A non-imaging wafer surface defect detection method and storage medium based on a single pixel

Abstract

This invention discloses a non-imaging wafer surface defect detection method and storage medium based on a single pixel. The method includes projecting a preset coded illumination pattern onto the surface of a wafer under test, so that the coded illumination pattern is reflected by the wafer surface; acquiring the light intensity value of the reflected pattern; performing data segmentation and labeling processing on the light intensity value and the defect pattern; selecting K highly correlated features using a feature selection algorithm and back-reaming to obtain a new combination of coded illumination patterns; projecting the new combination of coded illumination patterns onto other wafer surfaces under test, acquiring the reflected light intensity value, and extracting features; and using a random forest algorithm to detect wafer defects using the extracted features. This invention eliminates the need for prior imaging and can directly acquire information about the wafer surface under test. Therefore, it is not limited by the image quality of the object, reduces the influence of factors such as optical systems, noise, and image distortion, and especially avoids the problem of virtual images, thus improving the robustness of wafer surface defect detection.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection technology, and specifically to a non-imaging wafer surface defect detection method and storage medium based on a single pixel. Background Technology

[0002] In the traditional field of wafer surface defect detection, surface defect detection involves first imaging and then combining it with complex image processing algorithms. The detection quality is limited by the image quality itself, thus placing high demands on image resolution and machine computing power. Furthermore, high-resolution images have more stringent requirements for the imaging environment, and the imaging method is affected by factors such as optical systems, noise, and image distortion, which may lead to a decrease in the accuracy and reliability of defect detection. The non-imaging wafer surface defect detection scheme proposed in this invention does not require prior imaging, and therefore is not limited by the image quality of the object. In particular, it avoids the problems of localized highlighting in the imaging caused by an overly smooth wafer surface and the virtual image formed by reflections of surrounding objects during the imaging process, greatly improving the robustness of wafer surface defect detection.

[0003] Furthermore, traditional imaging wafer surface defect detection systems typically require high-resolution cameras and complex optical components to construct the system and acquire high-quality images, significantly increasing system costs. Traditional imaging systems also generate massive amounts of image data, necessitating large-scale data acquisition and storage, and processing and analyzing this image data consumes substantial computational resources and time. Additionally, traditional imaging systems often employ complex image processing algorithms, further increasing the complexity of system design and development. The non-imaging wafer surface defect detection system proposed in this invention does not rely on complex imaging equipment, greatly reducing the cost of wafer surface defect detection systems. It uses a single-pixel detector to collect wafer-related information, simplifying system design and implementation, reducing the complexity and maintenance costs of optical components, and improving system reliability. By eliminating redundant data and employing simpler and more direct feature extraction and classification methods, it does not require high system computing power, further streamlining system design and development complexity. Summary of the Invention

[0004] The present invention proposes a non-imaging wafer surface defect detection method, system and device based on a single pixel, which can at least solve one of the technical problems in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A non-imaging wafer surface defect detection method based on a single pixel includes the following steps:

[0007] A preset coded illumination pattern is projected onto the surface of the wafer under test, so that the coded illumination pattern is reflected by the surface of the wafer under test;

[0008] The light intensity value of the reflected pattern is collected, and the light intensity value and the defect pattern are processed by data segmentation and labeling.

[0009] K highly correlated features are selected using a feature selection algorithm, and a new coded lighting pattern combination is obtained by reverse engineering.

[0010] After projecting the new coded lighting pattern combination onto the surface of other wafers to be tested, the reflected light intensity value is collected and features are extracted. The extracted features are then used to detect wafer defects through a random forest algorithm.

[0011] Furthermore, the step of projecting a preset coded illumination pattern onto the surface of the wafer under test, so that the coded illumination pattern is reflected by the surface of the wafer under test, specifically includes:

[0012] Preset The coded lighting pattern is constructed using the following formula, first from the formula Generate the Hadama base pattern:

[0013]

[0014] in, This represents the inverse transform of Hadamard, and:

[0015]

[0016] in These are the coordinates of the Hadama domain;

[0017] Using difference To obtain a coefficient Two measurements are required; one measurement is performed via projection. Base-coded lighting pattern The other is obtained through its reverse. ] obtained;

[0018] Preset coded lighting pattern The light intensity is projected onto the surface of the wafer under test and measured using a single-pixel detector. The single-pixel light intensity measurement is mathematically equivalent to the inner product between the Hadamard diagram and the object.

[0019] From this, the corresponding Hadamard coefficient can be obtained. :

[0020]

[0021] in, and It corresponds to and Measurements of illumination, The coefficients are real values. The number of coefficients is the same as the number of pixels in the image; using difference. right A pixel-by-pixel image requires full sampling Second measurement.

[0022] Furthermore, the feature selection algorithm selects K highly relevant features, specifically determining the K most effective corresponding points based on the strong correlation obtained from the chi-square test, as follows:

[0023] The association between categorical variables is evaluated by comparing the differences between observed and expected values; the chi-square statistic of the contingency table is calculated.

[0024]

[0025] in It is the frequency of observations. It is the expected frequency, calculated based on the number of samples in each category and the total number of samples in the contingency table; specifically:

[0026]

[0027] in For the total count of the specified rows, It is the total count of the specified column. It is the total count of all observations;

[0028] By normalizing the chi-square statistics, the chi-square score is... By limiting it to a fixed range, comparisons can be made between different features without being affected by the number of feature values.

[0029]

[0030] The obtained chi-square scores are arranged in descending order, and the top K features with the strongest correlation with the target variable are selected as the selection results. Based on the correspondence between the Hadamard coefficient and the coded lighting pattern, a new combination of coded lighting patterns is deduced.

[0031] Furthermore, after projecting using the new coded illumination pattern, data is collected and feature extraction is performed. Before feature extraction, the obtained signal needs to be normalized to eliminate the influence of the data itself, overcome overfitting, and ensure that the features have the same scale and distribution, so as to facilitate comparison between different features. The normalization process is as follows:

[0032]

[0033] in These are the normalized eigenvalues. , These are the maximum and minimum values ​​of each feature, respectively.

[0034] Furthermore, after projecting the new coded lighting pattern, data is collected and feature extraction is performed, including extracting and calculating the standardized data features, as follows:

[0035] Peak value (PK): ;

[0036] Minimum (Min): ;

[0037] Peak-to-Peak value (PK-PK): ;

[0038] Average Energy (E): ;

[0039] Mean (A): ;

[0040] Average rectified value (Arv): ;

[0041] Root Mean Square (RMS) Energy: ;

[0042] Variance (var): ;

[0043] Kurtosis (Kur): ;

[0044] Skewness (Skew): ;

[0045] Clearance Factor (CL): ;

[0046] Shape Factor (Sf): ;

[0047] Crest Factor (Cf): ;

[0048] Impulse Factor (If): .

[0049] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0050] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0051] As can be seen from the above technical solution, the key to this invention is to apply a strategy for screening illumination coding patterns to the data preprocessing stage, combined with the advantage of obtaining information through non-imaging methods using a single pixel, and applying it to wafer defect detection, thus proposing a non-imaging wafer defect detection system. The proposed solution can acquire information about the wafer under test without any prior imaging. The application of the screening strategy achieves undersampling, reducing redundant data while acquiring wafer information, significantly reducing the amount of data required compared to traditional imaging-based wafer defect detection. The non-imaging method effectively avoids the problems of localized high brightness in imaging due to the excessive smoothness of the wafer surface and the virtual image formed by reflections of surrounding objects during imaging, which are problems that traditional imaging methods cannot directly avoid. The construction of the non-imaging wafer defect detection system does not rely on complex imaging equipment; it uses a single-pixel detector to collect wafer information, simplifying system design and implementation, reducing the complexity and maintenance costs of optical components, which is impossible with traditional imaging methods.

[0052] Compared to traditional imaging-based wafer defect detection methods, the non-imaging wafer defect detection scheme proposed in this invention does not require prior imaging and can directly acquire information about the surface of the wafer under test. Therefore, it is not limited by the image quality of the object and reduces the influence of factors such as optical systems, noise, and image distortion. In particular, it avoids the problems of localized highlighting in the image caused by an overly smooth wafer surface and the virtual image formed by reflections of surrounding objects during the imaging process, greatly improving the robustness of wafer surface defect detection. Furthermore, the proposed scheme filters out redundant data during the detection process, requiring less data and resulting in higher detection efficiency. Traditional imaging systems generate massive amounts of image data, requiring large-scale data acquisition and storage. Simultaneously, the total parameters and model size required during the detection process are smaller and more streamlined compared to traditional imaging-based wafer defect detection models, requiring less computational power. Traditional imaging-based wafer defect detection systems require high-resolution cameras and complex optical components to complete the system construction to obtain high-quality images, significantly increasing system costs. The proposed non-imaging wafer defect detection system has a simpler structure, reducing the complexity and maintenance costs of optical components and improving system reliability. Attached Figure Description

[0053] Figure 1 This is a flowchart of an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of an embodiment of the present invention;

[0055] Figure 3 These are partial wafer patterns according to embodiments of the present invention, wherein (a) represents no defects and (b) represents different types of defects;

[0056] Figure 4 This is a schematic diagram of the experimental apparatus according to an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram comparing the classification effects under different K values ​​in an embodiment of the present invention. Detailed Implementation

[0058] 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. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0059] like Figure 1 As shown, the non-imaging wafer detection method proposed in this embodiment of the invention is based on a single pixel and includes the following steps:

[0060] First, a pre-defined coded illumination pattern is projected onto the surface of the wafer under test, completing the information acquisition of the wafer surface. Simultaneously, this invention uses a single-pixel detector to collect the light intensity values ​​reflected from the reflection pattern as data samples. A portion of these samples is collected and matched with corresponding wafer defect types. This data is segmented and labeled for subsequent feature selection. A feature selection algorithm is used to filter the labeled data, where the filtered data values ​​are the features most relevant to the performance and interpretability of the prediction model. This invention obtains a new set of coded illumination patterns by filtering the compressed coded data and the coded illumination pattern. Subsequently, this invention uses the selected coded illumination pattern combination to project onto other wafer surfaces under test, similarly collecting the reflected light intensity values ​​and performing feature extraction. Finally, a random forest algorithm is used to classify the wafer defects. It is worth noting that the selected coded illumination patterns are sensitive to defects of the same type, meaning that this invention, through the execution of a single filtering strategy, can be applied to subsequent projection operations, ensuring the system's detection efficiency.

[0061] Specifically, it includes the following steps:

[0062] A preset coded illumination pattern is projected onto the surface of the wafer under test, so that the coded illumination pattern is reflected by the surface of the wafer under test;

[0063] The light intensity value of the reflected pattern is collected, and the light intensity value and the defect pattern are processed by data segmentation and labeling.

[0064] K highly correlated features are selected using a feature selection algorithm, and a new coded lighting pattern combination is obtained by reverse engineering.

[0065] After projecting the new coded lighting pattern combination onto the surface of other wafers to be tested, the reflected light intensity value is collected and features are extracted. The extracted features are then used to detect wafer defects through a random forest algorithm.

[0066] The following are detailed explanations:

[0067] like Figure 2 As shown, the non-imaging wafer detection method proposed in this invention fully utilizes the characteristics of single pixels, namely the correspondence between compressed and encoded data and encoded illumination patterns. By applying a new screening strategy to preprocess the encoded illumination patterns, new combinations of encoded illumination patterns are selected to achieve undersampling, thus achieving preliminary data simplification. The one-dimensional signal after compression and encoding is collected directly for feature extraction, further simplifying the complexity of feature engineering. Fourteen feature points are obtained through simple calculation, and finally, the classification is completed by combining the random forest algorithm.

[0068] Preset The coded lighting pattern is constructed using the following formula, first generating the Hadamard base pattern using formula (1):

[0069]

[0070] in, This represents the inverse Hadhamada transformation, and:

[0071]

[0072] in These are the coordinates of the Hadama domain;

[0073] To suppress noise and ensure more stable feature information before screening and encoding lighting patterns, we employ differential encoding. To obtain a coefficient Two measurements are required. One measurement is performed via projection. Base-coded lighting pattern The other is obtained through its reverse. ] obtained.

[0074] Preset coded lighting pattern The light intensity is projected onto the surface of the wafer under test and measured using a single-pixel detector. Single-pixel light intensity measurement is mathematically equivalent to the inner product between the Hadamard diagram and the object.

[0075] From this, the corresponding Hadamard coefficient can be obtained. :

[0076]

[0077] in, and It corresponds to and Measurements of illumination, The coefficients are real values. The number of coefficients is the same as the number of pixels in the image. Difference is used. right A pixel-by-pixel image requires full sampling Second measurement.

[0078] Although the one-dimensional light intensity signal obtained under full sampling contains a considerable amount of data, this state is not particularly advantageous due to the sparsity of defect distribution, with a significant amount of redundant data affecting detection efficiency. This invention starts with the compressed coded data before image reconstruction. Since the Hadamard coefficients after compression correspond to the coded illumination pattern, this provides a good approach for the undersampling processing module. This invention achieves undersampling through feature selection. Compared to feature extraction, the focus of feature selection is to optimize feature selection by applying scoring and ranking methods to retain the most suitable features representing the basic data characteristics.

[0079] The chi-square test, used in SelectKBest to assess the relationship between categorical variables, selects the K most relevant features from the original feature set to improve the performance and interpretability of the predictive model. It measures the association or independence between two categorical variables by comparing observed and expected frequencies. A significant relationship between the variables is determined by calculating the chi-square statistic and comparing it to the critical value in the chi-square distribution. The K most effective corresponding points are then determined based on the strong association obtained from the chi-square test. Its principle can be expressed as follows:

[0080] First, the chi-square statistic is calculated using a contingency table. The distribution of the contingency table involved in this invention is as follows:

[0081]

[0082] In this table, a_0, a_1, b_0, b_1, c_0, c_1, etc. represent the cross-frequency of the feature values ​​and the target variable values, which are used for subsequent observations and the calculation of expected frequencies.

[0083] The association between categorical variables is evaluated by comparing the differences between observed and expected values. The chi-square statistic for this contingency table is calculated as follows:

[0084]

[0085] in It is the frequency of observations. This is the expected frequency, calculated based on the sample size of each category and the total sample size in the contingency table. Specifically:

[0086]

[0087] in For the total count of the specified rows, It is the total count of the specified column. It is the total count of all observations.

[0088] By normalizing the chi-square statistic, the chi-square score ( This limits the comparison to a fixed range, allowing for comparisons between different features without being affected by the number of feature values.

[0089]

[0090] The obtained chi-square scores are sorted in descending order, and the top K features with the strongest correlation to the target variable are selected. To obtain coded illumination patterns consistent with the optimized K feature values, feature indexing is performed on each data point. After filtering these coded illumination patterns, they are used as a mixture of new coded illumination patterns. The new combination of coded illumination patterns replaces the previous ones. The encoded lighting pattern combination is projected to obtain highly correlated feature information while achieving undersampling. It is worth noting that the newly selected encoded lighting pattern combination is also sensitive to similar target patterns (including wafers and their corresponding surface defects), and the obtained data has strong correlation, which is similar to the effect of feature enhancement. Of course, this invention can also be understood as a dimensionality reduction process in data preprocessing, which can more efficiently compress the encoded data to obtain one-dimensional data, facilitating subsequent feature extraction in this invention.

[0091] After projecting using the new coded illumination pattern, data is collected and features are extracted. Before feature extraction, the obtained signals need to be normalized to eliminate the influence of the data itself, overcome overfitting, and ensure that the features have the same scale and distribution, so as to facilitate comparison between different features.

[0092]

[0093] in These are the normalized eigenvalues. , These are the maximum and minimum values ​​of each feature, respectively.

[0094] Therefore, this invention can obtain standardized feature data. Compared with high-dimensional feature data such as images, these feature representations are more concise, and the relevant feature extraction can be completed more efficiently. Finally, this invention performs extraction calculations on the standardized data features to obtain 14 relevant feature points, including energy, skewness, kurtosis, peak factor, and shape factor, as shown in the table below.

[0095]

[0096] By calculating the aforementioned feature points, high-dimensional data is transformed into a low-dimensional representation. This data abstraction process helps capture important features in the data, reduces redundant and irrelevant information, highlights the essential attributes of the data, and provides more refined and effective input for subsequent classifiers. Compared with the original feature data, these feature points focus more on the important features of the data, have better resistance to changes and noise between data, reduce noise and overfitting problems, and improve the generalization ability and interpretability of the model. Moreover, the smaller number of feature points improves the computational efficiency of the classifier in large-scale datasets and real-time application scenarios, while saving computational resources and storage costs.

[0097] After comparing the classification results of different methods, a random forest classifier was ultimately chosen for classification. The random forest classification algorithm is an ensemble learning method that successfully solves the overfitting problem and can handle feature selection and classification of datasets characterized by nonlinear variables. This algorithm has significant versatility and inclusiveness, better adapting to the data being examined. Furthermore, the random forest model exhibits excellent accuracy and generalization ability, placing it among the top machine learning classifiers. This invention employs cross-validation for result evaluation. Cross-validation allows for evaluation of results, improves data utilization, minimizes the possibility of overfitting, and enhances reliability and generalization ability in various applications.

[0098] It should be specifically noted that the single-pixel imaging basis pattern used in the embodiments of the present invention can be a basis pattern such as Hadamard, Fourier, or Wavelet; feature selection can also adopt other screening strategies, and feature extraction and classification algorithms can be replaced by other methods; the projection device can be a projector, LCD, DMD, or other spatial light modulators to modulate the projection pattern; the photodetector can be a photocell, photodiode, or other optoelectronic device that responds to light intensity information. The non-imaging wafer defect detection system of the embodiments of the present invention can adopt a passive structural imaging method.

[0099] like Figure 3 As shown, this invention obtained real silicon wafer image data from a semiconductor company. This data falls under the category of patternless wafers. To enhance the effectiveness of the dataset, the collected data was cropped and rotated. Each image is 64*64 pixels in size, with defect-free and defective wafers evenly distributed, serving as the target pattern for subsequent processing. This invention extracts five defect-free and five defective real silicon wafer data images as examples.

[0100] The non-imaging wafer defect detection system device of the present invention, such as Figure 4As shown, the filtered coded lighting pattern is projected onto the surface of the wafer under test by a projector. The one-dimensional light intensity signal reflected (or transmitted) from the object is detected by a single-pixel detector and converted into an electrical signal. The data acquisition card inputs the acquired signal into the computer. The data processing and classification work is completed through the above steps.

[0101] To demonstrate the feasibility of the proposed scheme in screening coded illumination patterns, this invention conducted a series of experiments, randomly assigning K values ​​from different directions. In this paper, K is set to 1000 for illustrative purposes. The method of this invention first focuses on the frequency aspect. This invention observes that the energy of ordinary signals tends to concentrate in the low-frequency range. Furthermore, these signals become even more concentrated in the low-frequency region after Hadamard transform. This invention finds that higher-energy-level frequency components play an important role in capturing signal feature information. Therefore, this invention extracts 1000 low-frequency coded illumination patterns from differential Hadamard coded illumination patterns, denoted as K''. Similarly, considering the presence of defects in silicon wafers, this invention concludes that high-frequency components are more effective in capturing signal details and rapid changes. In the Hadamard spectrum, the energy concentration in the high-frequency region indicates the presence of more detailed information in these frequency components. Therefore, this invention extracts 1000 high-frequency coded illuminations.

[0102] The pattern is also called K''. Finally, to ensure the comprehensiveness of the information, this invention selected 1000 projection patterns covering low, medium, and high mixed frequencies, denoted as K'''.

[0103]

[0104] As shown in the table, this invention selected three evaluation metrics—overall accuracy, recall, and false positive rate—to compare the effectiveness of different combinations in detecting real silicon wafer defects. The results show that the classification results achieved using high-frequency, mixed-frequency, and low-frequency coded lighting patterns are considered negligible. According to the data, the overall accuracy and recall of the combination of high-frequency, mixed-frequency, and low-frequency coded lighting patterns are approximately 80%, 85%, and 89%, respectively. On the other hand, the proposed method achieves approximately 96% on both evaluation metrics. Regarding the false positive rate, the combination of high-frequency, mixed-frequency, and low-frequency coded lighting patterns has a false positive rate of 23%, 17%, and 10.6%, respectively, while the proposed method has a false positive rate of 6.4%. These findings demonstrate the feasibility of the proposed screening and filtering strategy and further support the previous assertion that the screened coded lighting patterns exhibit equal sensitivity to similar defects, enabling more effective acquisition of feature information related to these defects for subsequent processing.

[0105] The analysis of this invention found that using K = 1000 can provide reliable information about the wafer surface. After processing the features extracted from this information by a random forest classifier, the recognition accuracy reached 96%. In other words, by implementing this screening strategy, this invention significantly reduces the number of illumination patterns that need to be encoded, while retaining the basic information of the target pattern. This achieves the effect of undersampling, simplifying the information acquisition process. Setting the K value to 1000, and the difference... Compared to the coded illumination mode, the low sampling rate of 12.2% allows this invention to retain effective information about the target mode. To further explore ways to reduce the limitations of the projected illumination coding mode, this invention conducted comparative experiments using different K values.

[0106] like Figure 5 As shown, through a careful selection process, Figure 5 Individual validation comparisons were performed on cases a, b, and c using different K values: 1000, 500, and 250. This invention consistently applied feature extraction to all selected cases and evaluated their performance using a random forest model. As shown in the figure, the results of this invention indicate that different K values ​​have almost no impact on the training set, as the accuracy remains close to 1 in all cases. However, significant differences emerged between different K values ​​when evaluating the validation set using cross-validation. Specifically, when K was set to 1000, the average accuracy of the validation set was approximately 95%, while for K of 500, the accuracy ranged from 89% to 92%, and for K of 250, the accuracy was 85%. These results demonstrate the adaptability of the proposed screening strategy, achieving diverse classification results by adjusting the K value. This aligns with the principle of feature extraction, where obtaining more data after feature extraction facilitates more effective analysis. Notably, even with a K value of 250, representing a low sampling rate of 3%, this invention still achieved a commendable 85% recognition accuracy, reaching the level of successful identification and detection. Overall, the results of this invention show that a higher K value can improve classification performance, indicating that more data features are extracted and integrated, which meets the throughput requirements of random forest classifiers.

[0107] In summary, this invention provides a non-imaging wafer surface defect detection method and system based on single-pixel (SPI). This method effectively reduces the amount of data required for detection by proposing a screening strategy for coded illumination patterns. Furthermore, the non-imaging single-pixel detection scheme acquires features by obtaining the intensity of reflected light signals. Compared to traditional imaging wafer surface detection methods, which have stringent requirements for high-resolution images and image processing algorithms, this method achieves efficient detection of silicon wafer defects under low undersampling conditions. In addition, this invention does not require imaging of the wafer surface, thus avoiding the high-brightness problem caused by an overly smooth wafer surface and the virtual image generated by reflections of surrounding objects, greatly improving the robustness of wafer surface defect detection. Moreover, the non-imaging wafer surface detection system proposed in this invention uses a single-pixel detector to collect information, eliminating errors caused by mutual interference between photosensitive elements that may exist in traditional high-resolution cameras, thereby improving the accuracy of defect detection. At the same time, the system structure using a single-pixel detector simplifies system design and implementation, reduces the complexity and maintenance costs of optical components, and improves system reliability. This invention has broad application potential in semiconductor manufacturing and optical component production.

[0108] In summary, traditional imaging wafer surface inspection methods first image the wafer with a high-resolution camera, then employ complex image processing algorithms or deep learning algorithms for feature extraction and further analysis. The wafer surface inspection method proposed in this invention eliminates the need for any prior imaging of the wafer, allowing information acquisition directly through a single-pixel detector. This invention uses data preprocessing to filter out corresponding coded illumination patterns sensitive to defects in similar wafers, which are then used for subsequent wafer information acquisition and classification. Furthermore, compared to complex image processing algorithms and computationally demanding deep learning algorithms, this invention employs the simplest feature extraction method and classification algorithm.

[0109] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0110] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0111] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the single-pixel-based non-imaging wafer surface defect detection methods described above.

[0112] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0113] This application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus.

[0114] Memory, used to store computer programs;

[0115] When the processor executes the program stored in the memory, it implements the above-described non-imaging wafer surface defect detection method based on a single pixel.

[0116] The communication bus mentioned in the aforementioned electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0117] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0118] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0119] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0120] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0122] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0123] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-imaging wafer surface defect detection method based on a single pixel, characterized in that, Includes the following steps, A preset coded illumination pattern is projected onto the surface of the wafer under test, so that the coded illumination pattern is reflected by the surface of the wafer under test; The light intensity value of the reflected pattern is collected, and the light intensity value and the defect pattern are processed by data segmentation and labeling. K highly correlated features are selected using a feature selection algorithm, and a new coded lighting pattern combination is obtained by reverse engineering. After the new coded lighting pattern combination is projected onto the surface of other wafers to be tested, the reflected light intensity value is collected and the feature is extracted. The extracted features are then used to detect wafer defects through the random forest algorithm. The step of projecting a preset coded illumination pattern onto the surface of the wafer under test, so that the coded illumination pattern is reflected by the surface of the wafer under test, specifically includes: Preset The coded lighting pattern is constructed using the following formula, first from the formula Generate the Hadama base pattern: in, This represents the inverse transform of Hadama, and: in These are the coordinates of the Hadama domain; Using difference To obtain a coefficient Two measurements are required; one measurement is performed via projection. Base-coded lighting pattern The other is obtained through its reverse. ] obtained; Preset coded lighting pattern The light intensity is projected onto the surface of the wafer under test and measured using a single-pixel detector. The single-pixel light intensity measurement is mathematically equivalent to the inner product between the Hadamard diagram and the object. From this, the corresponding Hadamard coefficient can be obtained. : in, and It corresponds to and Measurements of illumination, The coefficients are real values. The number of coefficients is the same as the number of pixels in the image; using difference. right A pixel-by-pixel image requires full sampling Second measurement.

2. The non-imaging wafer surface defect detection method based on a single pixel according to claim 1, characterized in that: The feature selection algorithm selects K highly relevant features, specifically determining the K most effective corresponding points based on the strong correlation obtained from the chi-square test, as follows: The association between categorical variables is evaluated by comparing the differences between observed and expected values; the chi-square statistic of the contingency table is calculated. in It is the frequency of observations. It is the expected frequency, calculated based on the number of samples in each category and the total number of samples in the contingency table; specifically: in For the total count of the specified rows, It is the total count of the specified column. It is the total count of all observations; By normalizing the chi-square statistics, the chi-square score is... By limiting it to a fixed range, comparisons can be made between different features without being affected by the number of feature values. The obtained chi-square scores are arranged in descending order, and the top K features with the strongest correlation with the target variable are selected as the selection results. Based on the correspondence between the Hadamard coefficient and the coded lighting pattern, a new combination of coded lighting patterns is deduced.

3. The non-imaging wafer surface defect detection method based on a single pixel according to claim 1, characterized in that: After projecting using the new coded illumination pattern, data is collected and feature extraction is performed. Before feature extraction, the obtained signal needs to be normalized to eliminate the influence of the data itself, overcome overfitting, and ensure that the features have the same scale and distribution, so as to facilitate comparison between different features. The normalization process is as follows: in These are the normalized eigenvalues. , These are the maximum and minimum values ​​of each feature.

4. The non-imaging wafer surface defect detection method based on a single pixel according to claim 3, characterized in that: After projecting the new coded lighting pattern, data is collected and feature extraction is performed, including extracting and calculating the standardized data features, as follows: Peak value (PK): ; Minimum (Min): ; Peak-to-Peak value (PK-PK): ; Average Energy (E): ; Mean (A): ; Average rectified value (Arv): ; Root Mean Square (RMS) Energy: ; Variance (var): ; Kurtosis (Kurtosis): ; Skewness (Skew): ; Clearance Factor (CL): ; Shape Factor (Sf): ; Crest Factor (Cf): ; Impulse Factor (If): .

5. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 4.

Images