TWI835524B - Machine training system of chip position correctness and machine training method of chip position correctness - Google Patents

Abstract

A machine learning system and a machine learning method for the correctness of the chip position, wherein the machine learning method includes: obtaining a chip image data and a socket image data from a training image data according to the brightness characteristics of a chip and a socket; according to a relative position relationship between the chip image data and the socket image data, determining whether a position offset phenomenon occurs; when the position offset phenomenon occurs, labeling the training image data as an unqualified data; and when the position offset phenomenon does not occur , labeling the training image data as a qualified data.

Description

Machine learning system for chip position accuracy and machine learning method for chip position accuracy

The invention relates to a training system and a training method for the correctness of the position of electronic components, and in particular to a machine learning system and a machine learning method for the correctness of the position of a chip.

The current chip positioning equipment first places the chip on the base through a robot arm, and then uses an optical rangefinder to sequentially emit laser light at two diagonal positions of the chip. When the optical rangefinder receives reflected light from two opposite corners, based on the time difference between the time point when the laser light is emitted and the time point when the reflected light is received, it can be estimated that the two diagonal positions of the wafer are related to each other. The distance between optical rangefinders.

When the distance between the first diagonal corner of the wafer and the optical rangefinder is equal to the distance between the second diagonal corner of the wafer and the optical rangefinder, it means that the wafer is placed correctly. On the contrary, when the distance between the first diagonal corner of the wafer and the optical rangefinder is not equal to the distance between the second diagonal corner of the wafer and the optical rangefinder, it means that the wafer is placed in an incorrect position. It takes some time every time the distance is measured by the optical rangefinder.

The technical problem to be solved by the present invention is to provide a machine learning system for the correctness of the wafer position and a machine learning method for the correctness of the wafer position in view of the shortcomings of the existing technology.

In order to solve the above technical problems, one of the technical solutions adopted by the present invention is to provide a machine learning system for wafer position accuracy, including: wafer positioning equipment; a server, electrically connected to the wafer positioning equipment, and the server positions from the wafer The device obtains multiple pieces of training image data, and the server performs the following operations on each training image data: distinguishes the chip image data and the base image data from the training image data based on the chip brightness characteristics and the base brightness characteristics; based on the chip image data The relative positional relationship between the data and the base image data is used to determine whether position shift occurs; when position shift occurs, the training image data is marked as unqualified data; and when position shift does not occur, training is marked The image data is qualified data; after multiple training image data are marked, the server uses a machine learning algorithm to train the multiple training image data in order to generate a machine learning model.

In order to solve the above technical problems, one of the technical solutions adopted by the present invention is to provide a machine learning method for the correctness of the wafer position, which is executed by the wafer positioning equipment and the server. The machine learning method includes: obtaining multiple data from the wafer positioning equipment. pieces of training image data; the server obtains multiple pieces of training image data from the chip positioning equipment, and the server performs the following operations on each training image data: distinguishes among the training image data according to the chip brightness characteristics and the base brightness characteristics. The chip image data and the base image data are generated; based on the relative positional relationship between the chip image data and the base image data, it is determined whether a position shift occurs; when a position shift occurs, the training image data is marked as unqualified data. ; and when position deviation does not occur, mark the training image data as qualified data; after multiple training image data are marked, the server uses a machine learning algorithm to train the multiple training image data to generate Machine learning model.

One of the beneficial effects of the present invention is that the machine learning system and the machine learning method for the correctness of the chip position provided by the present invention perform machine learning training based on complete image data, which can reduce the probability of misjudgment of the chip position. Compared with the past, after placing the chip on the base of the chip positioning equipment, in addition to taking the image, it is also necessary to measure the distance between the two diagonal positions of the chip through a laser rangefinder to confirm the placement of the chip. Is the placement correct? The machine learning system of the present invention only needs to directly capture and process images after the chip is placed on the base. The process is faster than laser ranging, so the process speed is accelerated.

In order to further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are only for reference and illustration and are not used to limit the present invention.

The following is a specific embodiment to illustrate the implementation of the "machine learning system for wafer position accuracy and the machine learning method for wafer position accuracy" disclosed in the present invention. Those skilled in the art can understand from the content disclosed in this specification. Advantages and effects of the present invention. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the concept of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depictions based on actual dimensions, as is stated in advance. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of the present invention.

It should be understood that although terms such as “first”, “second” and “third” may be used herein to describe various elements or signals, these elements or signals should not be limited by these terms. These terms are primarily used to distinguish one component from another component or one signal from another signal. In addition, the term "or" used in this article shall include any one or combination of more of the associated listed items depending on the actual situation.

FIG. 1 is a functional block diagram of the machine learning system for wafer position accuracy of the present invention. As shown in FIG. 1 , the machine learning system for wafer position accuracy includes a wafer positioning device 1 and a server 2 , and the server 2 is electrically connected to the wafer positioning device 1 . The wafer positioning equipment 1 includes a camera 11, and the camera 11 shoots toward the base to obtain training image data. When a chip is placed on the base, the training image data includes chip image data and base image data. The chip positioning device 1 transmits the training image data to the server 2. When the server 2 receives the training image data, the server 2 distinguishes the chip image data and the base image data from the training image data, and then the server 2 based on the relative positional relationship between the chip image data and the base image data, Determine whether position deviation occurs. When the position deviation occurs, server 2 marks the training image data as unqualified data. When no position deviation occurs, server 2 marks the training image data as qualified data.

FIG. 2 is a flow chart of the first embodiment of the machine learning method for wafer position accuracy of the present invention. As shown in FIG. 2 , regarding step S201 , the camera 11 of the wafer positioning equipment 1 acquires training image data. Regarding step S203, the server 2 obtains training image data from the wafer positioning device 1. Regarding step S205, the server 2 distinguishes the chip image data and the base image data from the training image data based on the chip brightness characteristics and the base brightness characteristics.

FIG. 3 is a schematic diagram of an embodiment in which the server 2 distinguishes the chip image data and the base image data from the training image data. As shown in Figure 3, when a chip is placed on the base, the camera 11 shoots toward the base to obtain training image data. The chip brightness characteristic of the chip is that multiple data points with a brightness value higher than the brightness critical value are concentrated at In the rectangular area, as for the base brightness characteristics of the base, multiple data points with brightness values higher than the brightness critical value are distributed in four arc-shaped areas C1 ~ C4, of which arc-shaped areas C1 and arc-shaped areas C2 are located at a diagonal On the line, the arc-shaped area C3 and the arc-shaped area C4 are located on another diagonal line, and the brightness value of the area surrounded by the arc-shaped areas C1 to C4 is lower than the brightness critical value. The server 2 distinguishes the chip image data M11 and the base image data M12 from the training image data M1 based on the chip brightness characteristics and the base brightness characteristics.

Regarding step S207, the server 2 performs principal component analysis (PCA) on the wafer image data to obtain the wafer center point. Regarding step S209, the server 2 performs principal component analysis on the base image data to obtain the base center point.

FIG. 4 is a schematic diagram of an embodiment in which the server 2 of the present invention uses principal component analysis to obtain the center point of the wafer and the base and the offset angle between the wafer and the base. As shown in Figure 4, after the server 2 performs principal component analysis on the wafer image data M11, it obtains the wafer center point P1 and the first long axis L1 and the first short axis W1 passing through the wafer center point P1. The dimensional coordinates are (X1, Y1). In the same way, after server 2 performs principal component analysis on the base image data M12, it obtains the base center point P2, the second long axis L2 and the second short axis W2 passing through the base center point P2, and the base center point P2 The two-dimensional coordinates are (X2, Y2).

Regarding step S211, the server 2 calculates the offset distance between the wafer center point P1 and the base center point P2. For example, when the two-dimensional coordinates of the wafer center point P1 are (X1, Y1) and the coordinates of the base center point P2 are (X2, Y2), the calculation formula of the offset distance is: .

Regarding step S213, the server 2 calculates the offset angle between the wafer and the susceptor. For example, the first long axis L1 is perpendicular to the first short axis W1, and the second long axis L2 is perpendicular to the second short axis W2. When the angle between the first long axis L1 and the second long axis L2 is 10 degrees, and the angle between the first short axis W1 and the second short axis W2 is also 10 degrees, at this time, the angle between the wafer and the base The offset angle is 10 degrees.

Regarding step S215, the server 2 normalizes the offset distance between the wafer center point P1 and the base center point P2 to obtain the offset distance ratio, where the standardized formula is (between the wafer center point P1 and the base center point P2 offset distance)/(diagonal length of the wafer), but the present invention is not limited to this. Through standardized processing, even if the resolution of the camera 12 changes, the offset distance ratio can accurately reflect the offset degree of the chip relative to the base.

Regarding step S217, the server 2 determines whether the offset distance ratio is greater than a preset critical distance ratio. If yes, proceed to step S219. If not, proceed to step S221. Regarding step S219, server 2 marks the training image data as unqualified data. Regarding step S221, the server 2 determines whether the offset angle is greater than the critical offset angle. If yes, proceed to step S223. If not, proceed to step S225.

Regarding step S223, the server 2 marks the training image data as unqualified data. Regarding step S225, the server 2 marks the training image data as qualified data.

After steps S219, S223 and S225, step S227 follows. Regarding step S227, the server 2 determines whether the quantity of training image data reaches a quantity threshold. If yes, proceed to step S229. If not, return to step S201.

Regarding step S229, the server 2 uses a machine learning algorithm to train multiple marked training image data to generate a machine learning model.

Figure 5 is a schematic diagram of an embodiment of the machine learning model of the present invention. As shown in Figure 5, server 2 uses a support vector machine (support vector machine) to train multiple labeled training image data to generate a support vector machine model. The support vector machine model includes a data grouping boundary L, and the opposite sides of the data grouping boundary L are the location qualified area PASS and the location unqualified area NG respectively. When the support vector machine model receives the test image data, the support vector machine model can determine that the test image data belongs to the position qualified area PASS or the position unqualified area NG. When the test image data belongs to the position qualified area PASS, it means that the chip is correctly placed on the base. On the contrary, when the test image data belongs to the unqualified position area, it means that the chip is not correctly placed on the base. Regardless of whether the test result is qualified or unqualified, the server 2 will send the test result back to the chip positioning equipment 1, and the chip positioning equipment 1 can adjust the positioning parameters according to the test results of the support vector machine model.

6A and 6B are flowcharts of a second embodiment of the machine learning method for wafer position accuracy of the present invention. As shown in FIG. 6A , regarding step S601 , the camera 11 of the wafer positioning equipment 1 acquires training image data. Regarding step S603, the server 2 obtains training image data from the wafer positioning device 1. Regarding step S605, the server 2 determines whether there is a wafer placed on the base. If so, step S607 is followed. If not, proceed to step S609. Regarding step S607, the server 2 determines whether the number of wafers placed on the base is single. Regarding step S609, server 2 marks the training image data as unqualified data.

Regarding how the server 2 determines whether there is a chip placed on the base, specifically, if the server 2 detects the brightness characteristics of the chip in the training image data, it means that there is a chip placed on the base. On the contrary, if the server 2 cannot detect the brightness characteristics of the chip in the training image data, it means that the chip is not placed on the base.

If the server 2 determines that the number of wafers placed on the base is single, step S611 follows. If the server 2 determines that the number of wafers placed on the base is not single, step S613 follows.

Regarding step S611, the server 2 distinguishes the chip image data and the base image data from the training image data based on the chip brightness characteristics and the base brightness characteristics. Regarding step S613, server 2 marks the training image data as unqualified data.

After step S611, step S615 follows. Regarding step S615, the server 2 performs principal component analysis on the wafer image data to obtain the wafer center point. Regarding step S617, the server 2 performs principal component analysis on the base image data to obtain the base center point.

As shown in FIG. 6B , regarding step S619, the server 2 calculates the offset distance between the wafer center point P1 and the base center point P2. Regarding step S621, the server 2 calculates the offset angle between the wafer and the susceptor.

Regarding step S623, the server 2 normalizes the offset distance between the wafer center point P1 and the base center point P2 to obtain an offset distance ratio.

Regarding step S625, the server 2 determines whether the offset distance ratio is greater than the critical distance ratio. If yes, proceed to step S627. If not, proceed to step S629. Regarding step S627, server 2 marks the training image data as unqualified data. Regarding step S629, the server 2 determines whether the offset angle is greater than the critical angle. If yes, proceed to step S631. If not, proceed to step S633.

Regarding step S631, server 2 marks the training image data as unqualified data. Regarding step S633, server 2 marks the training image data as qualified data.

After steps S609, S613, S627, S631 and S633, step S635 follows. Regarding step S635, the server 2 determines whether the quantity of training image data reaches a quantity threshold. If yes, proceed to step S637. If not, return to step S601.

Regarding step S637, the server 2 uses a machine learning algorithm to train multiple marked training image data to generate a machine learning model.

FIG. 7 is a flowchart of a method for the server 2 to determine whether multiple wafers are placed on the base, which means that FIG. 7 further describes the sub-steps of step S607 of FIG. 6 . As shown in FIG. 7 , regarding step S701 , the server 2 creates a frame surrounding the wafer image data. Regarding step S703, the server 2 obtains the first area enclosed by the frame and the second area of the wafer image data. Regarding step S705, the server 2 calculates the area difference between the first area and the second area. Regarding step S707, the server 2 calculates the area ratio of the area difference and the second area. Regarding step S709, the server 2 determines whether the area ratio is less than or equal to the critical area ratio. If yes, proceed to step S711. If not, proceed to step S713. Regarding step S711, the server 2 determines that only one wafer is placed on the base, and then proceeds to step S611. Regarding step S713, the server 2 determines that multiple wafers are placed on the base, and then proceeds to step S613.

Figure 8 is a schematic diagram of an embodiment of establishing a frame surrounding the chip image data. As shown in Figure 8, the server 2 creates a minimum frame C surrounding the chip image data M11. When the chip image data M11 is When the percentage of the area enclosed by the frame line C is greater than a critical percentage, it means that there is a high probability that only one chip is placed on the base. On the contrary, when the percentage of the area of the wafer image data M11 compared to the area enclosed by the frame line C is less than the critical percentage, it means that there is a high probability that multiple wafers are placed on the base.

FIG. 9 is a flow chart of the third embodiment of the machine learning method for wafer position accuracy of the present invention. The machine learning method of Figure 9 includes steps S901 to S927, where steps S901 to S913 are respectively the same as steps S201 to S213 of Figure 2, and steps S917 to S927 are respectively the same as steps S219 to S229 of Figure 2, with the difference being step S915. Regarding step S915, the server 2 determines whether the offset distance is greater than the critical distance. If yes, proceed to step S917. If not, proceed to step S919.

Furthermore, regarding the data transmission between the chip positioning device 1 and the server 2, for example, the chip positioning device 1 can first encode the training image data to form string data through base 64 encoding rules. Then, the chip positioning device 1 transmits the string data to the server 2 through the Transmission Control Protocol. When the server 2 receives the string data from the chip positioning device 1, it decodes the string data to obtain training image data, thereby solving the problem of unrecognizable cross-platform objects. The above data transmission methods are only examples, and the present invention is not limited thereto.

[Beneficial effects of the embodiment]

One of the beneficial effects of the present invention is that in the machine learning system and the machine learning method for the correctness of the chip position provided by the present invention, the server performs machine learning training based on complete image data, which can reduce misjudgments of the chip position. Probability. Compared with the past, after placing the chip on the base of the chip positioning equipment, in addition to taking the image, it is also necessary to measure the distance through the two diagonal positions of the laser rangefinder chip to confirm the placement of the chip. Is the placement correct? The machine learning system of the present invention only needs to directly capture and process images after the chip is placed on the base. The process is faster than laser ranging, so the process speed is accelerated. The chip provided by the present invention The machine learning system for position accuracy and the machine learning method for chip position accuracy can reduce the misjudgment rate by training an evaluation model based on complete image data.

The contents disclosed above are only preferred and feasible embodiments of the present invention, and do not limit the scope of the patent application of the present invention. Therefore, all equivalent technical changes made by using the description and drawings of the present invention are included in the application of the present invention. within the scope of the patent.

1: Wafer positioning equipment 2:Server 11:Camera C1~C4: Arc area P1: wafer center point L1: first long axis W1: first short axis P2: base center point L2: second long axis W2: Second short axis L: data grouping boundary PASS: location qualified area NG: location unqualified area M1: training image data M11: chip image data M12: Base image data C: frame line S201~S229: steps S601~S637: steps S701~S713: steps S901~S927: steps

FIG. 1 is a functional block diagram of the machine learning system for wafer position accuracy of the present invention.

FIG. 2 is a flow chart of the first embodiment of the machine learning method for wafer position accuracy of the present invention.

Figure 3 is a schematic diagram of distinguishing wafer image data and base image data through brightness characteristics.

FIG. 4 is a schematic diagram of obtaining the center point of the wafer, the center point of the base, and the offset angle between the wafer and the base through principal component analysis.

Figure 5 is an embodiment of a trained machine learning model of the present invention.

6A and 6B are flowcharts of a second embodiment of the machine learning method for wafer position accuracy of the present invention.

Figure 7 is a flowchart of a method for the server to determine whether multiple wafers are stacked on the base.

FIG. 8 is a schematic diagram of an embodiment of establishing a frame surrounding the wafer image data.

FIG. 9 is a flow chart of the third embodiment of the machine learning method for wafer position accuracy of the present invention.

S201~S229: steps

Claims (11)

A machine learning system for chip position accuracy, including: a chip positioning device; a server electrically connected to the chip positioning device, the server obtains a plurality of training image data from the chip positioning device, and the server Each training image data performs the following operations: distinguishing a chip image data and a base image data from the training image data according to a chip brightness characteristic and a base brightness characteristic, wherein the chip brightness characteristic has a brightness value higher than A plurality of data points with a brightness threshold value are concentrated in a rectangular area. The base brightness characteristic is that multiple data points with a brightness value higher than the brightness threshold value are distributed in four arc-shaped areas, two of which are arc-shaped. The area is located on one diagonal line of the rectangular area, and the other two arc-shaped areas are located on the other diagonal line of the rectangular area; according to the relative positional relationship between the wafer image data and the base image data , determine whether a position shift phenomenon occurs; when the position shift phenomenon occurs, mark the training image data as unqualified data; and when the position shift phenomenon does not occur, mark the training image data as qualified data ; After the plurality of training image data are marked, the server uses a machine learning algorithm to train the plurality of training image data to generate a machine learning model. A machine learning system for chip position accuracy as described in claim 1, wherein the server's determination of whether the position deviation occurs includes: performing principal component analysis on the chip image data to obtain a chip center point; Perform principal component analysis on the image data to obtain a base center point; calculate the wafer center point and an offset distance between the center points of the base; calculate an offset angle between the wafer and the base; normalize the offset distance to obtain an offset distance ratio; when the offset distance ratio is greater than When a critical distance ratio or the offset angle is greater than a critical angle, mark the training image data as unqualified data; and when the offset distance ratio is less than or equal to the critical distance ratio and the offset angle is less than or equal to the critical angle, mark the training image data as qualified data. The machine learning system for chip position accuracy as described in claim 1, wherein the machine learning algorithm is a support vector machine, and the machine learning model is a support vector machine model, and the support vector machine model includes a data grouping boundary, The opposite sides of the data grouping boundary are respectively a position qualified area and a position unqualified area. The machine learning system for chip position accuracy as described in claim 1, wherein the server also performs the following operations for each piece of training image data: creates a frame surrounding the chip image data; obtains a frame surrounded by the frame The first area and the second area of the chip image data; obtain the area difference between the first area and the second area; obtain the area ratio between the area difference and the second area; when the area ratio is greater than a critical area ratio , it is judged that multiple wafers are placed on the base; when the area ratio is less than or equal to the critical area ratio, it is judged that only a single wafer is placed on the base. The machine learning system for chip position accuracy as described in request item 1, wherein the server also performs the following operations on each training image data: the server determines whether the training image data has the chip brightness feature; when the training image The chip brightness feature does not exist in the data, and the training image data is marked as unqualified data. The machine learning system for chip position accuracy as described in claim 1, wherein the chip positioning device obtains the training image data and encodes the training image data into multiple string data. When the server receives the multiple string data, Pen string data At this time, the server decodes the plurality of string data to obtain the plurality of training image data. A machine learning method for chip position accuracy is executed by a chip positioning device and a server. The machine learning method includes: obtaining a plurality of training image data from the chip positioning device; and using the server to obtain a plurality of training image data from the chip positioning device. The multiple pieces of training image data are obtained, and the server performs the following operations on each training image data: distinguishes a chip image data and a base from the training image data based on a chip brightness feature and a base brightness feature. Base image data, in which the chip brightness feature is a plurality of data points with a brightness value higher than a brightness critical value concentrated in a rectangular area, and the base brightness feature is a distribution of multiple data points with a brightness value higher than the brightness critical value In four arc-shaped areas, two of the arc-shaped areas are located on one diagonal line of the rectangular area, and the other two arc-shaped areas are located on the other diagonal line of the rectangular area; according to the wafer The relative position relationship between the image data and the base image data is used to determine whether a position shift phenomenon occurs; when the position shift phenomenon occurs, the training image data is marked as unqualified data; and when the position shift phenomenon does not occur, When the offset phenomenon occurs, the training image data is marked as qualified data; after the multiple training image data are marked, the server uses a machine learning algorithm to train the multiple training image data to generate a machine Learning model. The machine learning method for chip position accuracy as described in claim 7, wherein the server's determination of whether the position deviation occurs includes: performing principal component analysis on the chip image data to obtain a chip center point; Image data Perform principal component analysis to obtain a base center point; calculate an offset distance between the wafer center point and the base center point; calculate an offset angle between the wafer and the base; calculate the offset The distance is normalized to obtain an offset distance ratio; when the offset distance ratio is greater than a critical distance ratio or the offset angle is greater than a critical angle, the training image data is marked as unqualified data; and when the offset distance When the ratio is less than or equal to the critical distance ratio and the offset angle is less than or equal to the critical angle, the training image data is marked as qualified data. The machine learning method for chip position accuracy as described in claim 7, wherein the server also performs the following operations on each training image data: creates a frame surrounding the chip image data; obtains a frame surrounded by the frame The first area and the second area of the chip image data; obtain the area difference between the first area and the second area; obtain the area ratio between the area difference and the second area; when the area ratio is greater than a critical area ratio , it is judged that multiple wafers are placed on the base; when the area ratio is less than or equal to the critical area ratio, it is judged that only a single wafer is placed on the base. The machine learning method for chip position accuracy as described in request item 7, wherein the server also performs the following operations on each training image data: the server determines whether the training image data has the chip brightness feature; when the training image The chip brightness feature does not exist in the data, and the training image data is marked as unqualified data. The machine learning method for chip position accuracy as described in claim 7, wherein the server's determination of whether the position deviation occurs includes: performing principal component analysis on the chip image data to obtain a chip center point; Perform principal component analysis on the image data to obtain a base center point; calculate an offset distance between the chip center point and the base center point; calculate an offset angle between the chip and the base; determine the Whether the offset distance is greater than a critical distance; when the offset distance ratio is greater than the critical distance, mark the training image data as incorrect Qualified data; when the offset distance is not greater than the critical distance, determine whether the offset angle is greater than a critical angle; when the offset angle is greater than the critical angle, mark the training image data as unqualified data; when the offset angle is greater than the critical angle, If the offset angle is not greater than the critical angle, the training image data is marked as qualified data.

Images