre-stores calibration reference environmental parameters (determined by factory calibration or initial calibration): standard temperature T0 and standard humidity H0. (3) Establishing an environmental drift response model: Based on the temperature / humidity drift characteristic curve provided by the sensor manufacturer, combined with the measured calibration data of this system in the laboratory environment, a linear compensation model is established. The formula is: E=α×(T-T0)+β×(H-H0), where: E is the environmental compensation factor, α is the temperature drift coefficient in mm / ℃, and β is the humidity drift coefficient in mm / %RH. Among them, α and β are system-level calibration constants, which are usually obtained by the equipment manufacturer through average calibration of multiple laser ranging sensors before leaving the factory and written into the system firmware or configuration file. (4) Substitute the real-time parameters to calculate the environmental compensation factor E: Using the above linear compensation model, substitute the current environmental parameters (real-time environmental temperature T and real-time environmental humidity H). (5) Input E as the feature value into the composite feature set.

[0039] Furthermore, the construction steps of the composite feature set include: (1) obtaining the calculated environmental compensation factor E. (2) performing environmental drift compensation on all original distance data, for example, if , , and obtain the compensated distance value, that is, the compensated distance value = 12.85 - 0.01782 ≈ 12.832 mm. (3) Extract statistical features, including: mean, difference, standard deviation, skewness and kurtosis. (4) Perform fast Fourier transform (FFT) and extract time-frequency domain features, including: main frequency component and high frequency energy ratio. (5) Construct the final composite feature set, such as composite feature set = [environmental compensation factor, mean, difference, standard deviation, skewness, kurtosis, main frequency component, high frequency energy ratio, real-time ambient temperature, real-time ambient humidity, compensated distance value], where the compensated distance value includes the distance value for environmental drift compensation of all original distance data.

[0040] S4: Perform a dual judgment on the composite feature set, the dual judgment including: using a dynamic threshold dynamically generated based on historical data to perform a preliminary screening of distance data; inputting the composite feature set corresponding to the data that has passed the preliminary screening into a pre-trained machine learning classification model for morphological classification judgment.

[0041] In one embodiment of this application, this step employs a dual-judgment architecture that links "dynamic threshold preliminary screening" (first-level judgment) and "machine learning classification model accurate judgment" (second-level judgment). The design goal of this progressive framework is to reduce the computational overhead of complex models through rapid filtering while ensuring an extremely low false negative rate, and to leverage the advantages of each to adapt to different scenarios.

[0042] 1. First-level judgment: rapid screening based on dynamic thresholds.

[0043] In the first stage of judgment, the data processing and judgment unit employs an adaptive, self-updating, and simplified threshold logic to quickly identify samples with severe and obvious morphological abnormalities, thereby achieving efficient preliminary filtering. Specifically, based on the acquired distance data, a rapid screening criterion is calculated (typically range, variance, or standard deviation are chosen as rapid features). Simultaneously, this dynamic threshold is calculated based on the statistical characteristics of a sliding window of historical qualified data collected and stored over a long period (e.g., retaining the most recent 500 ranging data samples ultimately confirmed as "morphologically normal"). The calculation process specifically incorporates the environmental parameters collected at the current moment, thereby providing real-time compensation for potential environmental drift of the ranging sensor (such as a laser ranging sensor).

[0044] The rapid screening logic is as follows: The current calculated criterion value (such as the range of this measurement) is compared with the dynamically generated latest dynamic threshold, taking into account environmental effects. If the range of this measurement significantly exceeds the "normal" fluctuation limit defined by the dynamic threshold (which can be set to the average of historical qualified data plus several times the standard deviation), then the paper is determined to have a serious defect in its shape (e.g., severe overall warping or severe detachment from the plane at a certain point), and the conclusion of "abnormal shape" is immediately drawn, while the process to the second step is terminated. This method efficiently filters out the vast majority of serious defects, greatly improving the system's response speed to "unqualified" cases and overall decision-making efficiency.

[0045] 2. Second judgment: Morphological classification judgment based on machine learning classification model.

[0046] For samples that pass the first stage of judgment and do not show obvious or extreme anomalies under the "preliminary screening" criteria, but may still have subtle, complex, or "suspicious" morphological defects, they are moved to the second stage of judgment, which is more refined and intelligent. At this point, a specially generated composite feature set X for the sample (it should be noted that all data within the composite feature set X needs to be normalized before being input into the machine learning classification model to eliminate dimensional differences) is input into a pre-trained machine learning classification model (such as Support Vector Machine (SVM), decision tree ensemble model, etc.) for judgment. Support Vector Machines (SVMs) are considered a powerful tool suitable for such structured composite features due to their excellent performance in handling small samples, nonlinear data, and high-dimensional data.

[0047] This machine learning classification model is based on a complex nonlinear mapping of a large number of learned "normal" and various "abnormal" samples (specific abnormality categories include overall warping, unilateral warping, folding, and center offset). It can use the composite feature set (statistical characteristics, frequency domain characteristics, distribution characteristics, etc.) constructed in the S3 step to explore deeper pattern associations and potential structures. It has a stronger recognition ability and higher detection confidence for defects with relatively subtle shapes that are difficult to distinguish by traditional thresholding methods or single-point judgment (such as slight unilateral upturning, small-scale convexity or folding in the center).

[0048] Furthermore, model fusion strategies can be employed to further enhance reliability. For example, the model can be constructed as a two-layer combination of Support Vector Machine (SVM) and Isolation Forest algorithm. SVM serves as the primary morphological classifier; the Isolation Forest algorithm is then added to specifically perform secondary anomaly detection on samples initially deemed "normal" by SVM. The Isolation Forest algorithm is an effective unsupervised (or semi-supervised) anomaly detection method, adept at identifying a minority of "isolated" samples whose patterns differ from the "majority." This dual machine learning verification mechanism (SVM used for supervised classification of broad categories, and the Isolation Forest algorithm used for unsupervised identification of anomalies outside the established patterns) effectively prevents potential misjudgments (false negatives) that may occur in the ambiguous classification boundaries of SVM, providing an additional and effective layer of protection for samples that appear "normal" but actually contain subtle, unlearned anomaly patterns. This is a key technical measure to reduce the false negative rate and enhance the ability to detect unknown and complex anomalies.

[0049] S5: Based on the results of the dual judgment, output the final judgment on whether the shape and position of the isolation paper are abnormal, and output the control instructions to control the subsequent wafer placement process.

[0050] In one embodiment of this application, after generating the final judgment result based on the aforementioned dual judgment framework, the system performs the final control and information notification tasks. The final judgment information is not just a simple "qualified" or "unqualified". In order to provide diagnostic guidance and facilitate subsequent process analysis, when the judgment result is "abnormal", the system can also associate the output of a machine learning classification model or the assistance of a specific algorithm to further distinguish the abnormality type, such as specifying "overall warping", "unilateral warping", "center offset", "folding", etc.

[0051] At the same time, the system immediately converts this judgment result into a closed-loop control instruction for the automated production line and executes it: (1) If the final judgment is "normal", a digital signal instruction is sent to the robot control system through the industrial bus (such as EtherCAT, Profinet, EtherNet / IP, etc.) to instruct it to continue execution or allow the wafer placement operation. (2) If the final judgment is "abnormal", regardless of whether the abnormality is detected directly by the first judgment screening or by the second machine learning classification model, the system will trigger the following operations: ① Send a "pause instruction" to the robot control system through the emergency stop or pause instruction channel to immediately pause or stop the current and subsequent related operations. ② Immediately activate the sound and light alarm device installed on the machine or operating area, such as flashing red light and buzzer sounding, or send a push alarm message to the equipment management terminal of the person in charge to remind the operator to check, adjust or replace the current workstation and isolation paper. This instruction execution logic ensures that high-risk work processes will stop immediately after potential problems are confirmed to prevent possible product damage.

[0052] To ensure that the entire detection system can effectively adapt to the slow changes on the production line caused by factors such as equipment aging, the use of new materials, and process adjustments, and to avoid the degradation of detection performance over time, the detection method of this invention is designed and incorporates an online model learning and updating mechanism.

[0053] During actual production testing, the system not only performs online judgments but also systematically and continuously collects complete raw data sequences (i.e., distance value sets and environmental parameters), calculated composite feature sets, and final online judgment results generated from each test in the background. This data initially forms a sample pool with "initial labels." Through manual sampling (i.e., operators periodically or based on system prompts manually checking the accuracy of some system judgment results) or long-term statistical analysis (e.g., tracking a batch of products for a period of time to determine if the wafer breakage rate is abnormally high in the later cleaning and packaging processes, thus inferring potential missed detections in the front-end testing), the system can acquire or gradually correct labels indicating that some data is "true" or highly reliable.

[0054] When new sample data with verified results (i.e., "qualified" or "unqualified," and potentially with more granular "defect type") accumulates to a predetermined trigger capacity (which can be set according to actual needs and computing resources, e.g., trigger capacity = 500N), a background incremental learning or retraining process is triggered. During periods of low computational load or dedicated service periods, the system will use this newly accumulated data to update, fine-tune, or retrain a portion of the core model of the machine learning classification model in S4.

[0055] Through this incremental learning and adaptive capability, the machine learning classification model, which was originally built on limited training data in the offline stage (initial deployment), can have its model parameters effectively updated and optimized over time, as the production line status evolves and different batches of materials are introduced. This enables the detection capability to evolve on its own, maintain and continuously adapt to the specific working conditions of the user's production line, greatly reducing the number of maintenance operations that require manual shutdown for system recalibration, and improving the overall lifecycle value and intelligence level of the detection system.

[0056] In summary, the wafer spacer paper shape and position detection method provided by this invention achieves high-precision and robust detection of the spacer paper's shape and position through a dual mechanism: a reflective sensor accurately identifies the spacer paper, multiple sets of distance sensors simultaneously collect distance data and combine it with environmental parameters to generate a composite feature set, and then a dynamic threshold initial screening and a machine learning classification model secondary judgment. On the one hand, the dynamic threshold combined with environmental compensation effectively eliminates interference from environmental factors such as temperature and humidity, reducing the false judgment rate. On the other hand, the introduction of the machine learning classification model not only accurately identifies various anomaly types such as overall warpage, single-sided warpage, folding, and center offset, but also continuously adapts to process fluctuations and material batch changes through an online learning and update mechanism, ensuring the long-term stability of detection performance. Simultaneously, the audible and visual alarms and robotic arm control command outputs under abnormal conditions can promptly prevent unqualified spacer paper from entering subsequent processes, effectively avoiding problems such as scratches and contamination of wafers due to spacer paper anomalies. This significantly improves the automation level and product yield of the wafer packaging process, providing reliable support for the refined and intelligent production of semiconductor manufacturing.

[0057] Example 2

[0058] Based on Embodiment 1 above, this embodiment also provides a wafer isolation paper shape and position detection system. Please refer to [link to previous embodiment]. Figure 2 This system is used to implement the steps of the wafer isolation paper shape and position detection method described in Embodiment 1 above. The main implementation modules of the system are: a sensor unit, a data processing and judgment unit, and a control execution unit. All units are organically coordinated through communication protocols such as fieldbus or industrial Ethernet.

[0059] In one embodiment of this application, the sensor unit is responsible for collecting the raw information required to perform logical judgments. The specific configuration consists of the following types of sensor components.

[0060] Reflection sensor: Preferably a miniature sensor with an adjustable distance sensing range, installed next to the suction cup or gripping module at the end of the pick-and-place robot (or mounted on a fixed bracket at the detection station) for non-contact scanning of the target surface. The reflection sensor outputs high and low levels based on the difference in light reflectivity of different materials, and the system uses this to determine whether the current detection point is a highly reflective wafer surface or a relatively matte release paper surface.

[0061] Multiple range sensors: This is the main component for acquiring the geometric information of the release liner. Laser triangulation sensors, laser confocal displacement sensors, or industrial-grade high-precision ultrasonic sensors can be used. The number of range sensors, M, is at least two (to detect flatness changes on a simple straight line), preferably four or more (to detect morphological anomalies at multiple points on a two-dimensional plane). The range sensors are grouped and mounted on a precision-machined rigid bracket or beam. This mounting structure is precisely positioned at a pre-calibrated working distance directly above the edge of the wafer carrier container opening, ensuring that the range sensors are vertically aligned (or meet a preset incident angle) with the preset measurement point on the edge of the release liner. The arrangement of the range sensors on the bracket, according to the invention's concept, should aim to stably and symmetrically capture the overall edge condition of the release liner. For example, a symmetrical geometric layout such as a square (four range sensors placed directly above the outer edges of the four virtual vertices), rectangle, or regular hexagon can be used. The number of range sensors determines the dimensionality and information richness of the final distance data.

[0062] Environmental parameter acquisition module: This typically includes one or more integrated sensor modules, with core components including at least a high-resolution digital temperature sensor and a relative humidity sensor. This module is usually integrated and installed near the sensor bracket of the detection system to ensure that the acquired temperature and humidity data effectively represent the real-time environmental conditions of the ranging sensor and the detected object. The environmental parameter acquisition module transmits real-time ambient temperature and humidity to the data processing and judgment unit via a digital interface (such as I2C or Modbus).

[0063] The data processing and judgment unit is implemented by an industrial computer (IPC), an embedded industrial control system (PLC), or a programmable automation controller (PAC) with equivalent computing power, and is equipped with a storage medium of sufficient capacity (such as SSD). It mainly includes the following functions: (1) receiving the reflection signal and synchronous trigger command feedback uploaded from the sensor unit, and after receiving the isolation paper confirmation signal, synchronously sending activation commands to all ranging sensors to complete the parallel acquisition and summarization of distance data. (2) performing preliminary denoising, unit conversion and verification on the acquired distance data and environmental parameters, and then dynamically constructing composite features according to the method provided in step S3. (3) performing dual judgment: performing judgment logic according to the process described in step S4: firstly, performing the first judgment, directly returning the result for data that is quickly screened and judged as "abnormal"; for data that passes the preliminary screening, calling the pre-trained machine learning classification model to perform morphological classification judgment, obtaining the final judgment and the detailed classification of defect types, and finally forming the decision result.

[0064] The control execution unit mainly includes the following functions: (1) When the data processing and judgment unit determines that the current detection target is abnormal, the control execution unit immediately drives the sound and light alarm light, alarm bell and / or the warning light on the top of the machine to flash and sound the alarm. At the same time, it can notify the workshop host computer or manufacturing execution system (MES) to pop up an alarm notification and record detailed information in the quality management log through digital I / O signal, and notify the on-site production, technical and management personnel to perform manual intervention and handling. (2) According to the detection judgment conclusion, the control execution unit outputs control instructions to the robot control system or robot controller through the preset industrial communication port and communication protocol. The control instructions include: a. Wafer placement "permission instruction": When the detection judgment is "normal" and the type has been determined to be without other faults, the control execution unit sends a digital signal or a specific communication message to the robot, instructing the robot to safely perform the wafer gripping and placement action on the current qualified isolation paper. b. Production line process "pause instruction": When the system confirms that there is a shape and position problem with the target isolation paper, the control execution unit will immediately and proactively send a sequence of instructions to the external robot control system to stop working or enter a safe pause state, so that the pick-up and put-down operation and related processes are immediately interrupted until the operator checks and handles the problem and manually resumes operation.

[0065] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting the shape and position of wafer isolation paper, characterized in that, include: At the wafer packaging station, a reflective sensor is used to identify whether the top layer material of the wafer carrier container is release paper. If it is identified as isolation paper, multiple sets of ranging sensors arranged directly above the edge of the wafer carrier container are driven to synchronously collect distance data from each point on the edge of the isolation paper to the ranging sensors, and environmental parameters are acquired in real time. Based on the collected distance data and environmental parameters, a composite feature set representing the shape and position state of the isolation paper is dynamically generated; The composite feature set includes at least: Statistical features extracted from distance data include the mean, the difference between the maximum and minimum values, and the variance or standard deviation. The dominant frequency component and / or high-frequency energy percentage extracted after performing time-frequency domain transformation on the distance data; Environmental compensation factors related to environmental parameters; Skewness values ​​reflecting the asymmetry of the distribution of distance data and / or kurtosis values ​​reflecting the sharpness; Perform a dual judgment on the composite feature set, the dual judgment including: The distance data is initially screened using dynamic thresholds generated dynamically based on historical data. The composite feature set corresponding to the data that has been initially screened will be input into a pre-trained machine learning classification model for morphological classification judgment. Based on the results of the dual judgment, the final determination of whether the isolation paper shape and position are abnormal is output, and the control instructions for controlling the subsequent wafer placement process are also output.

2. The wafer isolation paper shape and position detection method according to claim 1, characterized in that, The environmental parameters include at least real-time ambient temperature and real-time ambient humidity; the dynamically generated dynamic threshold is calculated by a sliding window based on maintained historical distance data, and environmental drift is compensated by combining environmental parameters during the calculation.

3. The wafer isolation paper shape and position detection method according to claim 2, characterized in that, The preliminary screening of distance data using a dynamic threshold dynamically generated based on historical data includes: Calculate the range, variance, or standard deviation of distance data; The range, variance, or standard deviation of the distance data are compared using a dynamic threshold calculated by a sliding window of historical distance data. If the range exceeds the threshold, the data is judged to be morphologically abnormal and the subsequent judgment is terminated. If the range does not exceed the threshold, the composite feature set corresponding to the initially screened data is input into a pre-trained machine learning classification model for further morphological classification judgment.

4. The wafer isolation paper shape and position detection method according to claim 1, characterized in that, The pre-trained machine learning classification model uses a support vector machine as the classifier and combines it with the isolated forest algorithm to perform secondary anomaly detection on results that are judged to be normal by the classifier.

5. The wafer isolation paper shape and position detection method according to claim 1, characterized in that, It also includes an online learning and update mechanism for the model: During the testing process, new sample data with true result labels are continuously collected; When the accumulated new sample data reaches a preset number, the incremental learning process is triggered to update the parameters of the pre-trained machine learning classification model to adapt to process fluctuations and material batch changes.

6. The wafer isolation paper shape and position detection method according to claim 1, characterized in that, If the final judgment on the shape and position of the isolation paper is abnormal as output by the result of the dual judgment, the abnormality type information is also output. The abnormality type information includes at least one of overall warping, single-sided warping, folding, and center offset.

7. The wafer isolation paper shape and position detection method according to claim 1, characterized in that, The multiple sets of ranging sensors are uniformly and symmetrically arranged in a rectangular or equilateral polygonal manner directly above the edge of the wafer carrier container, and the dimension of the distance data is consistent with the number of ranging sensors.

8. The wafer isolation paper shape and position detection method according to claim 1, characterized in that, The output control commands for subsequent wafer placement processes include: After outputting the final judgment that the shape and position of the isolation paper are abnormal, a pause command is sent to the robot control system and an audible and visual alarm is triggered. After outputting the final determination that the isolation paper is in the correct position, a command to allow wafer placement is sent to the robot control system.

9. A wafer isolation paper shape and position detection system, characterized in that, A method for detecting the shape and position of wafer isolation paper as described in any one of claims 1-8 includes: The sensor unit includes a reflective sensor for identifying whether the top layer material of the wafer carrier container is release paper, multiple sets of ranging sensors arranged directly above the edge of the wafer carrier container for collecting distance data, and an environmental parameter acquisition module for collecting environmental parameters. The data processing and judgment unit is used to dynamically generate a composite feature set representing the shape and position status of the isolation paper based on the collected distance data and environmental parameters; perform dual judgment on the composite feature set, the dual judgment including: using a dynamic threshold dynamically generated based on historical data to perform preliminary screening of the distance data; inputting the composite feature set corresponding to the data that passed the preliminary screening into a pre-trained machine learning classification model for morphological classification judgment; and outputting a final judgment on whether the shape and position of the isolation paper is abnormal based on the result of the dual judgment. The composite feature set includes at least: Statistical features extracted from distance data include the mean, the difference between the maximum and minimum values, and the variance or standard deviation. The dominant frequency component and / or high-frequency energy percentage extracted after performing time-frequency domain transformation on the distance data; Environmental compensation factors related to environmental parameters; Skewness values ​​reflecting the asymmetry of the distribution of distance data and / or kurtosis values ​​reflecting the sharpness; The control execution unit is used to send corresponding control commands to the external robot control system based on the final judgment output by the data processing and judgment unit.

Images