JP2023153608A - Welding inspection apparatus - Google Patents

Abstract

To easily realize a correct inspection of the welding quality.SOLUTION: A welding inspection device 10 comprises: a storage device 11 which stores a learned neural network 13 with image data of a surface of a weld zone P as an input and a dimension value of a cross-sectional shape of the weld zone P as an output; and an execution device 12 which calculates an output of the neural network 13 with the image data of the surface of the weld zone P being an inspection target as an input as a dimension value of the cross-sectional shape of the weld zone P. The neural network 13 is learned by using a data set including a measurement value of the dimension value of the cross-sectional shape of the weld zone P and the image data of the surface of the weld zone P in which the same dimension value is measured as teacher data.SELECTED DRAWING: Figure 1

Description

The present invention relates to a welding inspection device for inspecting welding processing quality.

Patent Document 1 describes a welding device that includes a visual camera that images the weld bead during welding. During welding, the welding device extracts the outline of the weld bead from the image data captured by the visual camera. The welding apparatus then corrects the positional deviation of the weld bead based on the extracted weld bead outline.

Japanese Unexamined Patent Publication No. 1-170583

The quality of welding can be accurately confirmed by looking at the cross-sectional shape of the weld. However, to check the cross-sectional shape of the welded part, it is necessary to cut the welded part. Therefore, the welding quality of products is generally ensured by controlling welding conditions such as outside temperature, welding current, arc voltage, and wire feed amount. However, in order to select appropriate welding conditions, it is necessary to investigate the cross-sectional shapes of a large number of welded parts under different welding conditions. Therefore, the selection process requires a large amount of man-hours.

A welding inspection device that solves the above problems includes a storage device that stores a trained neural network that receives image data of the surface of a welded part as input and outputs dimension values of the cross-sectional shape of the welded part, and an inspection target. and an execution device that calculates the output of the neural network input with image data of the surface of the welded portion as a dimension value of the cross-sectional shape of the welded portion. The above-mentioned neural network is trained using, as training data, a data set including measured values of the dimensions of the cross-sectional shape of the weld and image data of the surface of the weld where the same dimensions were measured.

The difference in the cross-sectional shape of the weld is also reflected in the surface texture of the weld. Therefore, by using the above data set as training data, it becomes possible to learn a neural network that receives image data of the surface of the welded part as input and outputs dimensional values of the cross-sectional shape of the welded part. Then, by using the neural network learned in this way, it becomes possible to obtain the dimension values of the cross-sectional shape of the welded part without actually cutting it. Therefore, the welding inspection apparatus described above has the effect of easily realizing accurate welding quality inspection.

1 is a diagram schematically showing the configuration of an embodiment of a welding inspection device. It is a figure which shows an example of the image of the welding part surface. FIG. 2 is a diagram schematically showing the configuration of a trained neural network stored in a storage device of the welding inspection device. It is a sectional view of a welding part.

Hereinafter, one embodiment of the welding inspection device will be described in detail with reference to FIGS. 1 to 4.
<Configuration of welding inspection device 10>
First, with reference to FIG. 1, the configuration of a welding inspection apparatus 10 of this embodiment will be described. The welding inspection device 10 of this embodiment inspects a welded portion P between an engine exhaust manifold and a catalytic converter.

The welding inspection device 10 includes an execution device 11 that executes arithmetic processing for inspecting the welded portion P. The welding inspection device 10 also includes a storage device 12 that stores inspection programs and a neural network 13. An imaging device 14, a display device 15, and an input device 16 are connected to the welding inspection device 10. The imaging device 14 is a CCD camera device that captures an image of the surface of the weld P as shown in FIG. 2 during a welding inspection. The display device 15 is a device for displaying test results and the like. The input device 16 is a device for manually inputting test data and the like. An example of the imaging device 14 is a CCD camera, an example of the display device 15 is a display, and an example of the input device 16 is a keyboard.

<Configuration of neural network 13>
Next, the configuration of the neural network 13 stored in the storage device 12 will be explained with reference to FIG. The neural network 13 includes an input layer consisting of N nodes, an intermediate layer consisting of M nodes, and an output layer consisting of L nodes. The input values X1, X2, ..., XN of each node of the input layer are data of welding conditions and image data of the surface of the welded part. The welding condition data includes, for example, outside temperature, welding current, arc voltage, and wire feed amount. On the other hand, the output values Y1, Y2, .

Here, input values to each node of the middle layer are U1, U2, ..., UM, and output values of each node of the middle layer are Z1, Z2, ..., ZM. Further, let "i" be a natural number from "1" to "N", "j" be a natural number from "1" to "M", and "k" be a natural number from "1" to "L". The input value Uj of each node of the intermediate layer is calculated as the sum of the input values X1, X2, . . . , XN of the input layer multiplied by a weight Vij. The output value Zj of each node in the intermediate layer is calculated as the return value of the activation function F that uses the input value Uj of the node as an argument. An example of activation function F is a sigmoid function. The output values Y1, Y2, . . . , YL of each node in the output layer are calculated as the sum of the output values Zj of each node in the intermediate layer multiplied by a weight Wjk.

FIG. 4 shows the cross-sectional shape of the weld P. In this embodiment, the bead height, bead width, and flank angle shown in the figure are used as the dimensional values of the cross-sectional shape of the welded portion P, which are the output values Y1, Y2, . . . , YL of the neural network 13.

<Learning of neural network 13>
Next, learning of the neural network 13 will be explained. Learning of the neural network 13 is performed using a data set constituted by image data of the surface of the weld P, data of welding conditions, and data of dimension values of the cross-sectional shape of the weld P as training data. In addition, in this embodiment, the welding inspection apparatus 10 performs learning of the neural network 13.

The creation of the training data dataset is performed in the following steps. Perform welding of the exhaust manifold and catalytic converter. The welding conditions at that time are inputted from the input device 16, and the surface of the welded portion P is imaged by the imaging device 14. The welded exhaust manifold and catalytic converter are cut, and a cross section of the welded portion P is imaged by the imaging device 14. As described above, data on welding conditions, image data on the surface of the welded portion P, and image data on the cross-sectional shape of the welded portion P are input to the welding inspection apparatus 10. The welding inspection device 10 acquires dimensional values of the cross-sectional shape of the weld P by image analysis of the cross-sectional shape. Then, the welding inspection device 10 creates a data set that includes welding condition data, image data of the surface of the welded portion P, and dimension values of the cross-sectional shape of the welded portion P.

The welding inspection device 10 performs learning of the neural network 13 by performing the following processing for each data set. That is, the welding inspection device 10 inputs the welding condition data in the data set and the image data of the surface of the welded part P to the neural network 13 as input values X1 to XN of the input layer. Then, the welding inspection device 10 calculates the output values Y1 to YL of the neural network 13 in response to the input. Thereafter, the welding inspection device 10 uses the error back propagation method to adjust each weight Vij so that the errors between the calculated output values Y1 to YL and the dimension values of the cross-sectional shape of the weld P in the data set become small. , Wjk are corrected. Then, the welding inspection apparatus 10 performs learning of the neural network 13 by repeating the correction processing of the weights Vij and Wjk until the above-mentioned error becomes equal to or less than a predetermined value. Note that such learning of the neural network 13 may be performed in a device other than the welding inspection device 10, and the learned neural network 13 may be stored in the storage device 12.

<Welding inspection>
Next, a welding inspection procedure using the neural network 13 learned in this way will be explained. When inspecting the weld P, welding conditions are input through the input device 16, and the surface of the weld P is imaged by the imaging device 14. The execution device 11 of the welding inspection device 10 calculates the output of the neural network 13, which receives the welding condition data and the image data of the surface of the welded portion P, as a dimension value of the cross-sectional shape of the welded portion P. Then, the welding inspection device 10 displays the calculation result on the display device 15.

<Actions and effects of embodiments>
The operation and effects of this embodiment will be explained.
The quality of welding can be determined by the cross-sectional shape of the welded portion P. On the other hand, the difference in the cross-sectional shape of the welded part P also appears in the surface texture of the welded part P. Therefore, by using the above data set as training data, it becomes possible to learn the neural network 13 which inputs the image data of the surface of the weld P and outputs the dimension value of the cross-sectional shape of the weld P. Then, by using the neural network 13 learned in this way, it becomes possible to obtain the dimension values of the cross-sectional shape of the welded part P without actually cutting it.

According to the welding inspection apparatus 10 of this embodiment described above, the following effects can be achieved.
(1) Accurate evaluation of welding quality based on the dimensional value of the cross-sectional shape of the welded part P becomes possible. In addition, in the welding inspection apparatus 10 of the said embodiment, the dimension value of a cross-sectional shape can be acquired without actually cutting the welded part P, or using a non-destructive inspection apparatus using an ultrasonic wave or magnetism. Therefore, welding inspections as described above can be easily carried out.

(2) The image data of the surface of the welded part P as well as the data of the welding conditions are input to the neural network 13. Therefore, the same neural network 13 can inspect welded parts P under different welding conditions.

(3) When creating a data set for training data, dimensional values of the cross-sectional shape of the welded portion P are obtained by image analysis of the cross-section. Therefore, it becomes easy to obtain the measured values of the dimensions of the cross-sectional shape when creating the teacher data.

<Other embodiments>
This embodiment can be modified and implemented as follows. This embodiment and the following modified examples can be implemented in combination with each other within a technically consistent range.

- In the above embodiment, when creating a dataset for teacher data, the measurement of the dimension value of the cross-sectional shape of the welded part P is obtained by image analysis of image data of the same cross-section. These dimension values may be measured using other methods.

- The welding inspection device 10 of the above embodiment may be used to inspect welds other than the welds P of the exhaust manifold and the catalytic converter.
- The welding inspection device 10 may be connected to a welding processing device, and data on welding conditions may be automatically input from the welding processing device to the welding inspection device 10.

- The type and number of parameters input into the welding inspection device 10 as welding condition data may be changed. Furthermore, the input to the neural network 13 may not include data on welding conditions.

- The neural network 13 may be configured so that the dimensional values of the cross-sectional shape of the welded part P other than those mentioned above, such as penetration depth and penetration width, are set as output values Y1 to YL.
- The neural network 13 may have a configuration having two or more intermediate layers.

- A function other than the sigmoid function may be used as the activation function F used in the neural network 13.

10...Welding inspection device 11...Execution device 12...Storage device 13...Neural network 14...Imaging device 15...Display device 16...Input device P...Welding part

Claims (1)

a storage device storing a trained neural network that receives image data of the surface of the welded portion as input and outputs dimension values of the cross-sectional shape of the welded portion;
an execution device that calculates the output of the neural network input with image data of the surface of the welded part to be inspected as a dimension value of the cross-sectional shape of the welded part;
It is equipped with
The neural network is trained using, as training data, a data set that includes measured dimensions of the cross-sectional shape of the weld and image data of the surface of the weld where the same dimensions were measured. Inspection equipment.

Images