Method for on-line prediction of incomplete fusion defects in gas metal arc welding based on dynamic characteristics of molten pool

Abstract

This invention discloses an online prediction method and system for incomplete fusion defects in gas metal arc welding (GMAW) based on the dynamic characteristics of the molten pool. The method includes: real-time acquisition of molten pool images during the welding process and obtaining the location information of incomplete fusion defects; labeling the acquired molten pool images according to the location information of the incomplete fusion defects; extracting the region of interest (ROI) from the acquired molten pool images and preprocessing it to obtain preprocessed molten pool images; combining multiple consecutive preprocessed molten pool images into a time-series molten pool dataset, using the label of the last image as the label of this dataset; establishing a time-series machine learning model, training, validating, and testing it using the time-series molten pool dataset to obtain a trained time-series machine learning model; and using the trained model to predict the welding process quality in real time. This invention achieves online real-time prediction of incomplete fusion defects, reduces rework costs, and helps improve welding quality and efficiency.

Description

Technical Field

[0001] This invention relates to the field of welding nondestructive testing technology, and in particular to an online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool. Background Technology

[0002] Automated welding technology has been widely applied in the production of key components in energy, shipbuilding, and aerospace industries. Gas metal arc welding (GMAW) exhibits unique advantages in joining medium-thick plates and in metal additive manufacturing due to its high deposition rate, high material utilization, and low cost. However, in GMAW, insufficient heat input, irregular geometry of the weld area, and incomplete contaminant removal can easily lead to incomplete fusion defects. These defects significantly reduce the weld's load-bearing area and cause severe stress concentration, resulting in serious hazards. Therefore, detecting incomplete fusion defects is crucial for improving the efficiency and quality of GMAW.

[0003] Currently, non-destructive testing methods such as ultrasonic testing, X-ray testing, eddy current testing, and infrared testing are mainly used in industrial production to detect non-fusion defects in low-carbon steel GMAW (Glass Wire Welded Aluminum). These methods cannot detect defects during the welding process and are primarily used for quality inspection, failing to provide process guidance for welding. Therefore, there is an urgent need for online prediction and early warning methods for non-fusion defects during welding. Some scholars have used various sensing methods to study optical, acoustic, thermal, and electrical parameters during the welding process, achieving preliminary online prediction of non-fusion defects under specific conditions. However, numerous factors contribute to non-fusion defects, and the coupling relationship between various defects and signals during welding is complex. Currently, there is no accurate and efficient online prediction method for non-fusion defects in GMAW. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide an online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool.

[0005] Technical solution: The online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool, as described in this invention, includes the following steps:

[0006] Real-time acquisition of molten pool images during the welding process; after welding is completed, flaw detection is performed on the welded parts to obtain information on the location of incomplete fusion defects.

[0007] The collected molten pool images are labeled according to the location information of the non-fusion defects. Molten pool images containing defects are labeled as "abnormal", and the rest are labeled as "normal".

[0008] Extract the region of interest from the acquired molten pool image and preprocess it to obtain a preprocessed molten pool image. Combine multiple consecutive preprocessed molten pool images into a set of time-series molten pool data. Use the label of the last image as the label of the set of molten pool data to generate a time-series molten pool dataset.

[0009] A temporal machine learning model was established and trained, validated, and tested using a temporal melt pooling dataset to obtain a trained temporal machine learning model. The model includes an encoder, a self-attention module, a ConvLSTM module, and a classification prediction head. The encoder extracts spatial features of single-frame images through convolutional layers and introduces skip connections to avoid gradient explosion or vanishing. The self-attention module performs adaptive average pooling on the feature map output by the encoder and calculates temporal attention weights for different images in each data set. These weights are then applied to the feature map output by the encoder to select images crucial for prediction. The ConvLSTM module uses a convolutional long short-term memory network to process the temporal features in the output of the self-attention module, capturing long-term dependencies in the sequence and outputting the hidden state of the last time step. The classification prediction head processes the hidden state of the last time step output by the ConvLSTM, calculates the prediction result through fully connected layers, and outputs a "normal" or "abnormal" prediction result.

[0010] A trained time-series machine learning model is used to predict the quality of the welding process in real time. If the model predicts that there is a lack of fusion, an alarm is triggered.

[0011] Furthermore, real-time acquisition of molten pool images during the welding process includes: using a molten pool camera to simultaneously start shooting at the moment of arc ignition and stop shooting at the moment of arc extinguishing, acquiring molten pool images during the welding process.

[0012] Furthermore, the acquired molten pool images are annotated based on the location information of the lack of fusion defects, including:

[0013] The location information of the non-fusion defect obtained from the flaw detection is associated with the sequence of molten pool images. Taking the arc initiation point of welding as the reference, the distance from the starting position of the defect to the arc initiation point is L0, the distance from the ending position of the defect to the arc initiation point is L1, the frame rate of the molten pool camera is f, and the speed of the welding torch is v. Then, the sequence interval of the molten pool image corresponding to the non-fusion defect is:

[0014]

[0015] Where, n 0, n1 is the image sequence number corresponding to the starting point of the currently detected non-fusion defect, and n1 is the image sequence number corresponding to the ending point of the currently detected non-fusion defect; images within the defect interval are marked as "abnormal", and the rest are marked as "normal".

[0016] Furthermore, the region of interest is extracted from the acquired molten pool image and preprocessed, including:

[0017] The YOLO v5 object detection algorithm is used to extract the region of interest in the molten pool image. The identified region of interest is cropped, and images with incorrect identification are corrected. Finally, black borders are filled at the edges of the cropped molten pool image, and it is converted to a uniform size.

[0018] Furthermore, multiple consecutive preprocessed molten pool images are combined into a set of time-series molten pool data. The label of the last image is used as the label of this set of molten pool data to generate a time-series molten pool dataset, including:

[0019] Using an interval sampling method, one image is selected for every n images. Referring to images from the previous t seconds at the current position, a time-series melt pool dataset with dynamic characteristics is formed. The total number of melt pool images in a dataset is:

[0020]

[0021] Where f is the frame rate of the molten pool camera, N images are selected at equal intervals before the current image; if there are fewer than N sampleable images before the first series of images, the first image is used to fill in the gaps.

[0022] Furthermore, equal amounts of "normal" and "abnormal" melt pool image data are selected to generate a time-series melt pool dataset.

[0023] The online prediction system for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool, as described in this invention, includes:

[0024] The defect location acquisition unit is used to acquire images of the molten pool in real time during the welding process, and to perform flaw detection on the welded parts after welding to obtain information on the location of non-fusion defects.

[0025] The image annotation unit is used to annotate the acquired molten pool image according to the location information of the non-fusion defect. Molten pool images containing defects are annotated as "abnormal", and the rest are annotated as "normal".

[0026] The temporal melt pool dataset generation unit is used to extract the region of interest from the acquired melt pool image and perform preprocessing to obtain a preprocessed melt pool image. Multiple consecutive preprocessed melt pool images are combined into a set of temporal melt pool data. The label of the last image is used as the label of the set of melt pool data to generate a temporal melt pool dataset.

[0027] The temporal machine learning model building and training unit is used to build a temporal machine learning model. It trains, validates, and tests the model using a temporal melt pool dataset to obtain a trained temporal machine learning model. The temporal machine learning model includes an encoder, a self-attention module, a ConvLSTM module, and a classification prediction head. The encoder extracts spatial features of a single frame image through convolutional layers and introduces skip connections to avoid gradient explosion or vanishing. The self-attention module performs adaptive average pooling on the feature map output by the encoder and calculates the temporal attention weights for different images in each data set. These weights are then applied to the feature map output by the encoder to select images crucial for prediction. The ConvLSTM module uses a convolutional long short-term memory network to process the temporal features in the output of the self-attention module, capturing long-term dependencies in the sequence and outputting the hidden state of the last time step. The classification prediction head processes the hidden state of the last time step output by the ConvLSTM, calculates the prediction result through fully connected layers, and outputs a "normal" or "abnormal" prediction result.

[0028] The real-time prediction unit is used to predict the quality of the welding process in real time using a trained time-series machine learning model. If the model predicts that there is a lack of fusion, an alarm will be triggered.

[0029] An electronic device for storing and executing the method includes:

[0030] Memory containing executable program code;

[0031] A processor coupled to the memory;

[0032] The processor calls the executable program code stored in the memory to execute the steps of the online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool.

[0033] A computer-readable storage medium for storing and executing the method, the computer-readable storage medium storing computer instructions, which, when invoked, are used to execute the steps of the online prediction method for incomplete fusion defects in gas metal arc welding based on dynamic characteristics of the molten pool.

[0034] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the online prediction method for incomplete fusion defects in gas metal arc welding based on dynamic characteristics of the molten pool.

[0035] Beneficial effects: Compared with the prior art, the significant technical effects of this invention are as follows: It realizes online real-time prediction of non-fusion defects in gas metal arc welding, overcoming the lag of traditional offline detection. It obtains the whole process information of defect formation by capturing the dynamic features of time-series molten pool images; it learns the dynamic information of molten pool images through a time-series machine learning model and performs online prediction of non-fusion defects; at the same time, the system adopts a closed-loop control mechanism, which can terminate abnormal welding processes in a timely manner to avoid defect expansion, significantly reducing rework costs and providing effective technical support for improving welding quality and production efficiency. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the method described in this invention;

[0037] Figure 2 This is a schematic diagram of the installation of a molten pool camera provided in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of a method for matching molten pool images with defect information obtained from flaw detection, provided in an embodiment of the present invention.

[0039] Figure 4 This is a schematic diagram of the region of interest extracted from the molten pool image provided in an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of a temporal machine learning model structure provided in an embodiment of the present invention;

[0041] Wherein: 1-Camera fixture, 2-Molten pool camera, 3-Welding torch, 4-Weld seam, 5-Location of defects found during flaw detection, 6-Substrate, 7-Region of interest in the molten pool image. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0043] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0044] As the core area of ​​the welding process, the molten pool contains a wealth of welding quality information. In recent years, with the development of computer vision technology and machine learning methods, predicting incomplete fusion defects using molten pool images has become a research hotspot. This method is expected to achieve real-time and accurate monitoring of welding quality.

[0045] This invention uses a time-series machine learning model combined with dynamic molten pool images to predict incomplete fusion defects in gas metal arc welding online. This allows for timely detection and handling of incomplete fusion defects during the welding process, which is of great significance and value for improving the production efficiency and welding quality of gas metal arc welding.

[0046] like Figure 1 As shown, the molten pool camera 2 is fixed to the side of the welding torch 3 along its direction of movement to capture images of the molten pool during the welding process and store them in a computer. A typical scenario is as follows. Figure 1 As shown, the molten pool camera 2 is fixed to the side of the welding torch 3 in the direction of movement by the camera fixing fixture 1. In this embodiment, the molten pool camera 2 is connected to the computer via a gigabit network cable. When the welding arc is started, the molten pool camera 2 begins to take pictures and stops taking pictures when the arc is extinguished, recording the molten pool image during the welding process and transmitting it to the computer.

[0047] like Figure 2 As shown, the online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool, as described in this invention, includes the following steps:

[0048] S1. Conduct welding experiments using a molten pool camera. Imaging begins synchronously with arc ignition and stops upon arc extinguishing, capturing images of the molten pool during the welding process and storing them in a computer. After welding, the welded parts are inspected to obtain information on the location of incomplete fusion defects. In this embodiment, the substrate to be welded (i.e., the welded parts) consists of two butt-jointed steel plates with a V-shaped bevel at the weld joint. Incomplete fusion defects are artificially generated by adding contaminants to the area to be welded. After welding, an ultrasonic flaw detector is used to inspect the welded parts and obtain information on the location of incomplete fusion defects.

[0049] S2. Annotate the acquired molten pool image based on the location information of the incomplete fusion defect. For example... Figure 3 As shown, a weld 4 is formed on the substrate 6 by welding. A non-fusion defect 5 detected by flaw detection is located on the weld 4. Taking the arc initiation point as a reference, the starting position of the non-fusion defect 5 obtained by flaw detection is L0 from the arc initiation point, and the ending point is L1 from the arc initiation point. The welding torch's movement speed is v, and the frame rate of the molten pool camera is f. Then, the ordinal interval of the molten pool image corresponding to the non-fusion defect is:

[0050]

[0051] Where, n 0,n1 is the image sequence number corresponding to the starting point of the currently detected non-fusion defect, and n1 is the image sequence number corresponding to the ending point of the currently detected non-fusion defect; images within the defect interval are marked as "abnormal", and the rest are marked as "normal".

[0052] S3. Preprocess the acquired molten pool image, extract the region of interest (ROI) 7 from the molten pool image, and remove redundant information. For example... Figure 4 As shown in the figure, this is an image of the molten pool captured by the molten pool camera in this embodiment. The YOLO v5 object detection algorithm is used to extract the region of interest (ROI) 7 outlined in the image. In this embodiment, the ROI is the molten pool region. A certain number of molten pool images containing features such as substrate, welding torch, welding wire, and molten pool are randomly selected from the molten pool images and combined into a miniature molten pool image dataset. The ROI target boxes are manually annotated on this dataset. The YOLO v5 model is used to train this dataset. After training, the model is applied to all molten pool images, and the identified regions (i.e., ROIs) are cropped. Finally, all cropped images are checked, and images with incorrect identification are manually corrected. Finally, black borders are added to the edges of the cropped molten pool images to make them square, and they are converted to a uniform size.

[0053] S4. Combine multiple consecutive molten pool images of uniform size obtained in step S3 into a set of time-series molten pool data. Use the label of the last image as the label of this set of molten pool data to generate a time-series molten pool dataset. Because the frame rate of the molten pool camera is relatively high, adjacent molten pool images are quite similar and contain a large amount of redundant information. If all molten pool images within a certain period are continuously selected to form a set of dynamic molten pool data, the data volume would be extremely large, severely slowing down the training speed of the machine learning model, or even making training impossible. Therefore, an interval sampling method is adopted, selecting one image from every n images, referring to images within t seconds before the current position, to form time-series molten pool data with dynamic features. The total number of molten pool images in a set of data is:

[0054]

[0055] Where f is the frame rate of the melt pool camera. N images are selected at equal intervals before the current image. A sliding sampling method is used to fully utilize each melt pool image. If there are fewer than N sampleable images before the first series of images, the first image is used to fill in the gaps. Since the number of melt pool images labeled "normal" is significantly greater than that labeled "abnormal," to prevent the model from biasing towards outputting "normal" predictions, the number of melt pool image data labeled "abnormal" is counted, and an equal amount of data is randomly selected from the melt pool image data labeled "normal" to generate a time-series melt pool dataset.

[0056] S5. Establish a time-series machine learning model. For example... Figure 5As shown, the temporal machine learning model used in this embodiment adopts an architecture of "encoder-self-attention module-ConvLSTM module-classification prediction head". The encoder extracts spatial features of a single frame image through convolutional layers. In each convolutional layer, a 3×3 convolution is performed, ReLU activation is applied, and batch normalization is performed. Skip connections are made every other convolutional layer to avoid gradient explosion or vanishing. The total number of convolutional layers is selected reasonably based on the task complexity and computing device performance; in this embodiment, it is set to 7. The self-attention module performs adaptive average pooling on the feature map output by the encoder and calculates the temporal attention weights for different images in each data set. These temporal attention weights are applied to the feature map output by the encoder to filter images crucial for prediction. The ConvLSTM module uses a convolutional long short-term memory network to process the temporal features in the output of the self-attention module. First, the cell state and hidden state are initialized, then the forget gate, input gate, and output gate are calculated, and the hidden state and cell state are updated. This step is performed along the time steps to capture long-term dependencies in the sequence, and the hidden state of the last time step is output. The classification prediction head performs adaptive average pooling on the hidden state of the last time step output by the ConvLSTM and flattens it into one dimension. The prediction result is calculated through a fully connected layer and then output.

[0057] S6. The established time-series machine learning model is trained, validated, and tested using a time-series melt pool dataset. The dataset is divided into training, validation, and test sets according to a ratio of 90%, 5%, and 5%, respectively. During each training round, the training data is input into the model for inference in batches. The batch size is set according to the computing device performance; in this embodiment, it is set to 8. After the model outputs the prediction results, the cross-entropy function is used as the loss function, and the output results are compared with the true values. The cross-entropy function H(p,q) is:

[0058]

[0059] Where p(x) i q(x) represents the actual probability distribution, i.e., the true label corresponding to the current data. i Let be the predicted probability distribution, i.e., the model's output on the current data, and m be the total number of training samples in this batch. The gradient of the loss function with respect to the model parameters is calculated using the Adams optimizer for iterative optimization. After each round of optimization, the validation set data is input into the model for inference, and the model's performance is observed. After all rounds of optimization, the test set data is input into the model for inference. If the model can achieve the prediction task, the process stops; if the model cannot achieve the prediction task, the hyperparameters during training are adjusted, and retraining is performed.

[0060] S7. Deploy the trained model on the industrial welding production line to monitor welding quality in real time.

[0061] The method proposed in this invention enables online real-time prediction of incomplete fusion defects in gas metal arc welding. Compared with existing technologies, it allows for timely detection and handling of incomplete fusion defects during the welding process, enabling the timely termination of abnormal welding processes to prevent defect expansion, significantly reducing rework costs, and significantly improving welding quality and production efficiency.

[0062] The online prediction system for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool, as described in this invention, includes:

[0063] The defect location acquisition unit is used to acquire images of the molten pool in real time during the welding process, and to perform flaw detection on the welded parts after welding to obtain information on the location of non-fusion defects.

[0064] The image annotation unit is used to annotate the acquired molten pool image according to the location information of the non-fusion defect. Molten pool images containing defects are annotated as "abnormal", and the rest are annotated as "normal".

[0065] The temporal melt pool dataset generation unit is used to extract the region of interest from the acquired melt pool image and perform preprocessing to obtain a preprocessed melt pool image. Multiple consecutive preprocessed melt pool images are combined into a set of temporal melt pool data. The label of the last image is used as the label of the set of melt pool data to generate a temporal melt pool dataset.

[0066] The temporal machine learning model building and training unit is used to build a temporal machine learning model. It trains, validates, and tests the model using a temporal melt pool dataset to obtain a trained temporal machine learning model. The temporal machine learning model includes an encoder, a self-attention module, a ConvLSTM module, and a classification prediction head. The encoder extracts spatial features of a single frame image through convolutional layers and introduces skip connections to avoid gradient explosion or vanishing. The self-attention module performs adaptive average pooling on the feature map output by the encoder and calculates the temporal attention weights for different images in each data set. These weights are then applied to the feature map output by the encoder to select images crucial for prediction. The ConvLSTM module uses a convolutional long short-term memory network to process the temporal features in the output of the self-attention module, capturing long-term dependencies in the sequence and outputting the hidden state of the last time step. The classification prediction head processes the hidden state of the last time step output by the ConvLSTM, calculates the prediction result through fully connected layers, and outputs a "normal" or "abnormal" prediction result.

[0067] The real-time prediction unit is used to predict the quality of the welding process in real time using a trained time-series machine learning model. If the model predicts that there is a lack of fusion, an alarm will be triggered.

[0068] An electronic device for storing and executing the method includes:

[0069] Memory containing executable program code;

[0070] A processor coupled to the memory;

[0071] The processor calls the executable program code stored in the memory to execute the steps of the online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool.

[0072] A computer-readable storage medium for storing and executing the method, the computer-readable storage medium storing computer instructions, which, when invoked, are used to execute the steps of the online prediction method for incomplete fusion defects in gas metal arc welding based on dynamic characteristics of the molten pool.

[0073] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the online prediction method for incomplete fusion defects in gas metal arc welding based on dynamic characteristics of the molten pool.

Claims

1. A method for online prediction of incomplete fusion defects in gas metal arc welding based on dynamic characteristics of the molten pool, characterized in that, Includes the following steps: Real-time acquisition of molten pool images during the welding process; after welding is completed, flaw detection is performed on the welded parts to obtain information on the location of incomplete fusion defects. The collected molten pool images are labeled according to the location information of the non-fusion defects. Molten pool images containing defects are labeled as "abnormal", and the rest are labeled as "normal". Extract the region of interest from the acquired molten pool image and preprocess it to obtain a preprocessed molten pool image. Combine multiple consecutive preprocessed molten pool images into a set of time-series molten pool data. Use the label of the last image as the label of the set of molten pool data to generate a time-series molten pool dataset. A time-series machine learning model is established, and it is trained, validated, and tested using a time-series melt pool dataset to obtain a well-trained time-series machine learning model. The temporal machine learning model includes an encoder, a self-attention module, a ConvLSTM module, and a classification prediction head. The encoder extracts spatial features of a single frame image through convolutional layers and introduces skip connections to avoid gradient explosion or vanishing. The self-attention module performs adaptive average pooling on the feature map output by the encoder and calculates the temporal attention weights for different images in each data set. These temporal attention weights are then applied to the feature map output by the encoder to select images that are key to the prediction. The ConvLSTM module uses a convolutional long short-term memory network to process the temporal features in the output of the self-attention module, capturing long-term dependencies in the sequence and outputting the hidden state of the last time step. The classification prediction head processes the hidden state of the last time step output by the ConvLSTM, calculates the prediction result through fully connected layers, and outputs a "normal" or "abnormal" prediction result. A trained time-series machine learning model is used to predict the quality of the welding process in real time. If the model predicts that a lack of fusion defect will occur, an alarm will be triggered.

2. The online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool according to claim 1, characterized in that, Real-time acquisition of molten pool images during the welding process includes: using a molten pool camera to simultaneously start shooting at the moment of arc ignition and stop shooting at the moment of arc extinguishing, acquiring molten pool images during the welding process.

3. The online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool according to claim 1, characterized in that, The acquired molten pool images are labeled based on the location information of the incomplete fusion defects, including: The location information of the non-fusion defect obtained from the flaw detection is associated with the sequence of molten pool images. Taking the arc initiation point of welding as the reference, the distance from the starting position of the defect to the arc initiation point is L0, the distance from the ending position of the defect to the arc initiation point is L1, the frame rate of the molten pool camera is f, and the speed of the welding torch is v. Then, the sequence interval of the molten pool image corresponding to the non-fusion defect is: Where n0 is the image sequence number corresponding to the current detected starting point of the non-fusion defect, and n1 is the image sequence number corresponding to the current detected ending point of the non-fusion defect; images within the defect interval are marked as "abnormal", and the rest are marked as "normal".

4. The online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool according to claim 1, characterized in that, Extract the region of interest from the acquired molten pool image and perform preprocessing, including: The YOLO v5 object detection algorithm is used to extract the region of interest in the molten pool image. The identified region of interest is cropped, and images with incorrect identification are corrected. Finally, black borders are filled at the edges of the cropped molten pool image, and it is converted to a uniform size.

5. The online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool according to claim 1, characterized in that, Multiple consecutive preprocessed melt pool images are combined into a time-series melt pool dataset. The label of the last image is used as the label of this dataset to generate a time-series melt pool dataset, which includes: Using an interval sampling method, one image is selected for every n images. Referring to images from the previous t seconds at the current position, a time-series melt pool dataset with dynamic characteristics is formed. The total number of melt pool images in a dataset is: Where f is the frame rate of the molten pool camera, N images are selected at equal intervals before the current image; if there are fewer than N sampleable images before the first series of images, the first image is used to fill in the gaps.

6. The online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool according to claim 1, characterized in that, Equal amounts of "normal" and "abnormal" melt pool image data are selected to generate a time-series melt pool dataset.

7. An online prediction system for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool, characterized in that, include: The defect location acquisition unit is used to acquire images of the molten pool in real time during the welding process, and to perform flaw detection on the welded parts after welding to obtain information on the location of non-fusion defects. The image annotation unit is used to annotate the acquired molten pool image according to the location information of the non-fusion defect. Molten pool images containing defects are annotated as "abnormal", and the rest are annotated as "normal". The temporal melt pool dataset generation unit is used to extract the region of interest from the acquired melt pool image and perform preprocessing to obtain a preprocessed melt pool image. Multiple consecutive preprocessed melt pool images are combined into a set of temporal melt pool data. The label of the last image is used as the label of the set of melt pool data to generate a temporal melt pool dataset. The temporal machine learning model building and training unit is used to build a temporal machine learning model, and to train, validate and test it using a temporal melt pool dataset to obtain a trained temporal machine learning model. The temporal machine learning model includes an encoder, a self-attention module, a ConvLSTM module, and a classification prediction head. The encoder extracts spatial features of a single frame image through convolutional layers and introduces skip connections to avoid gradient explosion or vanishing. The self-attention module performs adaptive average pooling on the feature map output by the encoder and calculates the temporal attention weights for different images in each data set. These temporal attention weights are then applied to the feature map output by the encoder to select images that are key to the prediction. The ConvLSTM module uses a convolutional long short-term memory network to process the temporal features in the output of the self-attention module, capturing long-term dependencies in the sequence and outputting the hidden state of the last time step. The classification prediction head processes the hidden state of the last time step output by the ConvLSTM, calculates the prediction result through fully connected layers, and outputs a "normal" or "abnormal" prediction result. The real-time prediction unit is used to predict the quality of the welding process in real time using a trained time-series machine learning model. If the model predicts that there is a lack of fusion, an alarm will be triggered.

8. An electronic device, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, are used to perform the steps of the online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the online prediction method for incomplete fusion defects in gas metal arc welding based on the dynamic characteristics of the molten pool as described in any one of claims 1-6.

Images