JPH11101620A - Soldering inspection device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To correctly make a further detailed judgment on a soldered state without being affected by the shape or the like of an electrode. SOLUTION: An image of a soldered electrode is obtained by a camera device 8, and this image is sent to a data processor. A CPU 61 computes a luminance histogram of the image. A neural network 65 is provided and is previously put through learning with the luminance histogram of the image as an input signal and with the visual evaluation value of the soldered state of the electrode as a teacher signal. The luminance histogram of the soldered electrode to be an inspection object is inputted to the neural network 65 to obtain the evaluation value of the soldered electrode.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a soldering inspection apparatus, and it is possible to determine not only the quality of soldering but also detailed judgment information such as the quality of soldering without depending on the shape of the electrode.
The present invention relates to a soldering inspection device that can be obtained with the accuracy of a seasoned inspector.

[0002]

2. Description of the Related Art As a soldering inspection apparatus, for example, Japanese Unexamined Patent Publication (Kokai) No. 6-174444 discloses that a density (luminance) histogram is registered in advance as a standard density histogram for each soldering state, and a soldered electrode is inspected. An apparatus has been proposed in which a density histogram is created from a grayscale image, and it is determined that the image is in a soldering state corresponding to a standard density histogram having the smallest sum of frequency differences.

[0003]

However, in the above-mentioned conventional soldering inspection apparatus, more detailed soldering in process control, such as a good product but a little solder, or a good product but a little solder, is used. There is a problem that it is difficult to obtain the attached information accurately.

Further, since the density histogram changes depending on the shape and size of the electrode to be soldered, there is a problem that an erroneous determination may occur depending on the electrode shape and the like.

Accordingly, the present invention has been made to solve such a problem, and a soldering inspection apparatus capable of accurately performing a more detailed determination of a soldering state without being affected by the shape of an electrode or the like. The purpose is to provide.

[0006]

To achieve the above object, the present invention provides means (8) for obtaining an image of a soldered electrode (9) and means (6, 6) for calculating a luminance histogram of the image. 103) and a neural network (65) in which the brightness histogram of the image of the soldered predetermined electrode (9) is used as an input signal, and the evaluation value obtained by visually observing the soldering state of the predetermined electrode is learned as a teacher signal. ), The luminance histogram of the soldered electrode (9) to be inspected is input to the neural network (65) to obtain the evaluation value of the soldered electrode (9).

In the present invention, the neural network learns by associating the evaluation value obtained by visually observing the soldering state of the experienced inspector with the luminance histogram at this time, and the learning result is reproduced during the inspection. It is possible to make a detailed determination of the soldering state as a veteran inspector.

In the case where there are a plurality of types of electrodes, a plurality of neural networks are provided corresponding to each electrode, and the neural network learns the soldering state of each electrode. An accurate determination of the soldering state can be made without being affected.

If there are a plurality of types of electrodes, if the type of the electrodes is learned as an input signal of the neural network in addition to the luminance histogram, even a single neural network is affected by differences in shape and the like. And it is possible to accurately determine the soldering state of a plurality of types of electrodes.

[0010] The reference numerals in parentheses indicate the correspondence with specific means described in the embodiment described later.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION

(First Embodiment) FIG. 1 shows the overall configuration of a soldering inspection apparatus. In the figure, a lead 11 of a discrete component 1 is inserted into a through hole 21 of a printed circuit board 2 and is connected and fixed to an electrode (not shown) on the printed circuit board 2 by a soldering portion 3. The printed circuit board 2 is mounted on a moving table 4, and one of a large number of discrete components 1 soldered on the printed circuit board 2 is sequentially transferred to a position directly below the lighting device 5 by the moving table 4 and positioned. You. The moving table 4 includes an X-axis stage 41 and a Y-axis stage 42, and can position the printed circuit board 2 on a two-dimensional XY plane. The X-axis stage 41 and the Y-axis stage 42 are servo-controlled by the output of the positioning circuit 72.

The illumination device 5 includes a plurality of annular illumination rings 51.
These illumination rings 51 to 54 are arranged concentrically at intervals in the vertical direction, and have a smaller diameter as they go upward. The illumination ring 51 is divided into four sections in the circumferential direction, and a large number of light emitting diodes are provided for each section. Other illumination rings 52 to 54 have the same structure. The illumination rings 51 to 54 are respectively provided with illumination control circuits 7
1 and the lighting control circuit 71
The light emitting diode group can be energized to emit light for each of the 54 sections. The uppermost illumination ring 54 is used when the printed circuit board 2 is aligned, and the remaining lower illumination rings 51 to 53 are used in the soldering inspection.

The discrete component 1 has its leads 11
(I.e., the through holes 21) are positioned on the central axis of the illumination device 5, and the illumination light emitted from each of the illumination rings 51 to 53 enters the soldered portion 3 of the lead 11 at a different angle. The incident angle of each illuminating light becomes larger as it is emitted from the upper illuminating ring 53, and is set in the range of 3 ° to 70 ° in the present embodiment.
As shown in FIG. 2, the “normal” soldering portion 3 has a mountain-shaped cross section rising from a gentle skirt portion to a steep top portion, and light L incident on the printed circuit board 2 at a large angle.
Numeral 1 is reflected at the foot of the soldering portion 3 and goes upward, and light L2 incident at a small angle is reflected at the top and goes upward. Therefore, in the “normal” soldered portion 3, high-luminance reflected light is generated almost upward on the entire surface of the electrode 9.

FIG. 3 shows the soldering part 3 in the case of "insufficient soldering".
In this case, in this case, the cross section of the soldered portion 3 is reduced in a similar manner as compared with the normal case, and the area where upward reflected light of high luminance is greatly reduced. FIG. 4 is a cross-sectional view of the “perforated” soldering portion 3 in which the through-hole 21 is partially exposed. In this case, in the region where the solder is attached, reflected light of high luminance is generated upward, The portion without solder has very low brightness. FIG. 5 is a cross section of the soldering portion 3 of “excessive soldering”. In this case, the inner peripheral portion of the soldering portion 3 has a flat surface and the outer peripheral portion has a steep surface.
The area of the moderately inclined surface that produces reflected light of high luminance going upward is also reduced. As described above, the intensity and distribution of the reflected light differ depending on the state of soldering.

In FIG. 1, an image pickup device 8 such as a CCD camera is provided above a lighting device 5 on its central axis.
The imaging device 8 receives reflected light upward from the soldering section 3 to obtain an image thereof. The image obtained by the imaging device 8 is sent to the data processing device 6. The data processing device 6 internally has input and output interfaces 62 and 6.
3, various memories 64 including a CPU 61, a video RAM,
And an input device 66 such as a keyboard and a display device 67 are connected. The CPU 61 controls the operations of the illumination control circuit 71 and the positioning circuit 72 described above, makes the neural network 65 learn according to a procedure described later, and evaluates the soldering using the learned neural network 65.

FIG. 6 shows the configuration of the neural network. The neural network 65 actually has A, A in accordance with the type of the electrode to be soldered, such as the shape and size.
B,..., N, and each of the neural networks 65A to 65N is composed of a neuron element network composed of neuron elements 654 for signal propagation and a learning control unit (not shown) for controlling learning thereof. . The neuron element network includes an input layer 651 to which an input signal is input, an output layer 653 to which an output signal is output, and both layers 651.
1 and 653, and is composed of an intermediate layer 652,
The neuron elements 654 of each of the layers 651 to 653 are connected to the neuron elements of the other layers with a predetermined strength (connection weight). The output signal changes due to the difference in the value of the connection weight. In the neural networks 65A to 65C configured in such a hierarchical structure, each neuron element 6
A process called “learning” is performed by changing the connection weight between the 54 by the learning control unit.

The output function of the input layer 651 can be, for example, a linear function, and the output functions of the intermediate layer 652 and the output layer 653 can be, for example, a sigmoid function. Further, the number of neuron elements 654 of the input layer 651 is provided by a number equal to the luminance value (for example, 0 to 255) of a luminance histogram described later, and the number of pixels at the corresponding luminance value is given to each neuron element 654 as an input signal. Middle layer 65
The number of the two neuron elements 654 is determined depending on the relationship with the determination accuracy, but is usually 10 to 20. Output layer 653
The number of neuron elements 654 is one when the determination value is one type as in the present embodiment.

Hereinafter, the learning procedure of the neural network will be described with reference to the flowchart of FIG. In step 101 of FIG. 7, an image obtained by the imaging device 8 is captured. In the following step 102, as shown in FIG. 8, a rectangular inspection window L is set around the soldering portion 3, and the soldering portion 3 (the electrode 9 on the printed circuit board 2) in the inspection window L is set. A masking process is performed to replace the area excluding the area with low-luminance pixels (hatched lines in FIG. 8). This is to prevent the pixel luminance value around the soldering section 3 from affecting the processing in the luminance smoothing processing described later. That is, since the processing result of the brightness smoothing process differs between the case where the periphery of the soldering portion 3 is a dark base material and the case where the wiring pattern is relatively bright, the influence thereof is eliminated. In the above image, lead 1
The central region N of the soldering portion 3 where the tip of the soldering portion 1 is exposed (see FIG. 2) has a low luminance region because there is almost no reflected light upward.

In step 103 of FIG. 7, a brightness smoothing process for smoothing the brightness of each part of the masked soldering image is performed. This luminance smoothing process is performed by, for example, a moving average method. The moving average method will be described with reference to FIG. 9. The luminance value of the pixel of interest P4 is replaced with the average value of the luminance values of eight pixels P0 to P3 and P5 to P8 around it, This is performed sequentially for pixels.
Note that, in addition to taking the average of the luminance values from eight neighboring pixels in the 3 × 3 pixel area as described above, the average of the luminance values is 2 in the 5 × 5 pixel area.
It can also be taken from four neighboring pixels. Such luminance smoothing processing is repeated about three times. The entire image is blurred by the smoothing process, and a narrow high-luminance portion in the outer peripheral boundary region of the soldering portion 3 is blurred, so that the influence on the luminance histogram is reduced.

In step 104 of FIG. 7, a luminance histogram is created for the image that has been subjected to the luminance smoothing process. This luminance histogram is obtained by taking the luminance on the horizontal axis and the number of pixels on the vertical axis. If the luminance histogram has severe irregularities, it is advisable to average the neighbors at three points before and after. FIG. 10 shows an example of the luminance histogram. In FIG. 10, a solid line X is a luminance histogram of a normal soldering portion, and a dashed line Y is of a soldering portion of "low soldering".

In step 105, the input layer 651 of the predetermined neural network 65A corresponding to the type of the electrode among the plurality of neural networks 65A to 65N provided as described above according to the type of the electrode to be soldered.
, The number of pixels for each luminance value is given as an input signal. The selection of the neural networks 65A to 65N is made based on a command signal or the like input from the input device 66 (FIG. 1). In the following step 106, the visual evaluation value previously performed by the experienced inspector on the soldering part 3 of the electrode 9 to be learned is stored in the neural network 65.
A is given to A as a teacher signal. In step 107, the connection weight between the neuron elements 654 in the neural network 65A is corrected so that the output signal matches the teacher signal. As a specific method of the correction, a known error back propagation method or the like is used.

In step 108, it is checked whether or not the output signal (evaluation value) of the neural network 65A matches the visual evaluation value.
Repeat step 07. When the output evaluation value matches the visual evaluation value, the learning is completed. Such learning is sequentially performed for all the neural networks 65A to 65N for various types of soldering states for each type of electrode 9. An example of each soldering state and its evaluation value is shown below.

(1) No solder 0.0 ≦ Evaluation value ≦ 0.1 (2) Under-soldering 0.1 <Evaluation value ≦ 0.3 (3) Underslightly 0.3 <Evaluation value ≦ 0.4 (4) ) Perfect non-defective product 0.4 <Evaluation value ≦ 0.6 (5) Excessive 0.6 <Evaluation value ≦ 0.7 (6) Excessive solder 0.7 <Evaluation value ≦ 0.9 (7) Excessive soldering 0 .9 <evaluation value ≦ 1.0

In the case of performing the soldering inspection using the neural network trained as described above, the soldering part 3 of the discrete component 1 to be inspected is transferred by the moving table 4 directly below the lighting device 5. Then, a luminance histogram of the soldered image is obtained in the same process as in steps 101 to 104 described above. When this luminance histogram is given as an input signal to one of the neural networks 65A to 65N corresponding to the type of the electrode to be inspected, the above-mentioned (1) to (1) to
The evaluation value shown in (7) is obtained. From this evaluation value,
Seven types of soldering modes from “no solder” to “excessive solder” can be determined.

As described above, according to the present embodiment, the neural networks 65A to 65N are trained so as to match the visual evaluation by the experienced inspector, and the state of soldering is evaluated using this. More detailed determinations of the above (1) to (7) can be performed as well as the excessive amount and the insufficient amount of the solder. Further, in the present embodiment, a plurality of neural networks are provided, and the neural network is selected according to the type of the electrode to be soldered to determine the soldering state. And accurate determination can always be made.

(Second Embodiment) In the first embodiment,
Although a plurality of neural networks are provided in accordance with the type of the electrode to be soldered, a single neural network 65 is used as shown in FIG. While inputting the number, the type of the electrode 9 is input as the input signal 2. In the present embodiment, the number of neuron elements 654 in the input layer 651 needs to be a number corresponding to the luminance value and the type of the electrode. However, when there are many types of the electrode 9, as in the first embodiment, The overall configuration can be simplified as compared with providing a neural network for each type.

(Third Embodiment) To make a more detailed judgment, as shown in FIG. 12, there are two types of teacher signal and output signal, and the teacher signal 1 and the output signal 1 are the same as those described above with "no solder". "To" Excessive solder "
For the teacher signal 2 and the output signal 2, the more detailed degree under each determination from “no solder” to “excessive solder” is, for example, an evaluation value evaluated in 10 steps. Thereby, the soldering state can be evaluated in more detail without depending on the type of the electrode.

[0028]

As described above, according to the soldering inspection apparatus of the present invention, a more detailed determination of the soldering state can be accurately performed without being affected by the shape of the electrodes.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating an overall configuration of a soldering inspection device according to a first embodiment of the present invention.

FIG. 2 is a vertical sectional view of a normal soldering portion.

FIG. 3 is a vertical cross-sectional view of a soldering portion of “insufficient solder”.

FIG. 4 is a vertical sectional view of a “perforated” soldered portion.

FIG. 5 is a vertical sectional view of a soldering portion of “excessive soldering”.

FIG. 6 is a schematic diagram showing a configuration of a neural network.

FIG. 7 is a flowchart showing a processing procedure in a CPU at the time of neural network learning.

FIG. 8 is a photographed image of a soldering portion.

FIG. 9 is an explanatory diagram of a luminance smoothing process.

FIG. 10 is a diagram showing a luminance histogram.

FIG. 11 is a schematic diagram illustrating a configuration of a neural network according to a second embodiment of the present invention.

FIG. 12 is a schematic diagram illustrating a configuration of a neural network according to a third embodiment of the present invention.

[Explanation of symbols]

3 soldering unit, 61 CPU (means for calculating a luminance histogram), 65, 65A, 65B, 65C neural network, 8 image pickup device (means for obtaining an electrode image), 9 electrodes.

[Procedure amendment]

[Submission date] December 8, 1997

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Name of invention

[Correction method] Change

[Correction contents]

[Title of the Invention] Soldering inspection device

Claims (1)

[Claims] A means for obtaining an image of a soldered electrode; a means for calculating a brightness histogram of the image; and for a predetermined soldered electrode, the brightness histogram of the image as an input signal; A neural network trained on the evaluation value of the soldered state as a teacher signal, and inputting a luminance histogram of the soldered electrode to be inspected to the neural network and A soldering inspection device characterized by obtaining an evaluation value.