Electric arc additive manufacturing weld defect detection method based on deep learning

Abstract

The invention provides an electric arc additive manufacturing weld defect detection method based on deep learning, and relates to the technical field of image recognition, and the method comprises the steps: adding a lightweight efficient attention mechanism and a feature fusion pyramid network in a neck network Neck, and carrying out the enhancement and fusion of features extracted from a backbone network. Secondly, partial convolution in the network is changed into two-dimensional convolution, and the multi-scale target detection performance of the model is enhanced by using deconvolution; and finally, a novel feature interaction detection head NFIDH is designed to improve the ability of model learning and defect feature selection, and the parameter quantity and the calculation quantity of the model are reduced, so that the performance of the model is better. According to the method, the surface defects of the welding seam can be detected and identified more quickly and accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition and deep learning, and relates to weld defect detection technology, in particular to a method for detecting weld defects in arc additive manufacturing based on deep learning. Background Art

[0002] Arc additive manufacturing is an efficient and low-cost metal additive manufacturing technology, which has broad application prospects in the fields of aerospace, national defense, and energy equipment. However, affected by the arc deposition forming characteristics and material characteristics, surface defects such as cracks and pores often appear on the weld deposition layer surface of WAAM manufactured parts. These surface defects not only affect the surface quality of WAAM parts but may also lead to a decline in the mechanical properties of WAAM parts. Quickly locating and identifying surface defects, and performing characterization and analysis, are important steps for subsequent improvement, suppression, and elimination of defects through process parameter optimization.

[0003] Traditional methods for detecting weld defects in arc additive manufacturing mainly include visual inspection and machine vision inspection. Due to the fact that the quantity, size, and location distribution of surface defects on the weld surface of arc additive manufactured parts have no significant rules, and affected by the welding process, the sample quantity is large and the hardware requirements are high, the detection efficiency and accuracy of traditional methods are relatively low, and traditional image processing technologies cannot adapt to the detection of weld surface defects in complex situations.

[0004] With the rapid development of artificial intelligence and the increasing maturity of deep learning algorithms, convolutional neural networks have better performance in object detection, such as networks like R-CNN and YOLO. The continuous development of deep learning technology provides new ideas and methods for weld defect detection. The object detection method based on deep learning can automatically learn the characteristics of defects from weld surface defect pictures, simplifying the large amount of manual preprocessing work required in traditional methods. Deep neural networks have excellent characteristics such as feature extraction ability and robustness, and these advantages improve the efficiency and accuracy of weld defect detection in arc additive manufacturing.

[0005] Currently, in the technical field of image recognition and deep learning, there are detection methods based on RetinaNet, SSD, YOLO, etc. Among them, YOLO is the most widely used and can efficiently detect weld surface defects. Summary of the Invention

[0006] To solve the above problems, the present invention provides a method for detecting arc additive manufacturing weld defects based on deep learning, which optimizes and improves the original YOLOv8 model. When used to detect arc additive manufacturing weld defects, it can improve its detection accuracy. At the same time, compared with the original YOLOv8 model, the number of model parameters and the amount of computation are significantly reduced. The accuracy rate of the model reaches 79.7%, which is suitable for deployment to embedded devices with limited resources and can effectively solve the above problems.

[0007] To achieve the above technical features, the object of the present invention is realized as follows: A method for detecting arc additive manufacturing weld defects based on deep learning includes the following steps: Step 1: Obtain pictures of weld surface defects, expand the weld defect data set through data augmentation, and label the defects, which are divided into a training set, a test set, and a validation set. Step 2: Establish an improved YOLOv8 weld defect detection model, which includes an embedded ELA attention mechanism, an HSFPN feature fusion pyramid network, and a ConvTranspose2d deconvolution module. A feature interaction detection head NFIDH is adopted in the head network to enhance the detection and recognition performance of the model for weld defects. Step 3: Use the weld defect training set to train the improved YOLOv8 network model, and use the validation set to evaluate the performance of the improved YOLOv8 detection model during the training process to obtain the optimal weld defect detection model. Step 4: Input the weld defect test set into the optimal weld defect detection model and output the test results.

[0008] Preferably, in the above Step 2, the specific operation steps of the improved YOLOv8 neural network model include: adopting the ELA attention mechanism and the HSFPN special effect fusion pyramid network in the YOLOv8 network structure to enhance and fuse the features extracted by the backbone network; introducing the ConvTranspose2d deconvolution to optimize the upsampling and feature fusion of the neck network; in the Head layer of the head network, obtaining the NFIDH feature interaction detection head by combining Group group normalization and DWConv depthwise separable convolution, and replacing the original detection head.

[0009] Preferably, in the above Step 2, the specific process of the improved YOLOv8 neural network model is: replacing the original C2f module with ELA-HSFPN; replacing the Conv module of the original neck network with Conv2d; adding a ConvTranspose2d deconvolution module to the neck network to optimize the original upsampling; replacing the original detection head in the head network with the NFIDH feature interaction detection head.

[0010] Preferably, the neck network is composed of an ELA-HSFPN module, a Conv2d module, a ConvTranspose2d module, a C2f module, etc.

[0011] Preferably, the ELA-HSFPN module replaces the C2f module in YOLOv8, improving the sensitivity and detection accuracy of the model to defect features while reducing the model's computational and storage costs.

[0012] Preferably, the Conv2d convolution has parameter sharing, sparse interaction, translational invariance in object detection, and can combine feature maps of different scales to extract richer features.

[0013] Preferably, the ConvTranspose2d deconvolution module improves the Neck based on the original neck network, upsamples the input features, and enhances the resolution of the feature maps.

[0014] Preferably, the NFIDH detection head module is mainly composed of a Conv_GN group normalization and convolution layer, a task Decoration task decomposition module, a DCNV2 deformable convolution network module, a Generator mask&offset generate mask and offset module, a depthwise separable convolution (DWConv), a DWConv_Seg regression operation module, and a DWConv_Cls classification operation module, etc., improving the model's ability to learn and select weld defect features while reducing the model's number of parameters and computational volume.

[0015] Preferably, the specific operation steps for establishing the weld defect dataset in step 1 are as follows: Use the roboflow public dataset and laboratory arc additive manufacturing equipment to produce weld defect samples, and use data augmentation to expand the weld defect dataset. Manually annotate the weld defect pictures using the LabelImg tool software, and classify the defect categories into surface pores, surface crack, overlap, weld, and spatter; and save them in the YOLOv8 data format. Then, randomly divide the annotated weld defect dataset into a training set, a test set, and a validation set according to a ratio of 7:2:1; the training set is used for model training, the test set is used for model testing, and the validation set is used for model validation; The mosaic data augmentation method, such as cropping, flipping, etc., is used to expand the dataset to mitigate overfitting.

[0016] Preferably, in step 3, the improved YOLOv8 model is trained. The specific steps include: inputting the weld defect training set and validation set into the improved YOLOv8 defect detection model. The experimental hardware configuration uses a 12th generation Intel Core (TM) i5-12490F processor and an NVIDIA GeForce RTX 3060 Graphics card. The operating system is Windows 10. The development platform uses Pycharm software, and the programming language is Python. The integrated development environment is Anaconda2023, Python3.8, and CUDA11.7. The training parameters for this experiment are set as follows: batch_size is set to 64, Epoch is set to 200, img size is set to 640 × 640, the initial learning rate is 0.0001, the momentum parameter is 0.937, and the weight decay coefficient is 0.0005.

[0017] Preferably, in step 4, the improved YOLOv8 defect detection model is loaded for detection processing, and the detection results are output.

[0018] The present invention has the following beneficial effects: 1. Based on YOLOv8, the present invention uses the ELA-HSFPN network to reconstruct the neck network of YOLOv8. It uses Conv2d convolution to replace the ordinary Conv convolution. To improve the upsampling and feature representation of the model, a ConvTranspose2d deconvolution module is added to the Neck network to improve the multi-scale object detection ability. The improved feature interaction detection head NFIDH module is used to reduce the model parameters and computational complexity, optimize the computational efficiency and structural complexity of the network, compress the model and computational complexity, and improve the feature extraction and recognition ability. The entire improved network model is called HEN-YOLOv8. The model of the present invention has good performance in terms of detection accuracy and speed.

[0019] 2. The present invention can complete the detection of weld defects in a short time, basically meeting the real-time requirements and having high performance.

[0020] 3. The present invention can be applied to various occasions according to the detection requirements of arc additive manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below with reference to the drawings and embodiments.

[0022] Figure 1 It is the structural diagram of the improved YOLOv8 network in the present invention.

[0023] Figure 2 It is the structural diagram of NFIDH in the present invention.

[0024] Figure 3 Schematic diagram of the Loss value during the network training of HEN-YOLOv8 in the embodiments of the present invention.

[0025] Figure 4 Schematic diagram of the Evaluate metrics during the network training of HEN-YOLOv8 in the embodiments of the present invention.

[0026] Figure 5 Detection result diagram of the present invention. Detailed implementation manners

[0027] The technical solutions in the embodiments of the present invention will be described in detail below. The described embodiments are only used to further illustrate the present invention, rather than all embodiments. Without departing from the design concept of the present invention, some non-essential improvements and adjustments made by those skilled in the art based on the above invention content shall fall within the protection scope of the present invention.

[0028] Embodiment 1: Step 1: Obtain pictures of weld surface defects, expand the weld defect data set through data augmentation, and label the defects, and divide them into a training set, a test set, and a validation set; Step 2: Establish an improved YOLOv8 weld defect detection model, which includes an embedded ELA attention mechanism, an HSFPN feature fusion pyramid network, and a ConvTranspose2d deconvolution module, and adopts a feature interaction detection head NFIDH in the head network to enhance the detection and recognition performance of the model for weld defects; Step 3: Use the weld defect training set to train the improved YOLOv8 network model, and use the validation set to evaluate the performance of the improved YOLOv8 detection model during the training process to obtain the optimal weld defect detection model; Step 4: Input the weld defect test set into the optimal weld defect detection model and output the test results.

[0029] Embodiment 2: As Figure 1 shown, a method for detecting weld defects in arc additive manufacturing based on deep learning includes the following steps: Step 1: Use the Roboflow public data set and the laboratory arc additive manufacturing equipment to produce and use an industrial camera to take pictures of weld surface defects, expand the weld defect data set through data augmentation, and manually label the defects using the LabelImg tool, and divide the weld surface defect image data set into a training set, a test set, and a validation set according to a set ratio; Step 2, establish an improved YOLOv8 weld defect detection model, which includes an embedded ELA attention mechanism, an HSFPN feature fusion pyramid network, and a ConvTranspose2d deconvolution module. A feature interaction detection head NFIDH is adopted in the head network to enhance the model's detection and recognition performance for weld defects; Among them, the ELA attention mechanism, also known as the efficient local attention mechanism, includes X Avg Pool global average pooling, Conv1d convolution, GroupNorm group normalization layer, and Sigmoid activation function part. It can effectively capture long-range dependencies while maintaining computational efficiency by using one-dimensional convolution to process feature maps. And it enhances the generalization ability of features by using Group Normalization (GN), thus better extracting the details and defect features of the input image. This helps improve the interpretability of the model and, by evaluating the error level in the image, helps the model accurately identify and distinguish defects on the weld surface. The HSFPN feature fusion pyramid network mainly consists of two parts. The first part is the feature selection module, and the second part is the feature fusion module. First, in the feature selection module, feature maps of different scales are evaluated. Subsequently, the high-level and low-level information contained in these feature maps will be integrated through the feature selection process. In this selection process, the feature maps will undergo two pooling operations: global average pooling and global max pooling. The features obtained through these poolings are then merged, and the sigmoid activation function is used to determine the weight values for each channel. Finally, the weights for each channel are obtained, and multiplying the weight information with the feature maps corresponding to each scale can generate the output feature map. In the process of feature fusion, the upsampled high-level features and low-level features are combined by pixel summation to enhance the semantic information of each layer. This module adopts a strategic feature fusion method, using high-level features as weights to selectively filter important semantic information in low-level features. The high-level features are scaled through bilinear interpolation and upsampled or downsampled as needed. Subsequently, the channel attention (CA) module is applied to transform the high-level features into corresponding attention weights, which are then used to filter the low-level features to ensure that the dimensions of the features remain consistent. Finally, the filtered low-level features are integrated with the high-level features, thereby enhancing the feature representation of the model and improving its overall performance. The C2f module mainly consists of an input convolution layer, a Bottleneck layer, a residual connection, and an output convolution layer, etc. The Conv2d two-dimensional convolution module mainly consists of an input feature map, a convolution kernel, a stride, padding, an output feature map, an activation function, and a bias term. By sliding the convolution kernel on the input feature map, features such as edges and shapes in the image can be captured. The parameters of the convolution kernel can be shared across the entire input feature map, featuring sparse interactions and translational invariance. By interacting only with a local area of the input feature map through the convolution kernel, the computational complexity of the model can be reduced to achieve sparse interactions. And when the position of the target in the input image changes, the convolution operation can still capture the features of the target to achieve translational invariance.The ConvTranspose2d transposed convolution module is mainly composed of an input feature map, a convolution kernel, a stride, padding, output padding, an activation function, and a bias term. It can perform upsampling while conducting feature learning to achieve multi-scale object detection. The Feature Interaction Detection Head (NFIDH) is the Head part of the model, including an input layer (P3, P4, P5), a Conv_GN group normalization and convolution layer, a Concat feature concatenation module, a Task Decoration task decomposition module, a Generator mask&offset mask and offset generation module, a DCNV2 deformable convolution, a module for predicting object classes and locations, a Scale scaling layer, etc.; and the module for predicting object classes and locations includes a multiplication operation (Multiply), a depthwise separable convolution (DWConv), a ReLU activation function (ReLU), a Sigmoid activation function (Sigmoid), a regression operation (DWConv_Seg), and a classification operation module (DWConv_Cls); First, the input feature map is processed through three feature extraction layers (P3, P4, P5) to extract feature information of different scales, and then the Conv_GN layer is used to complete feature extraction, reducing the computational complexity of the model and improving the stability and performance of model training; Second, Concat is used for feature fusion to obtain a richer feature map, which helps the network better understand the global information in the image; Third, the task decomposition operation will be performed twice to decompose the detection task into two subtasks; By generating masks and offsets and using DCNV2 to adaptively adjust the positions of the depthwise separable convolution kernels to better capture the shape of the target; After a series of DWConv layers and activation functions, DCNV2 will generate the final masks and offsets; The DCNV2 convolution includes a deformable convolution kernel, an offset generation module, and a feature map generation module for adaptive feature extraction and enhancing the robustness of the model. The DWConv convolution mainly includes a depthwise convolution and a pointwise convolution, which can generate an output feature map with the same number of input channels and restore cross-channel information interaction, reducing the amount of computation and the number of parameters and improving the computational efficiency of the model. Fourth, the object classes and locations are predicted by multiplying the masks and offsets with the feature map, and then the final prediction results are obtained through a series of convolution layers and activation functions; Finally, the scaling module is used to scale the detection results to adapt to targets of different scales, ensuring that the network can accurately detect targets of various sizes, thereby enhancing the model's ability to learn and select defect features while reducing the number of parameters and the amount of computation of the model.

[0030] Step 3: Use the weld defect training set to train the improved YOLOv8 network model, and use the validation set to evaluate the performance of the improved YOLOv8 detection model during the training process to obtain the optimal weld defect detection model. After improving the model using the above three methods, not only the number of model parameters and computational efficiency are optimized, but also the information redundancy in the network neck is reduced, and the feature recognition and extraction capabilities of the model head are improved, thereby improving the detection performance of the model. Step 4: Input the weld defect test set into the optimal weld defect detection model and output the test results.

[0031] As a preferred embodiment of the present invention, in Step 2, the specific process of improving the neural network model of YOLOv8 is as follows: The ELA attention mechanism and HSFPN special effect fusion pyramid network are adopted in the YOLOv8 network structure to enhance and fuse the features extracted by the backbone network; the ConvTranspose2d deconvolution is introduced to optimize the upsampling and feature fusion of the neck network; in the Head layer of the head network, the NFIDH feature interaction detection head is obtained by combining Group group normalization and DWConv depthwise separable convolution, replacing the original detection head.

[0032] In Step 2: The HEN-YOLOv8 network optimizes the computational performance and structure of the model through the Conv2d module, ConvTranspose2d module, ELA-HSFPN module and NFIDH module, and improves the feature extraction and recognition capabilities of the model.

[0033] It should be noted that this model uses Conv2d two-dimensional convolution to replace the original Conv ordinary convolution, and uses ConvTranspose2d deconvolution to enhance the multi-scale object detection performance of the model. This method first slides the convolutional kernel on the input data to extract effective features and parameter sharing, and then slides the convolutional kernel. The model can detect the change of feature positions in the input data, has sparse interaction and translational invariance, and can extract rich features. By adjusting the stride, padding and output padding, the size of the output feature map can be precisely controlled, maintaining the boundary features and feature learning, and effectively realizing multi-scale object detection.

[0034] Furthermore, while embedding the ELA efficient local attention mechanism, the ELA-HSFPN module introduces the HSFPN feature fusion pyramid network. The ELA-HSFPN module is introduced in the Neck layer of the neck network to replace the original C2f module, realizing feature enhancement and fusion. The ELA attention processes features through one-dimensional convolution and group normalization to enhance the generalization ability of the features. At the same time, feature vectors in the horizontal and vertical directions are obtained in the spatial dimension, and rich target location features are generated, avoiding the loss of channel attention information caused by dimensional reduction. In addition, the HSFPN network is divided into two parts: feature selection and feature fusion. Through the channel attention module and the selective feature fusion mechanism, the HSFPN can effectively screen and fuse feature maps of different scales, enhance the feature expression ability of the model, and optimize the performance and detection efficiency of the entire model.

[0035] Furthermore, compared with the original Head, the NFIDH detection head network adds parts such as the Conv_GN module, the TaskDecoration task decomposition module, the Generator mask&offset generation mask and offset module, and the DCNV2 deformable convolution. Specifically, through the three feature maps of P3, P4, and P5, after passing through the Conv_GN 3x3 module, the input features are fused through Concat splicing. And the Task-Decomposition module is used to decompose the obtained feature map into feature maps for classification and regression tasks to complete task decomposition. Then, the DCNV2 module, Multiply module, DWConv module, etc. are used to further process the feature map for the classification task to enhance the feature representation and output the classification result. In addition, the Multiply module, DWConv module, etc. are combined to process the feature map for the regression task to enhance the feature representation and finally output the regression result. Through the above operations, the improved NFIDH feature interaction detection head can effectively process feature maps of different scales, enhance the feature representation, improve the model's ability to learn and select weld defect features, reduce the parameters and computational amount of the model, and significantly improve the accuracy and efficiency of the arc additive manufacturing weld defect detection model.

[0036] In summary, the introduction of the Conv2d module, ConvTranspose2d module, ELA-HSFPN module, and NFIDH module optimizes the computational performance and structure of the model, reduces the number of model parameters and the computational amount, and improves the detection accuracy and detection efficiency of the model.

[0037] In step 4, the designed HEN-YOLOv8 neural network model is trained using the PyTorch deep learning framework. The set hyperparameters are as follows: the training period is 200, the batch size is 64, the initial learning rate is 0.0001, the momentum parameter is 0.937, the weight decay coefficient is 0.0005, and a linear decay learning rate adjustment strategy is used.

[0038] Performance analysis and comparison are carried out for the Neck network and other detection head networks of the present invention, as shown in Table 1: Table 1 Performance comparison of different modules

[0039] As can be seen from Table 1, the P, F1, and mAP@0.5 of the YOLOv8n+EHN network structure of the present invention are 78.2%, 77.8%, and 79.7% respectively. The comprehensive performance is superior to other structures, and a relatively low number of model parameters (2.897M) is maintained. Therefore, the HEN-YOLOv8 network of the present invention has better detection performance.

[0040] Analysis and comparison are carried out for the trained model and other detection models, as shown in Table 2: Table 2 Comparison of different models

[0041] As can be seen from Table 2, while having 2.897M model parameters, the HEN-YOLOv8 model of the present invention also achieves 79.7% mAP@0.5. Among the comparisons of several current mainstream algorithms, its detection accuracy, detection speed, and the number of model parameters are overall better.

[0042] As Figure 5 shown, the detection results of the present invention on the arc additive manufacturing weld defect dataset are given. It can be seen from the figure that the model can accurately identify 5 types of defects: surface pores, surface crack, overlap, weld, and spatter. The experimental results show that the detection results given by the present invention are accurately classified, and the predicted value positioning is almost the same as the true value, with high detection accuracy.

[0043] The above embodiments are only used to illustrate the technical solutions of the present invention and do not limit the present invention; although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that any simple modification, equivalent transformation, and modification made do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for arc additive manufacturing weld defect detection based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain images of weld surface defects, expand the weld defect dataset through data enhancement, annotate the defects, and divide them into training set, test set, and validation set; Step 2: Establish an improved YOLOv8 weld defect detection model, which includes an embedded ELA attention mechanism, an HSFPN feature fusion pyramid network, and a ConvTranspose2d deconvolution module. The feature interaction detection head NFIDH is used in the head network to enhance the model's detection and recognition performance of weld defects. Step 3, using the weld defect training set to train the improved YOLOv8 network model, and using the validation set to evaluate the performance of the improved YOLOv8 detection model during the training process to obtain the optimal weld defect detection model; Step 4: Input the weld defect test set into the optimal weld defect detection model and output the test results.

2. The arc additive manufacturing weld defect detection method based on deep learning according to claim 1, characterized in that: In the step 2, the specific operation steps of improving the neural network model of YOLOv8 include: using the ELA attention mechanism and HSFPN special effect fusion pyramid network to enhance and fuse the features extracted by the backbone network in the YOLOv8 network structure; introducing ConvTranspose2d deconvolution to optimize upsampling and feature fusion on the neck network; in the head network Head layer, by combining Group group normalization and DWConv deep separable convolution, the NFIDH feature interaction detection head is obtained to replace the original detection head.

3. The arc additive manufacturing weld defect detection method based on deep learning according to claim 2, characterized in that: In the step 2, the specific process of improving the neural network model of YOLOv8 is: replacing the original C2f module with ELA-HSFPN; replacing the original Conv module of the neck network with Conv2d; adding the ConvTranspose2d deconvolution module to the neck network to optimize the original upsampling; replacing the original detection head with the NFIDH feature interaction detection head in the head network.

4. The arc additive manufacturing weld defect detection method based on deep learning according to claim 3, characterized in that: The neck network consists of the ELA-HSFPN module, the Conv2d module, the ConvTranspose2d module and the C2f module.

5. The arc additive manufacturing weld defect detection method based on deep learning according to claim 4, characterized in that: The Conv2d convolution has parameter sharing, sparse interaction, translation invariance and the ability to combine feature maps of different scales in target detection to extract richer features.

6. The arc additive manufacturing weld defect detection method based on deep learning according to claim 5, characterized in that: The ConvTranspose2d deconvolution module improves Neck on the basis of the original neck network, upsamples the input features, and enhances the resolution of the feature map.

7. The arc additive manufacturing weld defect detection method based on deep learning according to claim 6, characterized in that: The NFIDH detection head module consists of the main components of Conv_GN group normalization and convolution layer, task Decoration task decomposition module, DCNV2 deformable convolution network module, Generator mask&offset mask and offset module, deep separable convolution, DWConv_Seg regression operation module and DWConv_Cls classification operation module.

8. The arc additive manufacturing weld defect detection method based on deep learning according to claim 1, characterized in that: The specific operation steps for establishing the weld defect data set in step 1 are as follows: using the roboflow public data set and the laboratory arc additive equipment to make weld defect samples, and using data enhancement to expand the weld defect data set, manually annotating the weld defect pictures using the LabelImg tool software, and dividing the defect categories into surface pores, surfacecrack, overlap, weld and spatter; and saving them in the YOLOv8 data format, and then randomly dividing the annotated weld defect data set into a training set, a test set and a validation set in a ratio of 7:2:1; the training set is used for model training, the test set is used for model testing, and the validation set is used for model validation; mosaic data enhancement methods, such as cropping and flipping, are used to expand the data set to reduce overfitting.

9. The arc additive manufacturing weld defect detection method based on deep learning according to claim 1, characterized in that: The step 3 trains the improved YOLOv8 model, and the specific steps include: inputting the weld defect training set and the verification set into the improved YOLOv8 defect detection model, the experimental hardware configuration uses the 12th generation Intel Core (TM) i5-12490F processor, NVIDIA GeForce RTX 3060 Graphics card, the operating system is Windows 10, the development platform uses Pycharm software, and the editing language is Python; the integrated development environment is Anaconda2023, Python3.8, CUDA11.7; the training parameter settings of this experiment are: batch_size is set to 64, Epoch is set to 200, img size is set to 640 × 640, the initial learning rate is 0.0001, the momentum parameter is 0.937, and the weight attenuation coefficient is 0.0005.

10. The arc additive manufacturing weld defect detection method based on deep learning according to claim 1, characterized in that: In step 4, the weld defect data set test set is loaded into the improved YOLOv8 defect detection model for detection processing, and the detection results are output.

Images