Wafer slip sheet detection method

Abstract

The invention provides a wafer slip sheet detection method, which comprises a laser emitter, an industrial camera and a light supplement lamp which are arranged on a machine table used by a wafer CMP process, and comprises the following steps: when the light supplement lamp is turned on, a visual identification system extracts a picture a from real-time video data sent by the industrial camera according to a set frame rate; a YOLOv8 model is used for processing, and whether the wafer slides out or not is judged based on output of the YOLOv8 model; when the light supplementing lamp is not turned on, the visual identification system extracts a picture b from real-time video data sent by the industrial camera according to a set frame rate, gray processing is carried out on the picture b to generate a gray-scale map, a communication area is searched on the gray-scale map, the searched communication area is verified to find a laser point area, and the gray value of the laser point area is calculated; and when the gray value of the laser point area is lower than the threshold value, determining that the laser point disappears, and sliding out the wafer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor chemical mechanical polishing, and particularly relates to a wafer slip detection method. BACKGROUND

[0002] Chemical mechanical polishing (CMP) is an important process for realizing in-plane planarization of semiconductor wafers. In the process of chemical mechanical polishing, the polishing head is buckled on the wafer, and the wafer is pressed on the polishing pad of the polishing disc. The polishing disc and the polishing head rotate at the same time to complete the polishing of the wafer. However, when the machine appears a problem, the wafer may be thrown out from below the polishing head during the rotation of the polishing disc and the polishing head. If the wafer slip is not found in time, and the polishing head and the polishing disc continue to rotate, the wafer may be broken.

[0003] In the prior art, in order to detect whether the wafer under the polishing head slips out, a laser emitter is arranged on the machine opposite to the polishing head. The laser emitted by the laser emitter is irradiated to the polishing disc of the wafer. The laser emitter can be arranged in multiple numbers, and the emitted laser can be on the wafer slip path. When the wafer slips out, the diffuse reflection of the laser becomes specular reflection, and the light intensity of the receiver suddenly becomes large. When the light intensity received by the receiving end exceeds a threshold value, it is determined that the wafer slips out, and the machine alarms.

[0004] However, in actual use, the back surface of the wafer is stained with polishing liquid, which may cause the intensity of laser reflection to be weak, or the wafer is a non-double-polished wafer, which may also cause the intensity of laser reflection to be weak. Therefore, the intensity of laser reflection when the wafer slips out cannot be obviously distinguished from the intensity of laser reflection on the polishing pad. Due to the interference of the above two working conditions, it is difficult to set a suitable laser reflection light intensity threshold value. If the threshold value is set too large, the wafer with polishing liquid on the back surface or the non-double-polished wafer may not alarm when it slips out. If the threshold value is set too small, the wafer may alarm even if it does not slip out. Therefore, the same threshold value may cause the situation that the wafer slips out without alarm and the wafer slips out with alarm at the same time, which may cause the set laser reflection light intensity threshold value to be invalid, the system cannot accurately find that the wafer slips out, and the problem of misjudgment is easily caused, thereby affecting the production of wafers.

[0005] Therefore, it is necessary to redesign a method for detecting whether the wafer slips out in the wafer CMP process, so as to solve the problem that the wafer slip detection method in the prior art is prone to misjudgment. SUMMARY

[0006] The present application provides a wafer slip detection method, which can more accurately detect whether the wafer slips out by assisting the judgment of wafer slip through two detection modes.

[0007] Other objects and advantages of the present application can be further understood from the technical features disclosed in the present application.

[0008] To achieve one or part or all of the above purposes or other purposes, a wafer slip detection method is provided in a technical solution of the present application. The method comprises a laser emitter, an industrial camera and a fill light lamp installed on a machine table used in a wafer CMP process. The laser emitted by the laser emitter is directed towards a polishing head position. The industrial camera photographs a polishing head and polishing pad adhesion position. According to whether the fill light lamp is turned on or not, a visual recognition system is used for processing and identifying whether the wafer slips out or not. When the fill light lamp is turned on, the visual recognition system extracts pictures from real-time videos sent by the industrial camera and processes them using a YOLOv8 model to output position box information of the polishing head and the wafer, and determines whether the wafer slips out or not. When the fill light lamp is not turned on, the visual recognition system extracts pictures from real-time videos sent by the industrial camera and generates a gray-scale image. The laser point area is found on the gray-scale image and the gray-scale value is calculated. According to the relationship between the gray-scale value of the laser point area and the set threshold, it is determined whether the wafer slips out or not.

[0009] When the fill light lamp is turned on, the visual recognition system extracts pictures a from real-time videos sent by the industrial camera at a set frame rate for preprocessing, removes noise in the pictures a and improves the contrast of the pictures a, and processes the preprocessed pictures a using the YOLOv8 model. The YOLOv8 model is trained by inputting a large number of wafer slip-out photos and non-slip-out photos. When the fill light lamp is not turned on, the visual recognition system extracts pictures b from real-time videos sent by the industrial camera at a set frame rate for gray-scale processing to generate a gray-scale image. The laser point area is found on the gray-scale image, and the centroid position is extracted. The found connected area is verified to find the laser point area.

[0010] The preprocessing of the pictures a includes multi-level noise reduction of the extracted pictures a. The median filter method is used to remove the salt and pepper noise in the pictures a, and the salt and pepper noise is the picture noise caused by liquid splashing. The bilateral filter method is used to remove the texture noise in the pictures, and the texture noise is the picture noise formed by the polishing pad.

[0011] The pictures a after denoising processing are subjected to Fourier transform to convert the brightness channel of the pictures a to the frequency domain. The periodic noise is expressed as a sharp peak of a specific frequency, and a band-stop filter is used to shield the high-frequency area noise corresponding to the periodic noise and the low-frequency noise within the radius set in the low-frequency area. The frequency domain signal after filtering processing is converted back to the spatial domain to obtain the denoised brightness channel. The HSV image corresponding to the denoised brightness channel is converted back to the BGR color space to output the final image.

[0012] The extracted picture a is improved in contrast, including converting the extracted picture a to an LAB color space, configuring a contrast limit threshold and a local region division number of a CLAHE algorithm, and dividing the extracted picture a into a plurality of local regions according to the local region division number for contrast improvement processing; and applying the CLAHE algorithm to perform contrast enhancement processing on the extracted picture a based on the set contrast limit threshold.

[0013] If the polishing pad is white, the contrast limit threshold is set to 2-2.5; if the polishing pad is black, the contrast limit threshold is set to 3-4; and the local region division number is 64-256, and the extracted picture a is divided into a plurality of local regions in the length and width directions.

[0014] The YOLOv8 model includes an input end, a backbone network, a neck network, a detection head, and an output end. The backbone network is a convolutional neural network model, used for extraction of wafer edge texture features, polishing head texture features, and polishing head global features. The extracted wafer edge texture features are subjected to multi-scale feature fusion and multi-scale output by the neck network of the YOLOv8 model. The P1 level is used for detecting wafer edge detail features, the P2 level is used for detecting wafer main body features, and the P3 level is used for detecting complete wafer position features. The detection head predicts target positions and categories based on three-scale output features of the neck network. The output end outputs bounding box position information and confidence. The extracted polishing head texture features and polishing head global features are subjected to feature fusion by the neck network of the YOLOv8 model. FPN network combined with PAN network is used for feature fusion. The detection head predicts target positions and categories based on the fused polishing head features. The output end outputs bounding box position information and confidence. Based on the position box information of the polishing head and the position box information of the wafer, it is determined whether the wafer slides out.

[0015] The neck network adopts a bidirectional feature pyramid structure, and performs upsampling, downsampling, and splicing operations on the output features. The P1 level is used for small target feature detection, the P2 level is used for medium target feature detection, and the P3 level is used for large target feature detection.

[0016] Based on the position box information of the polishing head and the position box information of the wafer, it is determined whether the wafer slides out, including converting the bounding box parameters into absolute coordinates, and determining whether the wafer exceeds the edge of the polishing head based on the absolute coordinates. If there are 4 frames in which the wafer exceeds the edge of the polishing head in the last 5 frames, it is determined that the wafer slides out.

[0017] The found connected region is verified to be a laser point region, including size verification, gray value verification, circularity verification, and intensity distribution verification. The intensity distribution verification is whether the gray value at the center point is the highest.

[0018] The PLC system of the machine is connected to the vision system through Ethernet, the algorithm module of the vision system is deployed on the edge computer device, when the light supplement lamp of the machine is turned on, the real-time data generated in the PLC light supplement lamp control register is read by the edge computer, and whether the machine turns on the light supplement lamp is determined according to the real-time data in the register to select different processing modes.

[0019] Compared with the prior art, the beneficial effects of the present application mainly include that the present application detects whether the wafer slides out based on the vision recognition system, reduces the shortcomings of laser detection in the prior art, and improves the accuracy of detection. At the same time, the vision recognition system of the present application identifies whether the wafer slides out through two different identification modes: in the case of allowing the light supplement lamp to be turned on, the light supplement lamp cooperates with the industrial camera to shoot real-time video and the real-time video is processed for visual recognition to determine whether the wafer slides out; in the case of not suitable for turning on the light supplement lamp, whether the red dot of the laser emitter on the grinding head disappears is identified by processing the real-time video shot by the industrial camera, and the wafer is determined to slide out if the red dot disappears. The present application can adapt to different production process requirements based on two different identification modes, the detection is more flexible and the detection success rate is high, and the misjudgment rate is reduced.

[0020] In order to make the above and other objects, features and advantages of the present application more apparent, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0022] Figure 1 The schematic diagram of the position of the laser emitter installed on the machine and the laser spot on the polishing pad of the present application.

[0023] Figure 2 The schematic diagram of the neck network of the YOLOv8 model of the present application. DETAILED DESCRIPTION

[0024] The foregoing and other technical contents, features and effects of the present application will be clearly presented in the following detailed description of a preferred embodiment with reference to the drawings. The directional terms mentioned in the following embodiments, such as up, down, left, right, front or back, etc., are only the directions of the drawings. Therefore, the directional terms used are used for illustration, not for limiting the present application.

[0025] Embodiment one

[0026] The wafer slip detection method provided by the embodiment one comprises a laser emitter, an industrial camera and a light supplement lamp installed on a machine table used in a wafer CMP process, the laser emitted by the laser emitter is directed to a position of a polishing head, and the laser emitted by the laser emitter forms a laser spot on a polishing pad, the laser spot is located in a slip direction of wafer slip, and a plurality of laser emitters can be arranged to adapt to different slip directions of the wafer. The industrial camera photographs a position where the polishing head and the polishing pad are attached, a visual recognition system is used to detect whether the wafer slips, and the visual recognition system selects a corresponding processing mode according to whether the light supplement lamp is turned on.

[0027] When the light supplement lamp is turned on, the visual recognition system extracts pictures from real-time video data sent by the industrial camera and processes them using a YOLOv8 model (You Only Look Once Version 8, released in 2023, the core features include high efficiency, light weight and dynamic sample allocation strategy, which generally includes a backbone network, a feature enhancement network and a detection head) to output position box information of the polishing head and the wafer, and analyze position information of the polishing head and the wafer according to the position box information of the polishing head and the wafer, and then determine whether the wafer slips.

[0028] When the light supplement lamp is not turned on, the visual recognition system extracts pictures from real-time video data sent by the industrial camera and generates a gray scale image, finds a laser spot area on the gray scale image and calculates a gray scale value, determines whether the wafer slips according to a relationship between the gray scale value of the laser spot area and a set threshold.

[0029] The following will be explained in detail with reference to the accompanying drawings. Figure 1 and the accompanying drawings. Figure 2 The present application will be explained in detail.

[0030] The wafer slip detection method of the present application comprises arranging a laser emitter, an industrial camera and a light supplement lamp on a machine table used in a wafer CMP process, the position of the laser emitter can be seen from Figure 1 which is located above the machine table, the laser emitted by the laser emitter is directed to a position of a polishing head, and forms a laser spot on a polishing pad, the laser spot is a circular red dot, the industrial camera and the light supplement lamp can be arranged around the laser emitter, the industrial camera is used to photograph a position where the polishing head and the polishing pad are attached

[0031] In order to test whether the wafer slides out, an industrial camera is used to shoot real-time video, and a visual recognition processing technology is used to identify whether the wafer slides out. In order to improve the visual recognition effect, the visual recognition system in embodiment one uses two recognition modes. One is to turn on the light compensation lamp, and the visual recognition system processes the real-time video to find out whether the wafer slides out. The other is to identify whether the laser point disappears in the real-time video when the light compensation lamp is not turned on, and then determine whether the wafer slides out.

[0032] The light compensation lamp is turned on under the premise that the wafer production process meets the demand. The light compensation lamp cooperates with the industrial camera to make the real-time video shot by the industrial camera clearer.

[0033] The light compensation lamp cannot be turned on during the CMP process of some wafers, so the wafer processing environment is relatively dark, and the video shot by the industrial camera is relatively blurred. In addition, the color of the wafer is also dark. If the color of the polishing pad is also dark at this time, the visual recognition effect is poor, and it is difficult to identify whether the wafer slides out. Therefore, the visual recognition system in embodiment one monitors whether the laser red dot on the polishing pad disappears in the real-time video image, and then determines whether the wafer slides out. When the wafer slides out, the red laser dot shines on the wafer, and the laser dot will suddenly disappear due to total reflection. Therefore, whether the wafer slides out can be determined according to whether the laser dot disappears.

[0034] The following explains in detail the two wafer processing modes.

[0035] When the light compensation lamp is turned on:

[0036] The light compensation lamp is generally yellow. The visual recognition system processes the real-time video data sent by the industrial camera. The video resolution shot by the industrial camera is not less than 720P, and the frame rate is not less than 30fps. The industrial camera can use a global shutter CMOS sensor, supports trigger acquisition, and can effectively capture high-speed moving scenes.

[0037] Among them, the PLC system of the machine, the industrial camera and the visual system are connected through Ethernet, and the algorithm module of the visual system is deployed on the edge computer device to run the visual recognition algorithm in real time. In order to ensure the efficiency of communication, a real-time communication interface is used to establish communication between the PLC system, the industrial camera and the edge computer device.

[0038] When the machine turns on the light compensation lamp, the real-time data generated in the PLC light compensation lamp control register is read by the edge computer, and whether the machine turns on the light compensation lamp is determined according to the real-time data in the register to select different processing modes.

[0039] After the visual recognition system receives the video sent by the industrial camera, the real-time video data sent is preprocessed according to a set frame rate to extract a picture a, wherein the preprocessing process includes removing noise in the picture a and improving the contrast of the picture a.

[0040] The noise removal in the picture a includes multi-stage noise removal of the picture a: the salt and pepper noise in the extracted picture a is removed by using a median filtering method, and the salt and pepper noise is the noise generated by the grinding liquid splashing.

[0041] The texture noise in the picture is removed by using a bilateral filtering method, and the texture noise is the picture noise formed by the polishing pad itself.

[0042] The picture a after the hierarchical noise removal processing is subjected to Fourier transform, specifically, the luminance channel of the picture a is converted to the frequency domain, at this time the periodic noise is expressed as a peak of a specific frequency, a band-stop filter is used to shield the high-frequency noise corresponding to the periodic noise and the low-frequency noise within a radius in the low-frequency region, the frequency domain signal after the filtering processing is converted back to the spatial domain to obtain the de-noised luminance channel, the HSV image (Hue, Saturation, Value, a color space created according to the intuitive characteristics of color, also known as a hexagonal cone model) corresponding to the de-noised luminance channel is converted back to the BGR color space (the color space used by default by OpenCV, which is a variant of RGB, in the RGB color space, colors are arranged in the order of red, green and blue, while in BGR, colors are arranged in the order of blue, green and red), and the final image is output.

[0043] The contrast of the extracted picture a is improved, including converting the extracted picture a (in BGR format) to the LAB color space (the Lab color model is composed of three elements, one element is luminance L, and a and b are two color channels. The color included in a is from dark green (low brightness value) to gray (medium brightness value) to bright pink (high brightness value); b is from bright blue (low brightness value) to gray (medium brightness value) to yellow (high brightness value), and such color mixing will produce a color with bright effect), configuring the contrast limit threshold of the CLAHE algorithm (Contrast limited adaptive histogram equalization, an image enhancement algorithm, using a release histogram equalization processing method, the originally randomly distributed pixel values are redistributed by accumulating the histogram, so that the pixel values are uniformly distributed in all available value range) and the number of local region division, and the extracted picture a is divided into a plurality of local regions according to the number of local region division for contrast improvement processing; the extracted picture a is subjected to contrast enhancement processing based on the set contrast limit threshold by applying the CLAHE algorithm, and the specific processing process of applying the CLAHE algorithm is an existing technology, which will not be described herein.

[0044] If the polishing pad is white, the set contrast limit threshold is 2-2.5; if the polishing pad is black, the set contrast limit threshold is 3-4; the number of local area divisions is 64-256, and the extracted picture a is divided into equal parts in the length and width directions, i.e. 8*8 to 16*16, wherein 8*8 means that the length and width directions are divided into 8 equal parts.

[0045] The preprocessed picture a is processed using a YOLOv8 model, the YOLOv8 model is trained by inputting a large number of wafer slip-out photos and non-slip-out photos, and after training, a YOLOv8 model capable of realizing wafer slip-out judgment can be obtained. The picture a is input to the trained YOLOv8 model for processing, and whether the wafer slips out is determined based on the output of the YOLOv8 model.

[0046] The YOLOv8 model in embodiment one includes an input end, a backbone network, a neck network, a detection head, and an output end. The backbone network is a convolutional neural network model, which includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The backbone network is used for extracting wafer edge texture features, grinding head texture features, and grinding head global features. The backbone network belongs to a conventional technology, and will not be described in detail. The extracted wafer edge texture features are subjected to multi-scale feature fusion and multi-scale output by the neck network of the YOLOv8 model. For the neck network structure for processing wafer edge texture features, refer to Figure 2 The neck network adopts a bidirectional feature pyramid structure, performs upsampling, downsampling, and splicing operations on the output features, the P1 level is used for small target feature detection, the P2 level is used for medium target feature detection, and the P3 level is used for large target feature detection. For example, the P1, P2, and P3 levels respectively divide the wafer edge texture features output by the backbone network into the following number of perception grids: 80*80, 40*40, and 20*20. After the above division, the 80*80 division method obtains smaller features, which can perceive small targets, such as wafer edge texture. The 20*20 division method can obtain large targets, such as the overall appearance of a wafer. Of course, the number of perception grids of the P1, P2, and P3 levels can also be redesigned according to actual needs during actual design.

[0047] When the texture features of the grinding head and the global features are processed, the extracted grinding head texture features and grinding head global features are fused by a neck network of a YOLOv8 model, and the features are fused by a FPN network (Feature Pyramid Network) combined with a PAN network (Path Aggregation Network), and the target position and category are predicted by a detection head based on the fused grinding head features, and the output end outputs the bounding box position information and the confidence; based on the position box information of the grinding head and the position box information of the wafer, it is determined whether the wafer slides out.

[0048] When it is determined whether the wafer slides out based on the position box information of the grinding head and the position box information of the wafer, the calculated position box information of the grinding head and the position box information of the wafer are converted into absolute coordinates, and it is determined whether the wafer exceeds the edge of the grinding head based on the absolute coordinates; if the wafer exceeds the edge of the grinding head for 4 frames in 5 consecutive frames, it is determined that the wafer slides out. The corresponding edge computing device sends alarm information to the PLC control system to perform a shutdown operation or an alarm operation.

[0049] When the fill light is not turned on:

[0050] The visual recognition system processes the real-time video data sent by the industrial camera, including extracting pictures b for grayscale processing according to the set frame rate. Since the processing is directly performed on the color image, the processing effect is not good. In embodiment one, the obtained pictures b are generated into grayscale images, and connected regions (the grayscale values of multiple regions are the same or close, and the threshold value of the close grayscale values can be set according to the actual situation) are found on the grayscale images, and the centroid positions (which can be used for positioning the connected regions) are extracted. The found connected regions are verified to find the laser point region, and the grayscale value of the laser point region is calculated. When the grayscale value of the laser point region is lower than the threshold value, it is determined that the laser point disappears and the wafer slides out.

[0051] When the found connected regions are verified to find the laser point region, size verification (laser point size), grayscale value verification (whether the grayscale mean value meets the requirements), circularity verification, and intensity distribution verification can be performed. The intensity distribution verification is to verify whether the grayscale value at the center point is the highest.

[0052] The above describes in detail a wafer slip detection method provided by the present application. In this paper, specific examples are used to describe the structure and working principle of the present application. The above examples are only used to help understand the method and core idea of the present application. It should be noted that for ordinary skilled persons in the technical field, without departing from the principles of the present application, the present application can be improved and modified in several ways, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A wafer slide detection method, comprising a laser transmitter, an industrial camera, and a fill light installed on a machine used in a wafer CMP process, wherein the laser transmitter emits laser light toward the position of the polishing head, and the industrial camera captures the position of the polishing head and the polishing pad. The method is characterized in that: Depending on whether the fill light is on, the visual recognition system is used to process and identify whether the wafer has slipped out; When the fill light is turned on, the visual recognition system extracts images from the real-time video sent by the industrial camera and processes them using the YOLOv8 model to output the position frame information of the grinding head and the wafer to determine whether the wafer has slipped out; When the fill light is not turned on, the visual recognition system extracts images from the real-time video sent by the industrial camera and generates a grayscale image, searches for the laser point area on the grayscale image and calculates the grayscale value, and determines whether the wafer has slipped out based on the relationship between the grayscale value of the laser point area and the set threshold.

2. A wafer slip detection method according to claim 1, characterized in that: When the fill light is on, the visual recognition system extracts image a from the real-time video sent by the industrial camera at a set frame rate for preprocessing. This process removes noise from image a and improves its contrast. The preprocessed image a is then processed using the YOLOv8 model. The YOLOv8 model is trained by inputting a large number of wafer slip-out photos and non-slip-out photos; When the fill light is not turned on, the visual recognition system extracts picture b from the real-time video sent by the industrial camera at a set frame rate and performs grayscale processing to generate a grayscale image. The connected area is searched on the grayscale image, and the center of mass position is extracted. The laser point area is verified for the found connected area.

3. A wafer slip detection method according to claim 2, characterized in that: The preprocessing of the image a includes performing multi-level noise reduction on the extracted image a, and using a median filtering method to remove salt and pepper noise in the image a, wherein the salt and pepper noise is image noise generated by liquid splashing; A bilateral filtering method is used to remove texture noise in the image, where the texture noise is image noise formed by the polishing pad.

4. A wafer slip detection method according to claim 3, characterized in that: The denoised image a is Fourier transformed to convert the brightness channel of the image a to the frequency domain. Periodic noise appears as spikes at a specific frequency. A band-stop filter is used to shield the high-frequency noise corresponding to the periodic noise and the low-frequency noise within a set radius in the low-frequency area. The filtered frequency domain signal is converted back to the spatial domain to obtain the denoised brightness channel. The HSV image corresponding to the denoised brightness channel is converted back to the BGR color space to output the final image.

5. The wafer slip detection method according to claim 2, wherein: Improving the contrast of the extracted image a, including converting the extracted image a into the LAB color space, configuring the contrast limit threshold and the number of local area divisions of the CLAHE algorithm, and dividing the extracted image a into several local areas according to the number of local area divisions for contrast improvement processing; The CLAHE algorithm is applied to perform contrast enhancement processing on the extracted image a based on a set contrast limit threshold.

6. The wafer slip detection method according to claim 5, characterized in that: If the polishing pad is white, the contrast limit threshold is set to 2-2.5; If the polishing pad is black, the contrast limit threshold is set to 3-4; The number of local area divisions is 64-256, and the extracted image a is divided into a number of local areas in equal parts in length and width directions.

7. The wafer slip detection method according to claim 1, characterized in that: The YOLOv8 model includes an input end, a backbone network, a neck network, a detection head, and an output end. The backbone network is a convolutional neural network model used to extract wafer edge texture features, grinding head texture features, and grinding head global features. The extracted wafer edge texture features are subjected to multi-scale feature fusion and multi-scale output by the neck network of the YOLOv8 model: Including, the P1 level is used to detect wafer edge detail features, the P2 level is used to detect wafer body features, and the P3 level is used to detect complete wafer position features. The detection head predicts the target position and category based on the three-scale output features of the neck network, and the output end outputs the bounding box position information and confidence; The extracted grinding head texture features and global features are fused through the neck network of the YOLOv8 model. The FPN network is combined with the PAN network for feature fusion. The detection head predicts the target position and category based on the fused grinding head features. The output end outputs the bounding box position information and confidence level. Based on the position frame information of the polishing head and the position frame information of the wafer, it is determined whether the wafer has slipped out.

8. A wafer slip detection method according to claim 7, characterized in that: The neck network adopts a bidirectional feature pyramid structure to perform upsampling, downsampling and splicing operations on the output features. The P1 layer is used for small target feature detection, the P2 layer is used for medium target feature detection, and the P3 layer is used for large target feature detection.

9. The wafer slip detection method according to claim 7, characterized in that: Determining whether the wafer has slipped out based on the position frame information of the grinding head and the position frame information of the wafer, including converting the bounding box parameters into absolute coordinates, and determining whether the wafer exceeds the edge of the grinding head based on the absolute coordinates; If the wafer exceeds the edge of the grinding head in 4 out of 5 consecutive frames, it is determined that the wafer has slipped out.

10. The wafer slip detection method according to claim 2, characterized in that: Verify the laser point area found in the connected area, including size verification, gray value verification, circularity verification and intensity distribution verification; The intensity distribution is verified by determining whether the grayscale value at the center point is the highest.

11. The wafer slip detection method according to claim 1, characterized in that: The machine's PLC system is connected to the vision system via Ethernet, and the algorithm module of the vision system is deployed on the edge computer device. When the machine turns on the fill light, the real-time data generated in the PLC fill light control register is read by the edge computer, and the real-time data in the register is used to determine whether the machine turns on the fill light to select different processing modes.

Images