Defect detection apparatus and method

Abstract

The present application provides a kind of defect detection method and equipment.The method includes obtaining the optical image of multiple groups of target units;The optical image of each target unit in the multiple groups of target units is converted into a spot distribution map;At least one spot feature of each spot in the spot distribution map is extracted based on the characteristic attribute of target defect;And whether the target unit exists the target defect is judged by determining whether the spot distribution map contains abnormal spot by using unsupervised machine learning processing the spot features of the multiple groups of target units.

Description

TECHNICAL FIELD

[0001] The present application relates to the packaging and testing of chips, and in particular to defect detection in the packaging and testing process. BACKGROUND

[0002] In the high volume manufacturing (HVM) environment of semiconductor chips, it is common to use machines to assemble and test the chips, during which process it is inevitable to affect the yield of the chips. For example, due to the foreign matter falling into the testing machine, the chip pin soldering points are damaged when the machine is operated, or the chip surface (i.e. the back of the chip) is left with a pressure mark. For example, FIG. 1A shows a schematic diagram of a captured image of the back of a chip processed by a testing machine, in which A-A is an enlarged view of the area where a pressure mark exists, and B-B shows the pressure mark left on the back of the chip, as shown in the figure, which also partially damages the soldering points. As can be seen, due to the machine hardware failure problem, the chips operated by the machine will have systematic pressure mark defects, which can have unpredictable effects. In addition, during the assembly and testing of the machine, the testing machine may also mistakenly stick foreign matter such as solder balls on the surface of the chip.

[0003] All these defects can adversely affect the quality and reliability of the chips, so they need to be detected before being sent to the customer. Since these physical defects can be very small, image analysis techniques are usually used to identify them. One commonly used method is rule-based image analysis, for example, after converting the optical image of the chip into a spot distribution map, by measuring the parameters of each spot in the spot distribution map, such as length, width, brightness or area, and comparing the measured parameter values with an empirical threshold to determine whether a defect has occurred. For example, for the pressure mark shown in FIG. 1A, the size (length, width or area, etc.) of the pressure mark is tested and compared with a threshold to determine the presence of such a pressure mark. Obviously, the rule-based method is only suitable for defects with fixed characteristics, and cannot adapt to unpredictable defect types, and even for the same defect such as foreign matter, there is still a possibility of missing detection when the shape of the foreign matter changes. Therefore, mechanically using fixed rules to judge defects can easily lead to false alarms or missed alarms. In the prior art, deep learning (such as based on convolutional neural networks) is also used to analyze chip images, so that it can automatically determine whether the currently processed chip has a defect. However, the current solution can be used to automatically learn and predict defects, but such deep learning needs to train the network model based on a large number of samples, and requires a large amount of calculation when implemented, so it is not conducive to real-time detection on the production line. SUMMARY

[0004] An improved detection scheme for automatically and effectively detecting defects occurred on chip units is proposed. According to one aspect of the present application, a defect detection method is provided, comprising: obtaining optical images of a plurality of target units; converting the optical image of each target unit in the plurality of target units into a spot distribution map; extracting at least one spot feature of each spot in the spot distribution map based on a characteristic attribute of a target defect to establish a spot feature dataset of the plurality of target units; determining whether the spot distribution map contains abnormal spots by processing the spot feature dataset of the plurality of target units using unsupervised machine learning to judge whether the target unit has the target defect.

[0005] According to another aspect of the present application, a defect detection device is provided, comprising an image obtaining module configured to obtain optical images of a plurality of target units; an image converting module configured to convert the optical image of each target unit in the plurality of target units into a spot distribution map; a feature determining module configured to extract at least one spot feature of each spot in the spot distribution map based on a characteristic attribute of a target defect to be detected to establish a spot feature dataset of the plurality of target units; and an anomaly identifying module configured to determine whether the spot distribution map contains abnormal spots by processing the spot feature dataset of the plurality of target units using unsupervised machine learning to judge whether the target unit has the target defect.

[0006] In one preferred example of the present application, the anomaly identifying module is further configured to perform clustering processing on the abnormal spots of the plurality of target units to determine whether the abnormal spots have fixed positional pattern characteristics; and when it is determined that the abnormal spots have fixed positional pattern characteristics, it is judged that the target unit has the target defect at the position indicated by the positional pattern. Here, the unsupervised machine learning used can be an Isolation Forest algorithm; and the clustering processing performed on abnormal spots can apply a Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0007] FIG. 1A shows a schematic diagram of a chip backside image;

[0008] FIG. 1B shows a schematic diagram of a spot distribution of a chip backside image;

[0009] Figure 2 shows a schematic diagram of a defect detection system according to the present application;

[0010] Figure 3 shows a schematic diagram of a defect detection device according to one example of the present application;

[0011] Figure 4 shows a flowchart of a defect detection method according to one example of the present application. Detailed Implementation

[0012] Exemplary embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention. In the embodiments of the invention, a 'target defect' is defined as damage to the chip surface or the attachment of foreign matter, such as indentations on the chip surface caused by a testing machine, damage to the solder balls of the pins along with the indentations (e.g., causing deformation, loss, etc. of normal solder balls on the pins), damage simply to the solder balls of the pins, or solder balls mistakenly implanted in non-pin locations, etc.

[0013] Figure 2 A block diagram of a detection system according to the present invention is shown. As shown, the system includes at least one light source 100, an image sensor 200 (e.g., a camera), and a defect detection device 300. The light source 100 illuminates the back side of the chip to be inspected, while the image sensor 200, under the control of the detection device 300, captures an image of the back side of the chip, hereinafter referred to as BImg, where chip pins are arranged on the back side of the chip, as shown in, for example, Figure 1A. According to an embodiment of the present invention, the light source 100 can emit light of different intensities or wavelengths under the control of the detection device 300, thereby making it easier to highlight the characteristic differences of defects such as solder balls or indentations relative to the back side of the chip when the light illuminates it. The light source 100 can also be multiple light sources that emit different intensities or wavelengths, and different light sources can be switched under the control of the detection device 300 to illuminate the chip.

[0014] On a production line, to improve processing speed, a single testing machine typically contains multiple test units (TUs) to package and test chips in parallel. According to embodiments of the present invention, chips packaged and tested by different test units are identified, thus associating each tested chip with a corresponding test unit (TU). In one example, the chips being tested can be grouped by test unit (TU), so that each group of chips is associated with a corresponding test unit (TU). For instance, a group of chips processed by test unit (TU1) on the testing machine is uniformly identified as TU1, while a group of chips processed by test unit (TU2) is identified as TU2, and so on.

[0015] The testing equipment 300 receives multiple sets of optical images (BImg) of the chips captured by the sensor 200 and analyzes these optical images to determine whether there are defects on the chips after they have been assembled and tested by the testing machine. Figure 3A block diagram of the detection device 300 according to one example is shown. As shown, the detection device 300 comprises an image acquisition module 301, an image conversion module 302, a feature determination module 303, and an anomaly identification module 304. The image acquisition module 301 is configured to trigger the sensor 200 to capture an optical image BImg of the backside of a chip with the arrangement of solder balls on the pins when the light source 100 illuminates the chip, and to receive and buffer the image BImg captured by the sensor 200. In another example of the present application, the image acquisition module 301 can also directly receive an image BImg of a batch of chips that is previously captured (e.g. by a third party).

[0016] According to one example of the present application, each of the buffered chip images BImg is associated with a test unit TU that processes the chip, for example, the images BImg can be buffered in a memory with the test unit label TU as index. For ease of illustration, it is assumed in the present example that the image acquisition module 301 receives images BImg of a batch of chips, where the batch of chips comprises N groups of chips, and each group of chips contains M chips, thus the image acquisition module 301 buffers N x M images BImg.

[0017] The image conversion module 302 is configured to convert the image BImg of each chip in the N groups of chips into a blob distribution pattern, where a conversion technique known in the art can be employed, for example, performing binarization on each image BImg, and segmenting the binarized image to obtain foreground and background using image segmentation techniques, where the background refers to a standard reference image of the backside of a normal chip without any defects. Then performing region connectedness on the same or similar pixels in the foreground to form a blob (Blob), thus a blob distribution pattern BlobPatten containing one or more blobs can be generated based on each BImg image, for example, Fig. IB shows a blob distribution pattern converted from the backside image BImg of Fig. 1A.

[0018] The feature determination module 303 extracts blob features BF in the blob distribution pattern BlobPatten of each chip, where the blob features BF can include, for example, (1) size features of each blob, such as the number of pixels in the blob, the size (e.g. length, width, or area, etc.) of the blob; (2) the gray scale or pixel brightness of the blob, which reflects the contrast of the pixel relative to the background; (3) statistical features of the blob, such as the size ratio (e.g. aspect ratio) in different dimensions, the distance Range between the brightest and darkest pixels in the blob, and the variation Variation of the gray scale of the pixels in the blob, the difference in gray scale between the edge pixels and the center pixel in the blob, etc. Of course, the blob features are not limited to the above examples, and other features such as shape, direction vector, etc. can also be selected as needed.

[0019] According to an example of the present application, the feature determination module 303 is configured to determine which type or types of blob features BF to extract for the purpose of anomaly defect identification detection based on the feature attributes or types of the target defects to be detected. For example, for a solder ball defect such as a solder ball that is pressed and damaged on the surface of a chip, since the material of the solder ball is a bright metal, the damaged surface of the solder ball has a high contrast with respect to the back surface of the chip, and therefore the blob features BF can be extracted based on the gray scale of the blob, and obviously the solder ball can be more obviously distinguished from the background noise of the chip due to the high gray scale characteristics of the solder ball, so that the detection of such solder ball foreign matter can be better performed. For a surface indentation defect caused by a machine when operating a chip, the length, width, aspect ratio Ratio or Range of the blob can be used as the blob features to be processed. Therefore, according to an embodiment of the present application, the feature determination module 303 can extract different blob features based on the target defect types set or desired by the user. As an example, the embodiment of the present application is described here with respect to a solder ball damage defect.

[0020] It is assumed that there are m blobs on the blob distribution pattern BlobPatten of a chip unit, wherein at least one of the blobs contains the pixels of a damaged solder ball. The feature determination module 303 extracts the gray scale distribution value of each blob, and for the purpose of description, the gray scale distribution corresponding to each chip is represented as follows:

[0021]

[0022] wherein Xi represents the gray scale distribution of the i-th blob, i.e. the set of gray scale values {x} of the pixels located in the blob, and wherein x represents the gray scale value of each pixel in the i-th blob. In this example, the feature determination module 303 determines at least one blob feature of each blob based on the set of gray scale values {x} of the blob, and as an example, three gray scale-related features of the blob are determined, including:

[0023] (1) the average value of the gray scale to represent the average gray scale of the i-th blob, wherein wherein the function Avg represents the calculation of the average value of the data set {x}, and any function known in the art can be used, such as the arithmetic average, the weighted average, etc.

[0024] (2) the gray scale distribution Variation of the blob, which can be represented by calculating the variance of the gray scale distribution {x} with respect to the average value , wherein the variance VAR can be calculated by the following formula:

[0025] wherein, x ij represents the gray scale of the j-th pixel of the i-th blob, and n represents the number of pixels of the i-th blob.

[0026] (3) The gray scale difference DIF between each pixel at the edge of the light spot and the center pixel, which can be calculated, for example, by the following formula:

[0027] where x m represents the gray scale of a pixel at the i-th light spot edge, and x0represents the gray scale or average gray scale of the light spot center, and m represents the number of pixels at the i-th light spot edge.

[0028] Thus, the feature determination module 303 determines the three-dimensional features

[0029] In addition, the feature determination module 303 further extracts the position information POSi of the i-th light spot. It should be noted that the number of light spots of each chip is not necessarily the same as each other, which depends on the conversion processing result of the pixel conversion module 302. Here, for ease of description, it is still assumed that the light spot distribution of each chip contains m light spots. Thus, for N groups of chips, the feature determination module 303 determines the light spot gray scale features of N×M×m light spots to form a light spot data set, which is denoted by .

[0030] It should be noted that in the data set , each data point, i.e. light spot, is associated with the chip where the light spot is located, the position POS in the chip, and the group where the chip is located, so that the chip source of the data point can be determined based on the data point.

[0031] The anomaly identification module 304 processes the data set to determine whether there is an abnormal data point, i.e. whether there is an abnormal light spot in this light spot set, so as to determine whether there is, for example, a solder ball anomaly on the back of the chip. According to an example of the present application, the anomaly identification module 304 performs anomaly detection on the data set by using machine learning, such as the Isolation Forest algorithm. According to this embodiment of the present application, the Isolation Forest tree does not need to be trained in advance, but the data points in the data set are randomly node-split according to the dimensions VAR i ,DIF i according to the Isolation Forest algorithm, so as to determine the tree height h(x) of the N×M×m data points, and further determine the average height Eh(x) of each data point in all trees, so that each data point can be evaluated based on the average height E(h(x)), for example, the evaluation score s(x) of each node can be determined according to the following formula:

[0032] s(x)=2 -E(h(x))

[0033] Therefore, when the dataset A certain data point The evaluation score s(xi) exceeds a threshold, for example, greater than s. MAX Then, the data point can be determined. This is an outlier, thus identifying the outlier. The corresponding i-th spot is an abnormal spot. In this invention, the method for evaluating abnormal data points is not limited to the example above; other methods known in the prior art can be used, such as normalizing the data point height h(x) to determine anomalies using a threshold (0, 0.5, 1). If a data point score s is very close to 1, it can be considered an abnormal point; if a data point score s is much less than 0.5, it can be considered a normal point; and if all data point scores s≈0.5, then there may be no obvious abnormal points in any of the samples. Assuming that in this example, the anomaly identification module 305 determines the 100th data point... If it is an outlier data point, then it is based on the data generated in the dataset. The correlation mechanism can determine the specific spot location of the 100th data point. For example, if the spot at position POS4 in the third chip of the second group of chips may be abnormal, the anomaly identification module 305 can output an anomaly alarm signal to alert the user. It should be noted that in processing N groups of chips, anomalies may be found in more than one chip. Therefore, the anomaly handling module 305 can output an alarm signal indicating the detection of multiple anomalies.

[0034] In another implementation of the invention, the tree height can be pre-specified for the isolated forest algorithm, for example, tree height = 8, i.e., according to three dimensions. VAR i ,DIF i For dataset The data points within the dataset are randomly cut 8 times, and then the evaluation score s(x) of each node is evaluated. Following this example, the 'contamination' parameter of the Isolation Forest algorithm is set to the range of 0.001–0.002. Therefore, when the ratio R of the total number of evaluated outlier data points to the total number of data points is between 0.1% and 0.2%, it can be confirmed that there are outliers in the dataset. When R is less than 0.1%, it means that there is a possibility of random defects in the chip, which can be ignored. When R is greater than 0.2%, it means that there is a possibility of systematic defects in the chip. In this case, the chip is considered to be defective. It should be noted that the parameter value of contamination is selected as 0.001-0.002, which is meaningful and effective. It can avoid unnecessary alarms caused by random defects and ensure that the missing report of problematic systematic defects is not missed. Thus, by particularly introducing the Isolation Forest machine learning algorithm in the spot information processing, very significant abnormal spot detection effect can be achieved.

[0035] In addition, in the above examples, three features of the spot are extracted with gray scale as the reference VAR i , DIF i The above embodiments are described, but the present application is not limited thereto. For example, under the premise of taking gray scale as the reference, the features related to gray scale can be increased or decreased. For example, only can be extracted to represent each spot, or the difference Range between the brightest and darkest pixels in the spot can be further increased, that is, four-dimensional features are used to represent each spot. In addition, for indentation defects, for example, one-dimensional or more dimensional physical size features such as length, width, and area of the spot can be extracted to represent each spot. In other embodiments of the present application, for the detection of defects, the gray scale related features and the physical size features of the spot can be combined to represent each spot.

[0036] It should also be noted that due to pixel conversion errors, noise, or the discovered defects are only random defects, false alarms or unnecessary alarms for random defects may still occur in the above abnormal detection. Therefore, in another preferred embodiment of the present application, the anomaly recognition module 305 further discriminates the determined abnormal data points, so as to only alarm the regular defects on the chip. For ease of description, here represents a set of abnormal data points determined by the anomaly recognition module 305. The anomaly recognition module 305 uses an unsupervised machine learning clustering algorithm to cluster the data point set Clustering is performed to group similar data points into a single class, thereby eliminating false detections caused by errors or noise. According to one example of the invention, the anomaly detection module 305 employs the Density-Based Noise Applied Spatial Clustering (DBSCAN) algorithm to perform clustering. DBSCAN, as an unsupervised machine learning algorithm, does not require pre-specifying the number of clusters and is adaptable to light spot data point sets of different shapes and numbers. According to an embodiment of the invention, the anomalous light spot dataset obtained after isolated forest processing... The density is greatly reduced, thus improving the performance of anomalous spot datasets. The DBSCAN algorithm can effectively identify abnormal light spots. According to this embodiment, when executing the DBSCAN algorithm, the search radius parameter epsilon is set to a value between 100 and 300 μm, and the minimum number of points (minpts) in each cluster is not less than 3. In a preferred example, epsilon is set to 100 μm or 200 μm. According to another embodiment of the invention, the maximum value of epsilon does not exceed the minimum distance between chip pins, preferably not exceeding half of the minimum distance.

[0037] Therefore, when the DBSCAN algorithm identifies at least one cluster, such as C1, and discrete data points that do not meet the clustering conditions, the anomaly identification module 305 can identify the positional pattern LP of the data points within cluster C1. This pattern LP indicates the distribution position of the data points (i.e., the light spots) on the chip within cluster C1. The pattern LP indicates that the data points within cluster C1 are distributed around a fixed position. This confirms that the defect occurring at the position indicated by the pattern LP is a real defect, and since it occurs in multiple chips, it is a common defect, such as solder ball damage left on the surfaces of multiple chips processed in a batch. Meanwhile, for data points that fail to cluster, it indicates that the data points are discrete and random. Therefore, the anomaly identification module 305 excludes the possibility of anomalies represented by discrete data points.

[0038] According to another embodiment of the present invention, based on the cluster C1 determined by the anomaly identification module 305, it can be further analyzed whether there are common hardware or machine problems that cause the formation of the cluster, that is, cause the occurrence of common defects at the locations indicated by the location patterns LP in the cluster. Therefore, as... Figure 3As shown, the device 300 further comprises a traceability analysis module 305 configured to obtain grouping information TU of the chips where each abnormal data point, i.e. blob, in the cluster C1 is located, wherein when the grouping information of all or at least a portion of the abnormal blobs indicates the same group, it indicates that there is a common fault in the test unit processing the chips in the group. For example, when the grouping information of a portion or all of the blobs in the cluster C1 is TU3, the traceability analysis module 305 can determine that there is a common fault in the No. 3 test unit for processing the No. 3 group of chips; meanwhile, when the grouping information of another portion of the blobs in the cluster C1 is TU7, the traceability analysis module 305 can determine that there is a common fault in the No. 7 test unit for processing the No. 7 group of chips. Thus, the traceability analysis module 305 determines that the No. 3 and No. 7 test units are faulty, which can be the source of the target defect, e.g. the solder ball. Therefore, the traceability analysis module 305 can form an alarm signal indicating that there is a common fault in the No. 3 and No. 7 test units of the test machine, thereby prompting the machine operator to check the No. 3 and No. 7 machines and eliminate the fault.

[0039] In the above embodiment, the solder ball damage defect is taken as an example and the gray scale correlation feature of the blob is determined as the feature representing the blob image, which has the advantage that the gray scale feature can well distinguish the solder ball from the background. In order to better highlight the difference between the defect and the normal background when performing anomaly detection, according to the present application, the detection system can be configured with multiple light sources 100, such as an oblique lighting system OLS, a coaxial lighting system CLS or an oblique and coaxial combined lighting system OCLS, etc., and the anomaly detection device 300 further comprises a light source setting module 306, such as Figure 3As shown, the light source setting module 306 is configured to select a proper light source based on the characteristic attributes of the target defect to be detected so as to enhance the feature difference of the target defect in the acquired optical image BImg with respect to the target unit as background. For example, for the solder ball damage defect that can exist on the chip, since it can reflect more light with respect to the chip surface, i.e. higher brightness, the light source setting module 306 can select the oblique coaxial combined illumination system OCLS as the irradiation light source 100, and irradiate the chip surface by activating the oblique coaxial combined illumination system, so that the sensor 200 can take a brighter image to suppress background noise. For the indentation defect caused by the machine, for example, since it is a physical structure damage to the chip surface, the oblique illumination system OLS or the coaxial illumination system CLS can be selected as the light source 100 to irradiate the chip. In another embodiment of the present application, the light source can also be adapted to different defect types by transforming the light intensity or spectrum of the same light source. For example, for the solder ball defect, the light source 100 can be made to emit light with higher light intensity to irradiate the chip surface; while for the defect with physical indentation such as indentation, the light source setting module 306 can control the light source 100 to emit light with lower light intensity, so as to better highlight the indentation area. Thus, the feature selection module 303 can receive the light source setting information of the light source setting module 306, and extract the light spot features BF accordingly.

[0040] In the above example, the abnormality detection is performed by analyzing the gray scale related features of the light spot, but obviously the present application is not limited thereto, and the detection of abnormal light spot can also be realized by combining the analysis of multiple features of the light spot such as the aspect ratio Ratio, the distance Range and the gray scale variation distribution of the pixels within the light spot. For example, the indentation detection, since it is caused by the hardware failure of the test unit TU, the indentation has a basically constant feature, and thus the indentation detection can be performed by combining the analysis of the light spot area and the gray scale distribution pattern of the light spot.

[0041] Figure 4 A flow chart of the defect detection method performed by the detection device 300 according to one example of the present application is shown. As shown, in step 401, the detection device 300 receives and buffers the optical images BImg of multiple groups of chips processed by multiple test units TU of the test machine from the sensor 200. According to one example of the present application, each chip image BImg buffered is associated with the test unit TU that processes the chip. As an example, the detection device 300 buffers NXM images BImg in total, where N corresponds to the number of test units TU, and M represents the number of chips processed by each test unit in this batch of detection.

[0042] In step 403, each of the N x M images BImg is converted into a blob pattern, for example using a binary image processing technique to form a blob pattern BlobPatten containing one or more blobs. Subsequently in step 405, based on the characteristic attributes or types of the target defects to be detected, one or several types of blob features BF are determined to be extracted for the abnormal defect recognition detection. According to the present application, the blob features BF that can be used include but are not limited to: (1) size features such as the number of pixels, the size of the blob, etc. of each blob; (2) optical features such as the gray scale or pixel brightness of the blob; and (3) statistical features of the blob such as the size ratio Ratio in different dimensions, the distance Range between the brightest and darkest pixels in the blob, the variation Variation of the pixel gray scale within the blob, the gray scale distribution feature of the edge of the blob, etc. As an example, when it is desired to detect whether there is a solder ball damage on the chip, in step 405, based on the blob gray scale values, at least three blob features BF of each blob are calculated Here represents the average gray scale of the i-th blob, VAR i represents the variance of the blob distribution of the i-th blob, and DIF i represents the pixel gray scale distribution of the edge of the i-th blob. In addition, in step 405, the position information POS of each blob is further extracted.

[0043] In step 407, based on the three-dimensional features BF of each blob calculated A blob data set is established where each data point represents a blob, and is represented by In the data set , each data point is associated with the chip where the blob is located, the position POS within the chip, and the group TU where the chip is located.

[0044] In step 409, the data set is processed using unsupervised machine learning to determine whether there is an abnormal blob in the blob set, so that it can be determined whether there is an abnormal solder ball on the back of the chip. According to an example of the present application, the Isolation Forest technique is used to perform anomaly detection on the data set . According to the Isolation Forest algorithm, the data points in the data set are randomly split according to the dimensions VAR i , DIF i , so as to determine the tree height h(x) of each data point N x M x m to determine whether there is an abnormal data point. For example, when the i-th data point in the data set is an abnormal data point, the tree height h(x) of the i-th data point is smaller than the tree height h(x) of the other data points. s(xi) exceeds a threshold, for example, greater than s MAX , it can be determined that the data point is an outlier, thereby determining that the i-th spot represented by is an abnormal spot. Due to the correlation mechanism established when the data set is generated, the specific spot position of the spot can be determined using the identified outlier, and thus an abnormal alarm signal can be output to prompt the user of the abnormality. In another implementation of the present application, the height of the tree can also be specified in advance for the isolation forest algorithm, for example, tree height = 8, i.e., according to three dimensions VAR i , DIF i , the data points in the data set are randomly cut 8 times, and then the evaluation score s(x) of each node is evaluated. According to the present example, the value range of the 'contamination' parameter of the isolation forest algorithm is set to (0.001, 0.002), i.e., greater than or equal to 0.001 and less than or equal to 0.002. Thus, when the ratio R of the total number of data points evaluated as abnormal data points to the total number of all data points is between 0.1% and 0.2%, it is determined that there are abnormal data points in the data set . When R exceeds 0.1% and 0.2% (whether less than or greater than this range), it is considered that there is no defect in the chip to be processed, and thus no alarm signal is generated.

[0045] According to another preferred example of the present application, the process can further identify the abnormal data points identified in step 409. To this end, as shown in Figure 4 , in step 411, a set of abnormal data points is established, where the elements come from the abnormal data points determined in step 409. Subsequently, in step 413, an unsupervised machine learning clustering algorithm is used to cluster the data point set The clustering processing is performed to group similar data points into a cluster so as to exclude false detections caused by errors or noises. According to an example of the present application, the density-based spatial clustering of applications with noise (DBSCAN) algorithm can be employed to perform the clustering processing, and particularly, the search radius epsilon is set to a range of 100-300 um, and the minimum number of points in each cluster minpts is not less than 3. Thus, when it is determined in step 413 that there is at least one cluster, for example, Cl, then step 415 is entered to further identify the location pattern LP of the data points within the cluster Cl, i.e., the distribution of the data points, i.e., the blobs, on the chip within the cluster Cl, which indicates that the data points within the cluster Cl are distributed around a fixed location, and thus it can be determined that the defects occurring at the location indicated by the location pattern LP are real common defects, for example, leaving solder ball damages on the surface of a plurality of chips processed in batch. Then, step 417 is entered to exclude the abnormal blobs represented by the discrete data points located outside the cluster, for example, Cl, from being abnormal. If no successful clustering is found in step 413, i.e., the data points in the data point set are discrete and do not satisfy the clustering condition, it indicates that the abnormal blobs represented by the data points in the data point set are accidental, and thus in step 417, the abnormal blobs represented by the abnormal data points in the data point set are excluded from being abnormal, and the abnormal detection for the batch of chips is ended.

[0046] According to another embodiment of the present application, in step 417, a defect root cause step can be further performed to determine whether there is a hardware or tester common problem causing the defects in the cluster, i.e., causing the common defects at the location indicated by the location pattern LP. To this end, in step 417, the grouping information TU of the chips on which the abnormal data points, i.e., the blobs, in the cluster Cl are located is further obtained, wherein when the grouping information of all or at least a part of the abnormal data points indicates the same group, it indicates that there is a common failure in the test unit processing the group of chips. For example, when the grouping information of a part of the blobs in the cluster Cl is TU3, it is determined that there is a common failure in the 3rd test unit processing the 3rd group of chips; and when the grouping information of another part of the blobs in the cluster Cl is TU7, it is determined that there is a common failure in the 7th test unit processing the 7th group of chips. Thus, an alarm signal indicating that the 3rd and 7th test units of the tester have common failures is output in step 417.

[0047] In the above embodiments, the chip is taken as the target unit to explain the concept and scheme of the present application, but the present application is not limited to this, and can be applied to any other individual unit defect detection to improve the early detection accuracy of defects. According to the detection scheme of the present application, not only the detection accuracy requirement can be ensured, but also the processing speed can be improved while reducing the complexity of the system.

[0048] In addition, although the exemplary method of the present application is described above in combination with specific examples, it can be understood that the units, modules and method steps described herein can be implemented as electronic hardware, computer software or a combination of both. In addition, another embodiment of the present application provides a machine readable medium having stored machine readable instructions, which, when executed by a processor, causes the processor to perform any one of the aforementioned methods disclosed herein. In addition, the present application has been described in detail above through the accompanying drawings and preferred embodiments, however, the present application is not limited to these disclosed embodiments, and based on the above multiple embodiments, those skilled in the art can know that more embodiments of the present application can be obtained by combining the above different embodiments, and these embodiments are also within the protection scope of the present application.

Claims

1. A defect detection method, comprising: Acquire optical images of multiple target units; The optical image of each target unit in the multiple groups of target units is converted into a light spot distribution map; Based on the characteristic attributes of the target defects, one or more different types of spot features are extracted from each spot in the spot distribution map to establish the spot feature set of the multiple target units; By using unsupervised machine learning to process the spot feature sets of the multiple groups of target units, it is determined whether the spot distribution map contains abnormal spots, thereby judging whether the target unit has a target defect.

2. The defect detection method as described in claim 1, further comprising: The unsupervised machine learning is used to perform clustering processing on the abnormal light spots of the multiple groups of target units to determine whether the abnormal light spots have fixed positional pattern features; When it is determined that the abnormal light spot has a fixed positional pattern characteristic, it is determined that the target unit has the target defect at the position indicated by the positional pattern.

3. The defect detection method as described in claim 2, wherein the extracted spot features include at least three features to establish at least a three-dimensional feature representation for each spot; The unsupervised machine learning process for processing the spot feature set includes: The isolated forest algorithm is applied to the spot feature set represented by the at least three-dimensional features to determine the abnormal spot dataset, wherein the contamination parameter used to determine abnormal spots in the isolated forest algorithm has a value range of 0.1% to 0.2%; The clustering process for the abnormal light spots of the multiple target units includes applying the density-based noise spatial clustering (DBSCAN) algorithm to the abnormal light spot dataset. In the DBSCAN algorithm, the search radius epsilon is between 100 and 300 μm, and the minimum number of points (minpts) is not less than 3.

4. The defect detection method according to any one of claims 1-3, wherein the light spot feature is extracted based on the characteristic attributes of the target defect and the characteristics of the light source, and wherein the optical image is captured under the illumination of the light source; The characteristic attributes of the target defect include: The optical characteristics of the target defect differ from those of the target unit in the background. The physical morphological characteristics of the target defect; The light spot characteristics include: (1) Length and width; (2) Average gray value; (3) Aspect ratio; (4) The distance between the brightest and darkest pixels in the light spot; (5) The grayscale difference between the edge and center of the light spot; and (6) Gray-scale changes of each pixel in the light spot.

5. The defect detection method according to any one of claims 1-3, wherein each of the plurality of target units is operated by one of the plurality of test units in the test machine. The defect detection method further includes: When it is determined that the target unit has the target defect, the correlation between the target unit with the target defect and the test unit that operates the target unit is determined; When different target units with the target defect are all associated with the same test unit, it is further determined that the same test unit has a hardware fault that causes the target defect.

6. The defect detection method as described in claim 5, further comprising: Based on the characteristic attributes of the target defect to be detected, one of a number of light sources is selected or the luminescence characteristics of the light source are adjusted to enhance the characteristic difference of the target defect relative to the target unit as the background in the acquired optical image.

7. A defect detection device, comprising: The image acquisition module is configured to acquire optical images of multiple target units; The image conversion module is configured to convert the optical image of each target unit in the plurality of target units into a light spot distribution map; The feature determination module is configured to extract one or more different types of spot features for each spot in the spot distribution map based on the feature attributes of the target defect to be detected, so as to establish the spot feature set of the multiple target units; The anomaly detection module is configured to determine whether the target unit has a target defect by using unsupervised machine learning to process the spot feature set of the multiple groups of target units to determine whether the spot distribution map contains abnormal spots.

8. The defect detection device as described in claim 7, wherein the anomaly identification module is further configured as follows: The unsupervised machine learning is used to perform clustering processing on the abnormal light spots of the multiple groups of target units to determine whether the abnormal light spots have fixed positional pattern features; When it is determined that the abnormal light spot has a fixed positional pattern characteristic, it is determined that the target unit has the target defect at the position indicated by the positional pattern.

9. The defect detection device of claim 8, wherein the extracted spot features include at least three features to establish at least a three-dimensional feature representation of each spot; The anomaly detection module is further configured as follows: The isolated forest algorithm is applied to the spot feature set represented by the at least three-dimensional features to determine the abnormal spot dataset, wherein the 'contamination' parameter used to determine abnormal spots in the isolated forest algorithm ranges from 0.1% to 0.2%; and The density-based noise spatial clustering (DBSCAN) algorithm is applied to the abnormal spot dataset, wherein the search radius epsilon in the DBSCAN algorithm is between 100 and 300 μm, and the minimum number of points minpts is not less than 3.

10. The defect detection device according to any one of claims 7-9, wherein the spot feature is extracted based on the characteristic attributes of the target defect and the characteristics of the light source, wherein the optical image is captured under the illumination of the light source; The characteristic attributes of the target defect include: The optical characteristics of the target defect differ from those of the target unit in the background. The physical morphological characteristics of the target defect; The light spot feature includes at least one of the following: (1) Length and width; (2) Gray value; and (3) Aspect ratio; (4) The distance between the brightest and darkest pixels in the light spot; (5) The grayscale difference between the edge and center of the light spot; (6) Gray-scale changes of each pixel in the light spot.

11. The defect detection device as described in any one of claims 7-9, wherein each of the plurality of target units is operated by one of the plurality of test units in the machine. The defect detection device further includes a source tracing analysis module, configured as follows: When it is determined that the target unit has the target defect, the correlation between the target unit with the target defect and the test unit that operates the target unit is determined; When different target units with the target defect are all associated with the same test unit, it is further determined that the same test unit has a hardware fault that causes the target defect.

12. The defect detection device as claimed in claim 11, further comprising: The light source setting module is configured to select one of multiple light sources or adjust the luminescence characteristics of a light source based on the feature attributes of the target defect to be detected, so as to enhance the feature difference of the target defect in the acquired optical image relative to the target unit as the background.

13. A machine-readable storage medium storing a machine-readable program, which, when executed by a machine, causes the machine to perform the defect detection method according to any one of claims 1-6.

Images