CN-121981990-A - Welding defect detection method, device and storage medium

CN121981990ACN 121981990 ACN121981990 ACN 121981990ACN-121981990-A

Abstract

The invention discloses a welding defect detection method, a device and a storage medium, belonging to the technical field of welding quality detection, wherein the method comprises the steps of obtaining a welding image of a battery connecting sheet to be detected; the method comprises the steps of inputting a welding image of a battery connecting sheet to be tested into a pre-trained welding detection model to obtain a welding detection result, wherein the welding detection result comprises a defect detection category and a corresponding detection frame, the welding detection model is obtained by replacing a staged unidirectional serial characteristic propagation path in a YOLOv model with a bidirectional parallel characteristic propagation topology, a sample set for training the welding detection model comprises a normal sample and a defect sample, the normal sample comprises a normal welding image acquired in a history welding process, the defect sample comprises a defect welding image acquired in the history welding process and a defect welding reconstruction image obtained by carrying out defect reconstruction on the normal welding image, and accurate and efficient welding defect detection is achieved.

Inventors

LU NIANSHENG
LI XINGSHUO
ZHANG WEI

Assignees

合肥国轩高科动力能源有限公司

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (10)

1. A method of detecting a solder mark defect, comprising: Acquiring a welding image of a battery connecting sheet to be tested; Inputting a welding image of a battery connecting sheet to be tested into a pre-trained defect detection model to obtain a welding defect detection result, wherein the welding defect detection result comprises a defect detection category and a corresponding detection frame; The defect detection model is obtained by replacing a staged unidirectional serial characteristic propagation path in a YOLOv model with a bidirectional parallel characteristic propagation topology, a sample set for training the defect detection model comprises a normal sample and a diversified defect sample, the normal sample comprises a normal welding image acquired in a battery connecting piece historical welding process, and the defect sample comprises a defect welding image acquired in the battery connecting piece historical welding process and a defect welding reconstruction image obtained by performing defect reconstruction on the normal welding image.
2. The solder mark defect detection method of claim 1, wherein the defect detection model is obtained by replacing PANet of YOLOv neck network with BiFPN.
3. The welding defect detection method according to claim 1, wherein the defect welding image and the normal welding image are obtained by acquiring historical welding processes of battery connecting sheets in real time by an industrial CCD camera and performing standardized processing on pixel values of all channels.
4. The welding defect detection method according to claim 1, wherein the defect welding reconstructed image obtained by performing defect reconstruction on the normal welding image comprises: Injecting defects into the normal welding and printing image by using the diffusion model to obtain a low-resolution defect welding and printing composite image; and reconstructing the low-resolution defect welding synthesized image by using a mixed attention algorithm to obtain a defect welding reconstructed image with the resolution consistent with the acquired image in the history welding process.
5. The method for detecting a solder mask defect according to claim 4, wherein the step of injecting defects into the normal solder mask image using the diffusion model to obtain a low resolution solder mask composite image comprises: Gradually adding Gaussian noise into a normal welding image to obtain a noise image, wherein the expression of adding Gaussian noise into the normal welding image in the step t is as follows: , Wherein, the Is a noise image at the time t, For an initial normal solder print image, Is the cumulative product of the noise scheduling parameters, Is standard Gaussian noise; Denoising the noise image by using a denoising algorithm, and introducing defect guide conditions in the denoising process to generate a low-resolution defect welding and printing composite image, wherein the defect guide conditions comprise preset defect prompt words.
6. The method for detecting a solder mark defect according to claim 4, wherein reconstructing the low-resolution defect solder mark composite image using a mixed-attention algorithm to obtain a defect solder mark reconstructed image having a resolution consistent with the image acquired during the history soldering process, comprises: Performing convolution operation on the low-resolution defect welding synthesized image to obtain an initial feature map; and carrying out local feature capture on the initial feature map by using a window attention mechanism, and dynamically adjusting the channel weight of the initial feature map by using a channel attention mechanism to obtain a defect welding reconstruction image with resolution increased to be consistent with the acquired image.
7. The solder mark defect detection method of claim 2, wherein the defect detection model comprises a backbone network, a neck network comprising the BiFPN, and a detection head; The training method of the defect detection model comprises the following steps: Extracting features of an input sample image through a backbone network to obtain multi-scale global semantic features; Fusing the multi-scale global semantic features through a neck network to obtain cross-scale fusion features; Obtaining a welding defect detection result according to the trans-scale fusion characteristic mapping by a detection head, wherein the welding defect detection result also comprises a confidence coefficient score; According to the defect detection category and the real defect category pre-marked by the sample image, the detection frame, the real frame pre-marked by the sample image and the confidence score, and the real label pre-marked by the sample image, loss calculation is carried out, wherein the defect category comprises no defects and defects; and adjusting parameters of the defect detection model based on the loss calculation result to obtain a trained defect detection model.
8. The method for detecting solder mark defects according to claim 7, wherein the fusing the multi-scale global semantic features through the neck network to obtain the trans-scale fused features comprises: carrying out multi-layer bidirectional feature extraction on the multi-scale global semantic features through BiFPN, and fusing the extracted multi-layer bidirectional features; The method comprises the following steps of: In the top-down path, up-sampling deep semantic features in the multi-scale global semantic features through Upsample layers, splicing the up-sampled features with features of corresponding scales in the multi-scale global semantic features, and then enhancing the spliced features through a C2f module; In the bottom-up path, downsampling is carried out on low-level semantic features in the multi-scale global semantic features through a separable depth convolution layer, the downsampled features are spliced with features of corresponding scales in the top-down path, and then the spliced features are enhanced through a C2f module, so that a cross-scale fusion feature vector is output.
9. A welding defect detecting device, characterized by comprising: The welding image acquisition module to be detected is used for acquiring welding images of the battery connecting sheet to be detected; The device comprises a defect detection result acquisition module, a detection frame and a detection frame, wherein the defect detection result acquisition module is used for inputting a welding image of a battery connecting sheet to be detected into a pre-trained defect detection model to obtain a welding defect detection result; The defect detection model is obtained by replacing a staged unidirectional serial characteristic propagation path in a YOLOv model with a bidirectional parallel characteristic propagation topology, a sample set for training the defect detection model comprises a normal sample and a diversified defect sample, the normal sample comprises a normal welding image acquired in a battery connecting piece historical welding process, and the defect sample comprises a defect welding image acquired in the battery connecting piece historical welding process and a defect welding reconstruction image obtained by performing defect reconstruction on the normal welding image.
10. A computer storage medium having stored thereon a computer program, which, when executed by a processor, implements the solder defect detection method according to any of claims 1-8.

Description

Welding defect detection method, device and storage medium Technical Field The invention belongs to the technical field of welding quality detection, and particularly relates to a welding defect detection method, a welding defect detection device and a storage medium. Background The lithium battery connecting sheet is used as a key component for internal electric connection of the battery, and the welding quality of the lithium battery connecting sheet directly influences the conductivity, the safety and the service life of the battery. Therefore, the quality control of the connecting sheet is important, especially the detection of the welding quality of the connecting sheet. In the manufacturing process of the lithium battery, if the production process is improper, the battery can not reach the production standard, and the normal operation of the production line is further affected. At present, abnormality detection of the connecting sheet mainly depends on the traditional machine vision technology, but the traditional machine vision technology can only process simple and regular image features, and an image collected by laser welding of the connecting sheet usually has rich color information including metallic luster, coating color, background texture and the like, and the factors can cause interference to the traditional graying treatment or simple color space conversion, so that a machine vision algorithm is difficult to accurately identify a defect area. With the development of deep learning technology, a defect detection algorithm based on a convolutional neural network is widely applied. For example, the Chinese patent with publication number CN114862777A, entitled a method and system for detecting welding of connecting sheet, uses an example segmentation algorithm to simultaneously realize positioning, classification and pixel level segmentation of defect targets. However, in practical industrial applications, these algorithms face a number of challenges. On one hand, the method has the defects that the number of defective samples in actual production is limited due to the relative maturity of the production process of the power lithium battery, so that the deep learning model is subjected to the problem of insufficient samples in the training process, the generalization capability and the detection precision of the model are seriously affected, and on the other hand, the abnormal detection algorithm can alleviate the problem of insufficient samples to a certain extent, but the detection speed requirement of the power lithium battery production site is extremely high, the processing of a single image is usually required to be completed in a millisecond level, and the real-time requirement is difficult to be met. Disclosure of Invention The invention aims to provide a welding defect detection method, a welding defect detection device and a storage medium, which are used for improving the feature fusion module of a YoLov model, enhancing the feature extraction and fusion capacity of the welding defect detection device, constructing a defect sample according to an acquired normal sample, training the improved YoLov model, and improving the defect detection precision and speed of the model. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: in a first aspect, the present invention provides a method for detecting a solder mark defect, including: Acquiring a welding image of a battery connecting sheet to be tested; Inputting a welding image of a battery connecting sheet to be tested into a pre-trained defect detection model to obtain a welding defect detection result, wherein the welding defect detection result comprises a defect detection category and a corresponding detection frame; The defect detection model is obtained by replacing a staged unidirectional serial characteristic propagation path in a YOLOv model with a bidirectional parallel characteristic propagation topology, a sample set for training the defect detection model comprises a normal sample and a diversified defect sample, the normal sample comprises a normal welding image acquired in a battery connecting piece historical welding process, and the defect sample comprises a defect welding image acquired in the battery connecting piece historical welding process and a defect welding reconstruction image obtained by performing defect reconstruction on the normal welding image. The method comprises the steps of taking a history normal welding image and a defect welding image set as a basis, solving the bottleneck problem of scarcity of defect samples in an industrial scene through a defect sample synthesis technology based on the normal welding image, training a defect detection model by using the normal samples and diversified defect samples at the same time, ensuring that the training data distribution is highly matched with an actual application scene, enabling the model to cover var