CN-122023263-A - Welding quality prediction method, model training method, device and chip

CN122023263ACN 122023263 ACN122023263 ACN 122023263ACN-122023263-A

Abstract

The invention belongs to the field of laser welding, and discloses a welding quality prediction method, a model training device and a chip, wherein the method comprises the steps of extracting welding spots based on a post-welding image to form a welding spot diagram; and inputting the neural network characteristics and the image characteristics into a welding quality prediction model to obtain a welding quality prediction value corresponding to the post-welding image. According to the invention, the welding quality prediction value of the welded materials corresponding to the welded image is predicted through the dual feature analysis of the welding quality prediction model, so that the accuracy of the welding quality prediction value is improved, and the accuracy of welding quality detection is further improved.

Inventors

LIN MENGCHEN
LIN ZONGRU
WU ZHENTING
WU YUE
LI RENBO
Yang Yaan

Assignees

富联裕展科技(深圳)有限公司

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (15)

1. A method of weld quality prediction, comprising: extracting welding spots based on the post-welding image to form a welding spot diagram; determining neural network characteristics and image characteristics corresponding to the post-welding images based on the welding spot diagram; and inputting the neural network characteristics and the image characteristics into a welding quality prediction model to obtain a welding quality prediction value corresponding to the post-welding image.
2. The method of claim 1, wherein the determining the neural network characteristics corresponding to the post-weld image based on the solder joint map comprises: Performing dimension reduction and feature extraction processing on the welding spot diagram based on a feature extraction model to obtain neural network features corresponding to the welding spot diagram; And determining the neural network characteristics corresponding to the post-welding image based on the neural network characteristics corresponding to the welding point diagram.
3. The method of claim 2, wherein the training method of the feature extraction model comprises: based on the history post-welding image, extracting a history welding spot to form a history welding spot diagram; projecting the historical welding spot diagram to a feature vector space to obtain a feature value; restoring the characteristic values in the characteristic vector space to obtain a reconstructed image; and determining the difference degree between the reconstructed image and the historical welding spot diagram, and training based on the difference degree input model to form the feature extraction model.
4. The method of claim 3, wherein the performing dimension reduction and feature extraction processing on the solder joint map based on the feature extraction model to obtain the neural network feature corresponding to the solder joint map includes: projecting the welding spot diagram to a feature vector space of the feature extraction model to obtain a feature value of the welding spot diagram; and determining the characteristic value of the welding point diagram as the neural network characteristic corresponding to the welding point diagram.
5. The method of claim 1, wherein the post-weld image has a plurality of welds, and wherein the determining image features corresponding to the post-weld image based on the weld map comprises: determining the total number of welding spots and the width and the height of each welding spot based on the welding spot graph; Determining a solder joint diameter based on the total number of solder joints and the width and height of each solder joint; and determining the diameter of the welding spot as the image characteristic corresponding to the post-welding image.
6. The method of claim 1, wherein the determining the image features corresponding to the post-weld image based on the solder joint map comprises: fitting the gray value of each pixel in the welding point diagram by using a preset probability density function to obtain the scale parameter and the shape parameter of the preset probability density function; Determining the gray distribution mode of the welding spot graph based on the scale parameter and the shape parameter of the preset probability density function; And determining the image characteristics corresponding to the post-welding image based on the gray distribution mode of the welding point diagram.
7. The method of claim 6, wherein the method further comprises: determining the gray standard deviation of the welding spot graph based on the scale parameter and the shape parameter of the preset probability density function; And determining the image features corresponding to the post-welding images based on the gray standard deviation of the welding point diagram.
8. The method of claim 1, wherein the determining the image features corresponding to the post-weld image based on the solder joint map comprises: Determining a gray level co-occurrence matrix of the welding spot diagram based on a gray level co-occurrence matrix algorithm; Determining texture complexity of the welding spot map based on the gray level co-occurrence matrix; and determining the image features corresponding to the post-welding images based on the texture complexity of the welding point diagram.
9. The method of any of claims 1-8, wherein the weld quality prediction value comprises a pullout force prediction value and a reliability of the pullout force prediction value, and wherein the weld quality prediction model comprises a gaussian process regression model; Inputting the neural network characteristics and the image characteristics into a welding quality prediction model to obtain a welding quality prediction value corresponding to the post-welding image, wherein the method comprises the following steps of: Performing standardization processing on the neural network characteristics and the image characteristics to obtain input values; calculating the mean value and variance of the input value through the kernel function of the Gaussian process regression model; And determining the mean value as a welding quality predicted value corresponding to the post-welding image, and determining the credibility of the welding quality predicted value based on the variance.
10. An apparatus for weld quality prediction, comprising: the welding spot extraction module is used for extracting welding spots based on the post-welding image to form a welding spot diagram; The characteristic module is used for determining neural network characteristics and image characteristics corresponding to the post-welding images based on the welding spot diagram; And the prediction module is used for inputting the neural network characteristics and the image characteristics into a welding quality prediction model to obtain a welding quality prediction value corresponding to the post-welding image.
11. A method of training a weld quality prediction model, comprising: Extracting welding spots based on the post-welding image to form a training welding spot diagram; Determining neural network characteristics and image characteristics based on the training weld points map; Forming a training sample by combining the neural network characteristics and the image characteristics of the training weld spot diagram based on the drawing force test value corresponding to the training weld spot diagram; constructing a training sample set by adopting the training samples corresponding to each training weld point diagram; and training an initial welding quality prediction model by adopting the training sample set, and taking the initial welding quality prediction model after iteration as the welding quality prediction model.
12. The training method of claim 11, wherein the initial weld quality prediction model is a gaussian process regression model; the training of the initial welding quality prediction model by the training sample set comprises the following steps: Performing standardization processing on all the training samples in the training sample set; Determining a kernel function of the gaussian process regression model; and optimizing the hyper-parameters of the kernel function by maximizing an edge likelihood function based on the standardized training samples to form the welding quality prediction model.
13. The training method of claim 12, wherein said determining a kernel of said gaussian process regression model comprises: judging whether a kernel function selection type input by a user is received or not; If yes, a kernel function of the Gaussian process regression model is determined based on the kernel function selection type; if not, determining the kernel function of the Gaussian process regression model based on a preset kernel function type.
14. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when run on a computer, causes the computer to perform the method of any of claims 1-9 or 11-13.
15. A chip comprising a processor and a memory for storing at least one instruction that, when loaded and executed by the processor, implements the method of weld quality prediction as claimed in any one of claims 1 to 9 or the training method of the weld quality prediction model as claimed in any one of claims 11 to 13.

Description

Welding quality prediction method, model training method, device and chip Technical Field The invention relates to the field of laser welding, in particular to a welding quality prediction method, a model training device and a chip. Background In modern industrial production, welding quality detection is a key link for ensuring product quality and safety. Traditional welding quality detection methods rely primarily on destructive testing, such as pullout force testing. These methods require physical destruction of the welded material to measure its strength and reliability, however, such methods are time consuming and laborious, and costly, and can only detect a small number of samples, failing to achieve comprehensive quality monitoring. Therefore, how to improve the efficiency of welding quality detection and reduce the detection cost is a technical problem to be solved by those skilled in the art. Disclosure of Invention In order to solve the problems of low welding quality detection efficiency and high cost in the prior art, the invention provides a welding quality prediction method, a model training device and a chip. A method of weld quality prediction, comprising: extracting welding spots based on the post-welding image to form a welding spot diagram; determining neural network characteristics and image characteristics corresponding to the post-welding images based on the welding spot diagram; and inputting the neural network characteristics and the image characteristics into a welding quality prediction model to obtain a welding quality prediction value corresponding to the post-welding image. Optionally, the determining, based on the solder joint map, a neural network feature corresponding to the post-solder image includes: Performing dimension reduction and feature extraction processing on the welding spot diagram based on a feature extraction model to obtain neural network features corresponding to the welding spot diagram; And determining the neural network characteristics corresponding to the post-welding image based on the neural network characteristics corresponding to the welding point diagram. Optionally, the training method of the feature extraction model includes: based on the history post-welding image, extracting a history welding spot to form a history welding spot diagram; projecting the historical welding spot diagram to a feature vector space to obtain a feature value; restoring the characteristic values in the characteristic vector space to obtain a reconstructed image; and determining the difference degree between the reconstructed image and the historical welding spot diagram, and training based on the difference degree input model to form the feature extraction model. Optionally, based on a feature extraction model, performing dimension reduction and feature extraction processing on the solder joint diagram to obtain a neural network feature corresponding to the solder joint diagram, including: projecting the welding spot diagram to a feature vector space of the feature extraction model to obtain a feature value of the welding spot diagram; and determining the characteristic value of the welding point diagram as the neural network characteristic corresponding to the welding point diagram. Optionally, the post-welding image has a plurality of welding spots, and the determining, based on the welding spot diagram, the image features corresponding to the post-welding image includes: determining the total number of welding spots and the width and the height of each welding spot based on the welding spot graph; Determining a solder joint diameter based on the total number of solder joints and the width and height of each solder joint; and determining the diameter of the welding spot as the image characteristic corresponding to the post-welding image. Optionally, the determining, based on the solder joint map, the image feature corresponding to the post-welding image includes: fitting the gray value of each pixel in the welding point diagram by using a preset probability density function to obtain the scale parameter and the shape parameter of the preset probability density function; Determining the gray distribution mode of the welding spot graph based on the scale parameter and the shape parameter of the preset probability density function; And determining the image characteristics corresponding to the post-welding image based on the gray distribution mode of the welding point diagram. Optionally, the method further comprises: determining the gray standard deviation of the welding spot graph based on the scale parameter and the shape parameter of the preset probability density function; And determining the image features corresponding to the post-welding images based on the gray standard deviation of the welding point diagram. Optionally, the determining, based on the solder joint map, the image feature corresponding to the post-welding image includes: Determining a gray level co-occ