IP Library › Granted Patent US 12670574
Granted Patent B2
US 12670574 · App. 18/540,123 · Granted Jun 30, 2026

Apparatus and method for predicting weld quality

Inventors: Hyeong Won Kim (Daejeon, KR); Jae Hun Kim (Daejeon, KR); Hyeon Jeong You (Daejeon, KR); Ki Dong Lee (Daejeon, KR)
Assignee: SK On Co., Ltd.
G06T7/0004G01N21/8851G06T7/50G06V10/764G06V10/7715G06V10/774G06V20/50G01N2021/8887G06T2207/20081G06T2207/20084G06T2207/30164G06V10/82
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12670574
App. No.
18/540,123
Granted
Jun 30, 2026
Kind
B2
Abstract

An apparatus for predicting weld quality includes an image sensor acquiring an image of a welding portion of a workpiece, a storage storing a shape detection model, a depth prediction model and a strength prediction model, and a controller configured to acquire shapes of a melt pool and a keyhole by receiving the image and inputting the image to the shape detection model, acquire a penetration depth of the keyhole by inputting the shapes of the melt pool and the keyhole to the depth prediction model, and acquire a tensile strength by inputting one of the shape of the melt pool, the shape of the keyhole and the shapes of the melt pool and the keyhole and the penetration depth of the keyhole to the strength prediction model. The weld quality of all products may be predicted during a welding process.

Claims (36)

1 . An apparatus for predicting weld quality, the apparatus comprising:

an image sensor acquiring an image of a welding portion of a workpiece;

a storage storing a shape detection model, a depth prediction model and a strength prediction model; and

a controller configured to acquire shapes of a melt pool and a keyhole by receiving the image and inputting the image to the shape detection model, acquire a penetration depth of the keyhole by inputting the shapes of the melt pool and the keyhole to the depth prediction model, and acquire a tensile strength by inputting the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole, or the shapes of both the melt pool and the keyhole to the strength prediction model,

wherein the shape detection model is an artificial intelligence model learned using a first learning dataset in which the image of the welding portion is first learning data and the shapes of the melt pool and the keyhole are first label data, and when the image of the welding portion is received, detects and outputs the shapes of the melt pool and the keyhole,

wherein the depth prediction model is an artificial intelligence model learned using a second learning dataset in which the shapes of the melt pool and the keyhole are second learning data and the penetration depth of the keyhole is second label data, and when the shapes of the melt pool and the keyhole are inputted, predicts and outputs the penetration depth of the keyhole,

wherein the strength prediction model is an artificial intelligence model learned using a third learning dataset in which the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole or the shapes of both the melt pool and the keyhole are third learning data and the tensile strength of the welding portion is third label data, and when the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole or the shapes of both the melt pool and the keyhole are inputted, predicts and outputs the tensile strength of the welding portion.

2 . The apparatus of claim 1 , wherein the shape detection model is an artificial intelligence model configured by a backbone structure in which feature maps of various scales are extracted while passing the image through multiple convolution layers, a neck structure in which the scales of the feature maps extracted in the backbone structure are up-sampled or down-sampled and combined, and a head structure in which classification and detection are performed using the feature maps combined in the neck structure.

3 . The apparatus of claim 1 , wherein the shape of the melt pool includes a position and size of the melt pool, and

the shape of the keyhole includes a position and size of the keyhole.

4 . The apparatus of claim 1 , wherein the tensile strength is acquired at each point from a welding start point to a welding end point of the workpiece.

5 . The apparatus of claim 1 , wherein the depth prediction model has a convolutional neural network structure which extracts features, and is an artificial intelligence model which extracts a scalar value by applying a linear function in a last layer of the convolutional neural network structure.

6 . The apparatus of claim 1 , wherein the strength prediction model has a deep neural network structure which extracts features of the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole or the shapes of both the melt pool and the keyhole, and is an artificial intelligence model which extracts a scalar value by applying a linear function in a last layer of the deep neural network structure.

7 . The apparatus of claim 1 , further comprising:

a distance sensor measuring the penetration depth of the keyhole of the welding portion in a process of welding a sample of the workpiece,

wherein the controller stores the measured penetration depth of the keyhole in the storage as label data for learning the depth prediction model.

8 . A method for predicting weld quality, the method comprising:

a model generation step of generating a shape detection model, a depth prediction model and a strength prediction model by learning artificial intelligence models using learning datasets; and

a quality prediction step of acquiring shapes of a melt pool and a keyhole by inputting an image of a welding portion of a workpiece to the shape detection model, acquiring a penetration depth of the keyhole by inputting the shapes of the melt pool and the keyhole to the depth prediction model and acquiring a tensile strength by inputting the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole, or the shapes of both the melt pool and the keyhole to the strength prediction model,

wherein the shape detection model is an artificial intelligence model learned using a first learning dataset in which the image of the welding portion is first learning data and the shapes of the melt pool and the keyhole are first label data, and when the image of the welding portion is received, detects and outputs the shapes of the melt pool and the keyhole,

wherein the depth prediction model is an artificial intelligence model learned using a second learning dataset in which the shapes of the melt pool and the keyhole are second learning data and the penetration depth of the keyhole is second label data, and when the shapes of the melt pool and the keyhole are inputted, predicts and outputs the penetration depth of the keyhole,

wherein the strength prediction model is an artificial intelligence model learned using a third learning dataset in which the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole or the shapes of both the melt pool and the keyhole are third learning data and the tensile strength of the welding portion is third label data, and when the penetration depth of the keyhole and one of the shape of the melt pool, the shape of the keyhole or the shapes of both the melt pool and the keyhole are inputted, predicts and outputs the tensile strength of the welding portion.

9 . The method of claim 8 , wherein the model generation step comprises:

a sample image sensing step of acquiring an image of the welding portion while welding the workpiece;

a distance sensing step of measuring the penetration depth of the keyhole formed in the welding portion;

a step of generating the shape detection model using the first learning dataset;

a step of generating the depth prediction model using the second learning dataset;

a step of generating the strength prediction model using the third learning dataset.

10 . The method of claim 8 , wherein the quality prediction step comprises:

an image sensing step of acquiring the image of the welding portion of the workpiece;

a shape detection step of acquiring the shapes of the melt pool and the keyhole by inputting the image to the shape detection model;

a depth prediction step of acquiring the penetration depth of the keyhole by inputting the shapes of the melt pool and the keyhole to the depth prediction model; and

a strength prediction step of acquiring a tensile strength by inputting the penetration depth of the keyhole acquired in the depth prediction step and one of the shape of the melt pool, the shape of the keyhole and the shapes of both the melt pool and the keyhole acquired in the shape detection step, to the strength prediction model.

11 . The method of claim 8 , wherein the shape of the melt pool includes a position and size of the melt pool, and

the shape of the keyhole includes a position and size of the keyhole.

12 . The method of claim 10 , wherein the tensile strength is acquired at each point from a welding start point to a welding end point of the workpiece.