JP4514230B2 - Component suction posture discrimination method and component suction posture discrimination system - Google Patents

Description

  The present invention relates to a component suction posture determination method and a component suction posture determination system for imaging a component sucked by a suction nozzle of an electronic component mounting machine with a camera and determining whether the suction posture of the component is normal suction or oblique suction by image processing technology. It is invention regarding.

  In general, in an electronic component mounting machine, a component is sucked by a suction nozzle, and the component is transferred onto a circuit board and mounted at a predetermined position on the circuit board. In addition, a camera that captures the part sucked by the suction nozzle from the lower surface side is installed, and the type of the part is confirmed by the part image picked up by the camera, or the deviation of the suction position of the part with respect to the suction nozzle is corrected. I have to.

  Normally, as shown in FIG. 18A, the component 2 is horizontally attracted to the suction nozzle 1, but as shown in FIG. 18B, the suction nozzle 1 for some reason. In some cases, the component 2 may be attracted diagonally. Since such oblique adsorption (abnormal adsorption) causes a mounting failure, as shown in Patent Document 1 (Japanese Patent Laid-Open No. 6-216484), an optical sensor is used as means for detecting oblique adsorption, and the adsorption nozzle 1 The light-emitting element 3 and the light-receiving element 4 of the optical sensor are arranged on both sides of the component 2 adsorbed on the surface so that the optical axis 5 passes through a position slightly lower than the lower surface of the component 2 in the normal adsorption posture. is there. As shown in FIG. 18 (a), if the suction posture of the component 2 is normal suction, the optical axis 5 is not blocked by the component 2, but as shown in FIG. 18 (b) When the suction posture 2 is oblique suction, the oblique suction is detected by blocking the optical axis 5 at the lower part of the component 2.

  However, in this configuration, it is necessary to provide an optical sensor (the light emitting element 3 and the light receiving element 4) as means for detecting oblique adsorption, and there is a disadvantage that the cost increases accordingly.

Therefore, as shown in Patent Document 2 (Japanese Patent Laid-Open No. 2006-114821), a plurality of seek lines that intersect with the outline of the component image in the normal suction posture assumed in advance are related to the center line of the component image. Some of them are set to symmetrical positions, and brightness change patterns on seek lines having a symmetrical positional relationship are compared to determine whether the suction posture of the component is normal suction or diagonal suction.
JP-A-6-216484 (first page, etc.) Japanese Unexamined Patent Publication No. 2006-114821 (page 4, etc.)

  However, in the technique of the above-mentioned Patent Document 2, since the design developer sets a discrimination standard for discriminating between normal adsorption and abnormal adsorption (oblique adsorption), the discrimination standard varies depending on the design developer's subjectivity. The discrimination rate of abnormal adsorption will deteriorate.

  In order to solve this problem, as shown in the specification of Japanese Patent Application No. 2006-149799, the present applicant previously collects a large number of normal suction component images and abnormal suction (diagonal suction) component images, After extracting feature quantities from each part image, the statistical distribution of the feature quantity data is evaluated by a discriminant analysis method, and a discrimination criterion for discriminating between normal adsorption and abnormal adsorption is set and stored in the memory. After that, every time a component is sucked to the suction nozzle while the electronic component mounting machine is in operation, a feature amount is extracted from the component image picked up by the camera, and the component is sucked from the feature amount according to the determination criterion. A technology has been developed to determine whether the posture is normal or abnormal.

  However, in the discrimination method using this discriminant analysis method, although the discrimination rate of abnormal adsorption is improved as compared with the prior art, it is still 93.1 [%], and it is still necessary to improve the discrimination rate of abnormal adsorption. .

  The present invention has been made in view of such circumstances. Therefore, the object of the present invention is to change or add software only to a system that accurately detects abnormal suction of components sucked by a suction nozzle of an electronic component mounting machine. Another object of the present invention is to provide a component suction posture determination method and a component suction posture determination system that can be configured inexpensively.

In order to achieve the above object, according to the first and third aspects of the present invention, a part picked up by a pick-up nozzle of an electronic component mounting machine is picked up by a camera, and the pick-up posture of the part is normal pick-up or abnormal pick-up by image processing technology. In order to determine whether or not, a lot of normal suction part image data and abnormal part image data collected in advance are learned as neural network data as teacher data, and captured by the camera while the electronic component mounting machine is in operation. Image data of a part is input to the neural network, and based on an output value of the neural network, it is determined whether the suction posture of the part is normal suction or abnormal suction. There are four types of postures: normal adsorption, abnormal adsorption by right diagonal adsorption, abnormal adsorption by left diagonal adsorption, and abnormal adsorption by lateral adsorption. It is configured to have four output layers that output numerical values according to the degree of similarity between teacher data and input data of each pattern divided into turns , and compare the output values of each output layer with normal adsorption. It is characterized by distinguishing from abnormal adsorption.

  In this way, a system for detecting abnormal suction of components can be configured at low cost by only changing or adding software to an existing electronic component mounting machine without requiring addition of an optical sensor or the like. In addition, the learning effect by the neural network can accurately determine whether the suction posture of the component is normal suction or abnormal suction, and the discrimination rate of abnormal suction can be higher than that of the discriminant analysis method.

As a result of observing various abnormal adsorption images in the course of studying the present invention, the abnormal adsorption patterns are roughly divided into three patterns (abnormal adsorption by right diagonal adsorption, diagonal left), as shown in FIGS. It was found that it can be classified into abnormal adsorption by adsorption and abnormal adsorption by lateral adsorption .

Therefore, in the inventions according to claims 1 and 3 , the component adsorption posture is divided into four patterns of normal adsorption, abnormal adsorption by right diagonal adsorption, abnormal adsorption by left diagonal adsorption, and abnormal adsorption by lateral adsorption. 4 output layers for outputting numerical values corresponding to the degree of similarity between the teacher data of the pattern and the input data are provided . In this way, since abnormal suction can be determined for each pattern, the frequency of misjudging normal suction as abnormal suction can be reduced to almost zero. The cycle time loss due to the stoppage of the electronic component mounting machine can be almost eliminated.

In this case, as in claims 2 and 4 , it is only necessary to select an output layer having the maximum output value from a plurality of output layers and discriminate between normal adsorption and abnormal adsorption. For example, if the output layer with the maximum output value is classified into a normal adsorption group, it is determined as normal adsorption, and the output layer with the maximum output value is classified into an abnormal adsorption group. For example, it may be determined as abnormal adsorption.

  Hereinafter, two Examples 1 and 2, which embody the best mode for carrying out the present invention, will be described.

A first embodiment as a reference example related to the present invention will be described with reference to FIGS.
First, a schematic configuration of the entire electronic component mounting machine will be described with reference to FIG.
The X-axis slide 11 is provided so as to be slidable in the X-axis direction (left-right direction in FIG. 1) by the X-axis ball screw 12, and the Y-axis slide 13 is moved by the Y-axis ball screw 14 relative to the X-axis slide 11. It is provided so as to be slidable in the Y-axis direction (perpendicular to the plane of FIG. 1).

  The Y-axis slide 13 is provided with a suction head 15, and a suction nozzle 17 is attached downward to a suction holder 16 provided on the suction head 15 so as to be movable up and down. On the lower surface of the suction holder 16, a backlight 19 is provided for brightening the background of the component 18 sucked by the suction nozzle 17.

  The X-axis slide 11 is provided with a camera 22 such as a CCD camera facing downward from the lower surface side of the component 18 sucked by the suction nozzle 17 via a pair of reflecting mirrors 20 and 21. The deviation of the suction position of the component 18 with respect to the suction nozzle 17 is corrected based on the component image captured by the camera 22.

  One reflecting mirror 20 is provided so as to be positioned below the component 18 sucked by the suction nozzle 17, and the component 18 sucked by the suction nozzle 17 is illuminated below the reflecting mirror 20 from its lower surface side. The front light 23 is provided. The reflecting mirror 20 is subjected to a half mirror process so that the light from the front light 23 is transmitted upward.

  This electronic component mounting machine control device (not shown) images the component 18 adsorbed to the adsorption nozzle 17 with the camera 22 from the lower surface side, and whether the adsorption posture of the component 18 is “normal adsorption” by image processing technology. Determine whether it is “abnormal adsorption”. Hereinafter, the method for determining the suction posture will be described in detail.

In the first embodiment, the suction posture of the component 18 is determined as follows using a three-layer neural network in which the number of output layers is one.
As shown in FIG. 2, the position and angle of the component in the image captured by the camera 22 are detected, angle correction / position correction is performed, and only the component portion is cut out with a size of, for example, 32 × 24 [pixel]. A gray scale image is used as component image data. A part of part images used as teacher data and test data in the discrimination test described below is shown in FIGS.

FIG. 3 shows an image example of normal adsorption.
FIG. 4 shows an example of an image of left diagonal suction. Left diagonal adsorption is an adsorption state in which the left electrode side surface of the component is visible, and this is also a kind of abnormal adsorption.
FIG. 5 shows an example of an image of right diagonal suction. The right diagonal adsorption is an adsorption state in which the right electrode side surface of the component is visible, and is a kind of abnormal adsorption.

FIG. 6 shows an example of lateral suction image. Lateral suction is a suction state in which the original side surface of a component is seen downward, and this is also a kind of abnormal suction.
The abnormal suction images used in this example are only those in which abnormal suction could not be detected by the conventional detection method based on the external size.

  As shown in FIG. 7, the number of component images used in this example is 3657 normal suction images, 6644 abnormal suction images (breakdown includes 3268 left diagonal suction images and 3332 right diagonal suction images). , 44 horizontal suction images), a total of 10301 sheets. From these images, select only 377 normal suction images and 753 abnormal suction images (including 368 left diagonal suction images, 362 right diagonal suction images, and 23 horizontal suction images). Thus, a total of 1130 images are used as teacher data. The remaining images are used for performance evaluation of the neural network as test data.

In the first embodiment, a three-layer neural network shown in FIG. 8 is used.
In this three-layer neural network, the number of neurons in the input layer is 32 × 24, the number of neurons in the intermediate layer is 128, and the number of output layers is one. As the output value of the output layer is closer to “1.0”, “normal adsorption” is indicated, and as “0.0” is indicated, “abnormal adsorption” is indicated.

  As a learning method, a back propagation method was used. In learning, the parameters of the neural network were set as a learning coefficient α = 0.9 and a bias update coefficient β = 0.7. The slope of the sigmoid function was U0 = 5.0.

  In learning by the back propagation method, as shown in FIG. 9A, the same teacher data is repeatedly learned N times, and then the next teacher data is learned. In addition, the order of the teacher data for learning is changed randomly at each learning. In this example, the number of repetitions N = 10.

  When learning each teacher data, as shown in FIG. 9B, an error is added by parallel / rotating the image. Further, as shown in FIG. 9C, for one teacher data, an image inverted in the horizontal direction, an image inverted in the vertical direction, and an image inverted in the horizontal and vertical directions are learned at the same time.

  The teacher data learning by the three-layer neural network described above is executed according to the learning program of FIG. 10 by the control device of the electronic component mounting machine. The learning program in FIG. 10 is executed before performing the electronic component mounting work. When this program is started, first, in step 101, the normal suction component image as shown in FIG. 3 and the abnormal suction (left diagonal suction, right diagonal suction, lateral suction) as shown in FIGS. ) A large number of parts images.

  Thereafter, the process proceeds to step 102, and after detecting the position / angle of the part in the image, the process proceeds to step 103, where the position / angle of the part in the image is corrected according to the xy coordinates, and the x-axis and y-axis Are matched with the vertical and horizontal center lines of the component, and only the component portion in the image is cut out with a size of 32 × 24 [pixel]. At this time, the x-axis is set in the longitudinal direction of the component, and angle correction and position correction are performed so that the electrodes at both ends of the component are located at symmetrical positions with respect to the y-axis, and only the component portion is cut out.

  In the next step 104, a predetermined number of images are selected from the normal suction image and the abnormal suction image, respectively, and teacher data is created. Thereafter, the process proceeds to step 105, where the teacher data is input to the neural network shown in FIG. 8, and the normal suction image and the abnormal suction image are learned.

Thereafter, the process proceeds to step 106, where it is determined whether or not the number of learning is equal to or greater than a predetermined number. If the number of learning is less than the predetermined number, the process returns to step 105 and the neural network learning is repeated. Thereby, the learning of the neural network is repeated until the number of times of learning becomes a predetermined number or more. The number of learning repetitions (predetermined number) may be set to the number of learnings necessary to keep the average error rate (see FIG. 12) within the target value.
The learning program in FIG. 10 may be executed by a computer different from the control device for the electronic component mounting machine.

  On the other hand, the component suction posture determination program in FIG. 11 plays a role as the suction posture determination means in the claims together with the learning program in FIG. This program is started every time a component is sucked to the suction nozzle 17 during operation of the electronic component mounting machine, and the suction posture of the component is determined as follows.

  When this program is started, first, in step 201, an image of a component sucked by the suction nozzle 17 is captured by the camera 22 and captured. Thereafter, the process proceeds to step 202, and after detecting the position / angle of the part in the image, the process proceeds to step 203, where the position / angle of the part in the image is corrected according to the xy coordinates, and the x-axis and y-axis Are matched with the vertical and horizontal center lines of the component, and only the component portion in the image is cut out with a size of 32 × 24 [pixel]. At this time, the x-axis is set in the longitudinal direction of the component, and angle correction and position correction are performed so that the electrodes at both ends of the component are located at symmetrical positions with respect to the y-axis, and only the component portion is cut out.

  Then, in the next step 204, the extracted image data is input to the neural network of FIG. 8, and in the next step 205, it is determined whether or not the output value of the output layer of the neural network is greater than or equal to the determination threshold value. Then, it is determined whether the suction posture of the component is normal suction or abnormal suction. As the output value of the output layer is closer to 1.0, normal adsorption is indicated, and as the output value is closer to 0.0, abnormal adsorption is indicated.

  The inventor performed the neural network learning using the teacher data shown in FIG. 7 by the method described above, and performed a discrimination experiment using the test data shown in FIG. In this discrimination experiment, the result of measuring the change in average error rate using a neural network for each number of learning times is shown in FIG.

  Examining the results of this discrimination experiment, it was found that the average error rate was stable within about 0.5 [%] after the number of learnings exceeded about 1500 times. As shown in FIG. 13, there are 27 cases that were actually judged to be abnormal adsorption despite being normal adsorption, and conversely, those that were abnormal adsorption were judged to be normal adsorption. There were 17 cases.

The discrimination rate in this discrimination experiment was calculated by the following equation, a high discrimination rate of 99.5 [%] was obtained, and the error rate was 0.5 [%].
Discrimination rate = (3253 + 5874) / (3280 + 5891) × 100 [%]
= 99.5 [%]
Error rate = 100-99.5 [%]
= 0.5 [%]

  According to the first embodiment described above, a large number of normally picked-up component image data and abnormally picked-up component image data are learned by a neural network as teacher data, and the camera is operated while the electronic component mounting machine is in operation. Since the image data of the part imaged in 22 is input to the neural network and it is determined whether the suction posture of the part is normal suction or abnormal suction based on the output value of the neural network, the suction posture of the part is It can discriminate between normal adsorption and abnormal adsorption with high accuracy, and the abnormal adsorption discrimination rate can be increased as compared with the discriminant analysis method. In addition, there is an advantage that a system for detecting abnormal adsorption of components can be configured at low cost by only changing or adding software to the current electronic component mounting machine without adding an optical sensor or the like.

  In Example 1 above, it was found that when images with normal adsorption were misidentified as abnormal adsorption, the images clearly contained normal adsorption. On the other hand, when observing images in which abnormally adsorbed ones were misidentified as normal adsorbed, it was found that those images were clearly abnormally adsorbed. The reason for this is that despite the fact that there are several types of abnormal adsorption, they are grouped together into one “abnormal adsorption” pattern, resulting in some confusion in the neural network. It is presumed that it is because.

  As a result of observing various abnormal adsorption images in the course of studying the present invention, the abnormal adsorption patterns are roughly divided into three patterns (abnormal adsorption by right diagonal adsorption) as shown in FIGS. It was found that it can be classified into an abnormal adsorption due to left diagonal adsorption and an abnormal adsorption due to lateral adsorption).

  Therefore, in the second embodiment of the present invention, as shown in FIGS. 14 and 15, the abnormal adsorption pattern is divided into three patterns: abnormal adsorption by right diagonal adsorption, abnormal adsorption by left diagonal adsorption, and abnormal adsorption by lateral adsorption. In this way, the component adsorption posture is divided into four patterns: normal adsorption, abnormal adsorption by right diagonal adsorption, abnormal adsorption by left diagonal adsorption, and abnormal adsorption by lateral adsorption. Four output layers for outputting numerical values corresponding to the degree of similarity with input data are provided.

  In the second embodiment, the three-layer neural network shown in FIG. 14 is used, the number of neurons in the input layer is 32 × 24, the number of neurons in the intermediate layer is 128, the number of output layers is 4, and four outputs are output. The output layer having the maximum output value is selected from the layers 1 to 4 to discriminate between normal adsorption and abnormal adsorption.

  In this case, as shown in FIG. 15, the output layer 1 outputs a numerical value corresponding to the degree of similarity between the normal suction teacher data and the input data, and the output layer 2 is similar to the left diagonal suction teacher data and the input data. The output layer 3 outputs a numerical value according to the degree of similarity between the right diagonal suction teacher data and the input data, and the output layer 4 outputs the similarity between the side suction teacher data and the input data. The numerical value corresponding to is output. As a learning method, a back propagation method was used, and learning was performed in the same manner as in Example 1.

For example, if the output value of the output layer 1 is larger than the output values of the other output layers 2 to 4, it is determined as “normal adsorption”, and the output value of any of the output layers 2 to 4 is the output value of the output layer 1. If it is larger than that, it is determined as “abnormal adsorption”. This will be explained with a specific example.
Output value of output layer 1 (normal adsorption) = 0.3
Output value of output layer 2 (left diagonal adsorption) = 0.8
Output value of output layer 3 (right oblique adsorption) = 0.1
Output value of output layer 4 (lateral adsorption) = 0.2
The output value (0.8) of the output layer 2 (left oblique adsorption) is the maximum, and the output value (0.8) of the output layer 2 (left oblique adsorption) is the output layer 1 ( Since it is larger than the output value (0.3) of “normal adsorption”, it is determined as “abnormal adsorption”.

  Also in the second embodiment, the neural network was learned using the teacher data shown in FIG. 7, and a discrimination experiment was conducted using the test data shown in FIG. In this discrimination experiment, the result of measuring the change in average error rate using a neural network for each number of learning times is shown in FIG.

  From the result of this discrimination experiment, it was found that the average error rate was stable within about 1.5 [%] from the time when the number of learning exceeded about 1200 times. As shown in FIG. 17, although there was actually normal adsorption, none of them was erroneously determined to be abnormal adsorption, but what was abnormally determined to be abnormal adsorption was 134. There was an example.

The discrimination rate and error rate in this discrimination experiment are calculated by the following equations.
Discrimination rate = (3280 + 5757) / (3280 + 5891) × 100 [%]
= 98.5 [%]
Error rate = 100-98.5 [%]
= 1.5 [%]

  In the first embodiment in which the number of output layers is one neural network, the error rate is about 0.5 [%]. In the second embodiment, the number of output layers of the neural network is four. As a result, the error rate increased to about 1.5%. However, there has been no example of misclassifying a normal suction image as abnormal suction.

  As described above, in the second embodiment, the images that are erroneously determined are all images that are erroneously determined to be abnormally attracted from those that are abnormally attracted. When these misclassification images are observed, the misclassification images may contain a large number of adsorption states that have no problem in terms of mounting accuracy even if the component is mounted on the circuit board. found. Therefore, the substantial misclassification rate seems to be lower. Therefore, as in the second embodiment, the number of output layers of the neural network is increased to four of “normal adsorption”, “left diagonal adsorption”, “right diagonal adsorption”, and “lateral adsorption”, thereby improving the performance. A component abnormal suction discriminator can be configured.

  By the way, misjudging normal adsorption as abnormal adsorption leads to cycle time loss due to component loss and electronic component mounting machine stoppage, so the frequency of misidentifying normal adsorption as abnormal adsorption can be reduced as much as possible. desirable. On the other hand, as described above, examples of misidentifying an abnormally adsorbed item as normal adsorbed include many cases that do not cause a substantial problem in mounting.

  In consideration of the fact that abnormal adsorption patterns are roughly divided into three patterns as in the second embodiment, the number of output layers of the neural network is increased to four, thereby eliminating inconsistencies in the neural network. As a result, there is no example of misidentifying normal adsorption as abnormal adsorption, and it is possible to eliminate the loss of components due to the erroneous determination of abnormal adsorption and the cycle time loss due to the stop of the electronic component mounting machine.

  When a human sees an obliquely picked up image, it is considered that there is a characteristic part that leads to determination of “abnormal” somewhere in the image. As in the first embodiment, when the output layer is constituted by one neural network, it is presumed that the image area determined to be abnormal adsorption covers a very wide range.

  On the other hand, when the abnormal adsorption determination pattern is increased by using four output layers as in the second embodiment, in the case of the left diagonal adsorption, the reaction is performed on the image area of the left electrode portion, and conversely, In the case of diagonal adsorption, it reacts to the image area of the right electrode part, in the case of lateral adsorption, it reacts to the image area above and below the part, etc. A stable classifier can be configured.

  As in the first embodiment, when there is one output layer, the right electrode portion, the left electrode portion, and the upper and lower regions of the component are comprehensively determined, so even in the case of normal adsorption, the components are slightly different. Therefore, it is presumed that there is a possibility of being erroneously determined as abnormal adsorption cumulatively.

As a further, the present invention is not limited to the electronic component mounting apparatus configured as shown in FIG. 1, it can be carried out by applying to the electronic component mounting apparatus of various configurations.

It is a longitudinal cross-sectional view which shows the structure of the electronic component mounting machine used in Example 1, 2 of this invention. It is a figure explaining the process which extracts the component image of 32x24 [pixel] from the image imaged with the camera. It is a figure which shows the example of an image of normal adsorption | suction. It is a figure which shows the example of an image of left diagonal adsorption | suction. It is a figure which shows the example of an image of right diagonal adsorption. It is a figure which shows the example of an image of side adsorption. It is a figure which shows the number of teacher data and the number of test data about each of normal adsorption | suction and normal adsorption | suction (left diagonal adsorption, right diagonal adsorption, side adsorption). 1 is a diagram schematically illustrating a structure of a three-layer neural network according to Embodiment 1. FIG. It is a figure explaining the learning method of a neural network. It is a flowchart which shows the flow of a process of a learning program. It is a flowchart which shows the flow of a process of a components adsorption | suction attitude | position discrimination | determination program. In Example 1, it is a figure which shows the result of having measured the change of the average error rate using the neural network for every learning frequency. It is a figure which shows the discrimination experiment result by the neural network of Example 1. FIG. It is a figure which shows roughly the structure of the 3 layer type | mold neural network of Example 2. FIG. It is a figure which shows the ideal output value of the output layers 1-4 of the three-layer neural network of Example 2. FIG. In Example 2, it is a figure which shows the result of having measured the change of the average error rate using the neural network for every learning frequency. It is a figure which shows the discrimination experiment result by the neural network of Example 2. It is a figure explaining the structure of the conventional components adsorption | suction attitude | position identification system, (a) is a figure explaining normal adsorption | suction, (b) is a figure explaining diagonal adsorption | suction.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 11 ... X-axis slide, 13 ... Y-axis slide, 17 ... Adsorption nozzle, 18 ... Parts, 19 ... Back light, 20, 21 ... Reflector, 22 ... Camera, 23 Front light

Claims (4)

In the component suction orientation determination method, which picks up the part sucked by the suction nozzle of the electronic component mounting machine with a camera and determines whether the suction posture of the component is normal suction or abnormal suction by image processing technology, a large number of normal suction collected in advance The part image data and the abnormal part image data are learned by the neural network as teacher data, and the image data of the part imaged by the camera during operation of the electronic component mounting machine is input to the neural network, A component suction posture determination method for determining whether a suction posture of the component is normal suction or abnormal suction based on an output value of a neural network,
The neural network divides the component adsorption posture into four patterns: normal adsorption, abnormal adsorption by right diagonal adsorption, abnormal adsorption by left diagonal adsorption, and abnormal adsorption by lateral adsorption. Teacher data and input data for each pattern The component suction posture is characterized in that it has four output layers that output a numerical value corresponding to the degree of similarity with each other, and compares the output values of each output layer to discriminate between normal suction and abnormal suction How to determine. 2. The component suction posture determination method according to claim 1, wherein an output layer having a maximum output value is selected from the plurality of output layers to determine normal suction and abnormal suction. A camera that picks up an image of a component picked up by a pick-up nozzle of an electronic component mounting machine, and a pick-up posture determination unit that performs image processing on image data of the component picked up by the camera and determines whether the pick-up posture of the component is normal suction or abnormal suction In the component adsorption posture discrimination system with
The suction posture discriminating means learns a number of normal suction part image data and abnormal suction part image data collected in advance as neural data using a neural network, and picks up an image with the camera while the electronic component mounting machine is in operation. A component suction posture determination system that inputs image data of a component to the neural network and determines whether the suction posture of the component is normal suction or abnormal suction based on an output value of the neural network,
The neural network divides the component adsorption posture into four patterns: normal adsorption, abnormal adsorption by right diagonal adsorption, abnormal adsorption by left diagonal adsorption, and abnormal adsorption by lateral adsorption. Teacher data and input data for each pattern The component suction posture is characterized in that it has four output layers that output a numerical value corresponding to the degree of similarity with each other, and compares the output values of each output layer to discriminate between normal suction and abnormal suction Discriminating system. 4. The component suction posture determination system according to claim 3 , wherein an output layer having a maximum output value is selected from the plurality of output layers to determine normal suction and abnormal suction.

Images