Metal real-time deposition defect detection method based on artificial intelligence

Abstract

The invention relates to a metal real-time deposition defect detection method based on artificial intelligence, and the method comprises the steps: obtaining a deposition defect data set, and extracting key deposition features through a backbone network; capturing high-level semantic features in the sedimentary features through an efficient hybrid encoder and performing multi-scale feature fusion; by utilizing an intersection-union ratio perception query selection method, corresponding classification scores are distributed to the sedimentary features with different intersection-union ratios in the training process so as to constrain model learning, and more accurate sedimentary features are selected in the decoding stage by optimizing object query; the optimization object is mapped to the classification confidence and the bounding box, and training convergence is accelerated through a denoising module in a decoder. The method has the beneficial effects that an artificial intelligence algorithm model for end-to-end real-time target detection is designed by taking deposition defects as objects, the training speed is high, the generalization ability is high, and the method is suitable for deposition application scenes such as electric arc additive manufacturing needing efficient, real-time and accurate detection.

Description

Technical Field

[0001] The present invention relates to the technical field of additive manufacturing, and more precisely, it relates to a method for real-time deposition defect detection of metals based on artificial intelligence. Background Art

[0002] Metal additive manufacturing and metal weld welding are effective methods for the integrated deposition manufacturing and welded connection of metal structures in industries such as construction machinery, aerospace, and automotive manufacturing. The former directly manufactures metal structures integrally by heating wire materials at high temperatures, and the latter connects metal structures by high-temperature heating welding. Deposition defects during the deposition process not only affect the structure appearance but also endanger its performance and safety. Therefore, real-time detection and evaluation of deposition defects are crucial for ensuring the safety and reliability of industrial facilities.

[0003] Traditional deposition defect detection generally collects data through methods such as X-rays, laser scanning, and ultrasonic waves, and uses computer-aided algorithms for data processing. However, the accuracy of these detection methods highly depends on experience, resulting in low efficiency, high subjectivity, and unreliable detection results, and unable to achieve real-time detection effects. Therefore, there is an urgent need to develop an efficient and accurate real-time end-to-end target detection method for deposition defects to be applicable to practical applications in various complex scenarios.

[0004] With the development of deep convolutional neural networks (CNNs), object detection algorithms have obtained powerful capabilities in feature extraction and learning complex patterns; CNNs have significantly improved the accuracy of image recognition and have high applicability in deposition detection; currently, various model algorithms and detection methods have been developed based on this. However, the existing detection model algorithm structures are often too complex and require multi-stage training, resulting in low computational efficiency, and thus there are huge challenges in its application for real-time detection.

[0005] Especially in complex metal deposition application scenarios with high temperatures and high speeds such as wire arc additive manufacturing (WAAM), real-time and accurate detection of deposition defects is crucial for structural safety and production efficiency. Without real-time detection and adjustment, due to the overlap of deposition layers, deposition defects may be overlooked, leading to hidden defects. However, most existing real-time detectors rely on CNN-based architectures and require non-maximum suppression (NMS) for post-processing, which slows down the detection speed and introduces hyperparameters that cause instability in computational efficiency and effectiveness.

[0006] In summary, it is very necessary to study a method for real-time deposition defect detection of metals based on artificial intelligence to achieve efficient, real-time, and accurate detection in multi-layer and multi-pass complex deposition defect application scenarios with high temperatures and high speeds such as arc additive manufacturing. Summary of the Invention

[0007] The object of the present invention is to propose a real-time deposition defect detection method for metals based on artificial intelligence in view of the deficiencies of the prior art.

[0008] In a first aspect, there is provided a real-time deposition defect detection method for metals based on artificial intelligence, including:

[0009] S1. Obtain a deposition defect data set and extract key deposition features using a backbone network;

[0010] S2. Combine an attention mechanism and multi-layer convolution operations to form an efficient hybrid encoder, and capture high-level semantic features in the deposition features and perform multi-scale feature fusion through the efficient hybrid encoder;

[0011] S3. Use the intersection over union (IoU)-aware query selection method to assign corresponding classification scores to deposition features with different IoUs during the training process to constrain model learning, and select more accurate deposition features at the decoding stage by optimizing object queries;

[0012] S4. Use a decoder with an auxiliary prediction head to map the optimized object to classification confidence and bounding boxes, and accelerate the training convergence through the denoising module in the decoder.

[0013] Preferably, in S1, it further includes: expanding the deposition defect data set through a data augmentation method;

[0014] Preferably, in S1, for the deposition defect data set, it is marked as "good" or "bad" according to the presence or absence of defect forms and annotated; the deposition defects include forms such as deposition burrs, deposition holes, and deposition dents; the deposition defect data set is divided into a training set and a validation set, which are respectively used for model training and performance evaluation;

[0015] Preferably, in S1, the backbone network is an improved PResNet50, in which the initial 7×7 convolutional layer is replaced by three 3×3 convolutional layers, and a 2×2 average pooling layer is introduced in the residual block to replace part of the 1×1 convolutional layer; the improved PResNet50 is used to enhance the retention of deposition key information and improve the efficiency of deposition feature extraction;

[0016] Preferably, in S2, the efficient hybrid encoder includes an intra-scale feature interaction module based on the attention mechanism and a cross-scale feature fusion module based on the convolutional neural network; the intra-scale feature interaction module is used to capture high-level semantic features in the deposition features; the cross-scale feature fusion module is used to fuse multi-scale features and enhance the overall deposition feature information;

[0017] Preferably, in S2, for shallow sedimentary features, a convolutional layer is used for extraction; for deep sedimentary features, corresponding to high-level sedimentary semantic features, a sedimentary feature map output by an efficient hybrid encoder is used for cross-scale fusion to perform multi-scale sedimentary feature extraction and improve the efficiency of sedimentary feature extraction;

[0018] Preferably, in S2, a parallel structure is used to extract sedimentary features of different degrees from the sedimentary feature map, and sedimentary feature maps of multiple different scales are combined to enhance the fusion effect of sedimentary features;

[0019] Preferably, in S3, precision, recall, and mean average precision are used as evaluation metrics, and the intersection over union between the predicted and actual bounding boxes is used as the evaluation threshold; the calculation formulas for precision, recall, average precision, and mean average precision are as follows:

[0020]

[0021]

[0022] Among them, P represents precision, R represents recall, AP represents average precision, mAP represents mean average precision, TP, FP, and FN respectively represent "true positive", "false positive", and "false negative", which are defined according to the intersection over union between the predicted bounding box and the ground truth; is the recall at any point on the precision P-recall R curve ; k is the interpolation point number; N is the total number of interpolation points; is the precision when the recall is ; ΔR(k) is the change in recall between the (k - 1)-th and k-th interpolation points; n is the total number of classes in multi-class object detection; i is the class number; the bounding box is marked as TP when the intersection over union is greater than the threshold, representing the number of correctly recognized objects; the bounding box is marked as FP when the intersection over union is less than or equal to the threshold, representing the number of misrecognized objects;

[0023] Preferably, in S3, when detecting the speed, the number of frames per second is used as the main metric for evaluating the detection speed, represented by FPS;

[0024] Preferably, in S4, during the training process, the ground truth box is denoised and used as part of the decoder input to accelerate the training convergence speed; and a piecewise decay schedule and linear warm-up are used to set the learning rate to address the training hyperparameter problem.

[0025] In a second aspect, an artificial intelligence-based real-time metal deposition defect detection system is provided for performing the method according to any one of the first aspect, including:

[0026] An acquisition module for acquiring a deposition defect dataset and extracting key deposition features using a backbone network;

[0027] A composition module for constructing an efficient hybrid encoder by combining an attention mechanism and multi-layer convolutional operations, and capturing high-level semantic features in the deposition features and performing multi-scale feature fusion through the efficient hybrid encoder;

[0028] An assignment module for using the intersection over union (IoU)-aware query selection method to assign corresponding classification scores to deposition features with different IoUs during the training process to constrain model learning, and selecting more accurate deposition features during the decoding phase by optimizing object queries;

[0029] A mapping module for mapping the optimized object to classification confidence and bounding boxes using a decoder with an auxiliary prediction head, and accelerating training convergence through a denoising module in the decoder.

[0030] In a third aspect, a computer storage medium is provided, in which a computer program is stored; when the computer program runs on a computer, the computer is caused to execute the method according to any one of the first aspect.

[0031] In a fourth aspect, an electronic device is provided, including:

[0032] A memory for storing a computer program;

[0033] A processor for executing the computer program to implement the method according to any one of the first aspect.

[0034] The beneficial effects of the present invention are:

[0035] 1) The artificial intelligence-based real-time metal deposition defect detection method provided by the present invention takes deposition defects as the object, designs an artificial intelligence algorithm model for end-to-end real-time object detection (RT-DETR), has a fast training speed and strong generalization ability, and is applicable to deposition application scenarios such as arc additive manufacturing that require efficient, real-time, and accurate detection.

[0036] 2) The artificial intelligence-based real-time metal deposition detection method provided by the present invention has higher detection accuracy compared with traditional detection methods, can more accurately identify and classify targets, and effectively solves the problems of low detection accuracy and slow detection efficiency in the deposition defect detection during the arc additive process and the weld welding process by combining the conversion architecture with real-time processing optimization.

[0037] 3) The metal real-time deposition detection method based on artificial intelligence provided by the present invention extracts key deposition features through the backbone network to improve the detection accuracy of deposition defects, improves the capture and fusion of deposition features through the efficient hybrid encoder to improve the accuracy and efficiency of the detection model, improves the prediction framework accuracy through the intersection over union (IoU)-aware query selection, and improves the classification and bounding box prediction accuracy through the decoder with an auxiliary prediction head, realizing the real-time detection of deposition defects in metal materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a schematic diagram of the overall process of the metal real-time deposition defect detection method based on artificial intelligence of the present invention;

[0039] Figure 2 is a schematic diagram of the overall real-time detection framework;

[0040] Figure 3a is a typical schematic diagram of a target deposition defect image with "good" quality;

[0041] Figure 3b is a schematic diagram of burrs with "poor" quality in a target deposition defect image;

[0042] Figure 3c is a schematic diagram of holes with "poor" quality in a target deposition defect image;

[0043] Figure 4a is a schematic diagram of the stitching of deposition defect images after Mosaic data augmentation;

[0044] Figure 4b is a schematic diagram of the stitching of another deposition defect image after Mosaic data augmentation;

[0045] Figure 5 is a schematic diagram of the architecture of the PResNet50 backbone network;

[0046] Figure 6 is a schematic diagram of the architecture of the intra-scale feature interaction (AIFI);

[0047] Figure 7 is a schematic diagram of the architecture of the cross-scale feature fusion (CCFM);

[0048] Figure 8 is a schematic diagram of the architecture of the IoU-aware query selection and decoder in the end-to-end real-time object detection (RT-DETR);

[0049] Figure 9a is a curve of the evolution process of mAP@0.5 during the real-time detection training process;

[0050] Figure 9b is a curve of the evolution process of mAP@0.5:0.95 during the real-time detection training process;

[0051] Figure 10a is the original deposition image of the arc additive manufacturing process;

[0052] Figure 10b is the marked deposition image of the arc additive manufacturing process;

[0053] Figure 10c is the image of the real-time detection result of deposition defects in the arc additive manufacturing process. Detailed implementation manners

[0054] The present invention will be further described below in conjunction with embodiments. The description of the following embodiments is only used to help understand the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0055] Embodiment 1:

[0056] To solve the problems of the prior art, Embodiment 1 of the present application provides a method for real-time detection of metal deposition defects based on artificial intelligence, as Figure 1 and Figure 2 shown, including:

[0057] S1. Obtain a deposition defect data set, and extract key deposition features by using a backbone network.

[0058] In S1, the deposition defects include forms such as deposition burrs, deposition holes, and deposition dents, which are mainly caused by improper deposition current and deposition speed. The deposition defects will seriously affect the integrity and safety of arc additive manufacturing structural parts or welded joint structural parts.

[0059] As shown in Figure 3, the deposition defect data set is marked as "good" ( Figure 3b ) or "bad" ( Figure 3c ) according to the presence or absence of defects such as burrs ( Figure 3a ) and holes ( Figure 3b - Figure 3c ), and is annotated; the deposition defect data set should cover common deposition defects, and each type of deposition defect is carefully annotated to ensure accuracy and diversity; these target deposition defect images provide detailed information on possible quality problems during arc additive manufacturing or welded joint welding processes, making the deposition defect data set highly representative and practical; for model training and performance evaluation, the deposition defect data set is divided into a training set and a validation set in a ratio of 8:2, and this division method can ensure sufficient training data and reliable validation results.

[0060] In S1, through the multi-layer convolution operation module of the PResNet50 backbone network, the perception ability and extraction efficiency of the model algorithm for sedimentation features are improved; through the enhanced residual block design module of the PResNet50 backbone network, the retention of sedimentation information is improved, and the detection accuracy of sedimentation defects is enhanced. Among them, as Figure 5 shown, the PResNet50 backbone network is also improved. The 7×7 convolutional layer is replaced by three 3×3 convolutional layers to perform operations, enhancing the extraction of sedimentation features while maintaining the same output; in the first residual block of the second, third, and fourth stages, the 1×1 convolutional layer (kernel size 1×1, stride 2) is replaced by a 2×2 average pooling layer (kernel size 2×2, stride 2) and a 1×1 convolutional layer to perform operations, so as to better retain the key sedimentation information and improve the efficiency of sedimentation feature extraction; through the above operations, smaller feature maps are used to significantly reduce the computational cost and improve the timeliness of sedimentation defect detection.

[0061] In addition, in S1, it also includes: expanding the sedimentation defect data set through data augmentation methods. As shown in Figure 4, for sedimentation defect detection, combining pixel-level data augmentation methods can provide more comprehensive and effective training data, improving the performance and generalization ability of the model; using the Mosaic data augmentation method to achieve data set expansion. First, randomly select four pictures, and through random scaling, cropping, and arrangement, splice the four pictures together to form a new picture ( Figure 4a 、 Figure 4b ); the sedimentation defects of the four pictures may appear in different positions of the new picture. Therefore, the training model needs to learn to detect objects at different positions and scales, making it more adaptable and robust.

[0062] S2: Combine the attention mechanism and multi-layer convolution operations to form an efficient hybrid encoder, and capture high-level semantic features in the sedimentation features and perform multi-scale feature fusion through the efficient hybrid encoder.

[0063] In S2, the efficient hybrid encoder is used to improve sedimentation feature capture and fusion. Specifically, the efficient hybrid encoder includes an intra-scale feature interaction (AIFI) module based on the attention mechanism and a cross-scale feature fusion (CCFM) module based on the convolutional neural network (CNN); the intra-scale feature interaction module is used to capture high-level semantic features in the sedimentation features to distinguish various sedimentation defects and reduce computational redundancy; the cross-scale feature fusion module uses convolutional blocks to fuse multi-scale features, enhancing the overall sedimentation feature information and improving the accuracy and efficiency of the detection model.

[0064] As Figure 6 、 Figure 7As shown, in step S2, the deposition features S3, S4, and S5 respectively come from the second, third, and fourth stages of the backbone network in step S1; compared with the shallower deposition features S3 and S4, the deeper deposition feature S5 contains more advanced and richer deposition semantic features, which is crucial for the Transformer decoder to distinguish different deposition features; since the semantic features of the shallow features in the deposition features are relatively few, for the in-scale feature interaction on the deposition feature S5 ( Figure 6 ), for the deposition features S3 and S4, a convolutional layer is used for processing to extract the deposition features; the deposition features are cross-scale fused through the feature maps output by the encoder ( Figure 7 ), to achieve multi-scale deposition feature acquisition and significantly reduce the computational cost;

[0065] In the feature fusion path, multiple convolutional fusion blocks are inserted to merge adjacent feature maps into a new feature map; the network module adopts a parallel structure to extract different degrees of deposition features from these deposition feature maps, and performs addition and dimensionality transformation on these deposition feature maps; the three different-scale deposition feature maps are combined together to enhance the effect of deposition feature fusion;

[0066] S3. Using the intersection over union (IoU)-aware query selection method, corresponding classification scores are assigned to the deposition features with different IoUs during the training process to constrain the model learning, and more accurate deposition features are selected in the decoding stage by optimizing the object queries.

[0067] S4. Using a decoder with an auxiliary prediction head to map the optimized object to the classification confidence and the bounding box, and accelerating the training convergence through the denoising module in the decoder.

[0068] Example 2:

[0069] Based on Example 1, Example 2 of the present application provides a more specific artificial intelligence-based real-time metal deposition defect detection method, including:

[0070] S1. Obtain a deposition defect dataset and use the backbone network to extract key deposition features.

[0071] S2. Combine the attention mechanism and multi-layer convolutional operations to form an efficient hybrid encoder, and capture the advanced semantic features in the deposition features and perform multi-scale feature fusion through the efficient hybrid encoder.

[0072] S3. Using the intersection over union (IoU)-aware query selection method, corresponding classification scores are assigned to the deposition features with different IoUs during the training process to constrain the model learning, and more accurate deposition features are selected in the decoding stage by optimizing the object queries.

[0073] In S3, the intersection over union (IoU)-aware query selection is used to improve the accuracy of the prediction framework. Specifically, high and low classification scores are assigned to sedimentary features with high and low IoU values respectively to constrain the model learning and improve the accuracy of the prediction framework, effectively solving the problem of inconsistent classification scores and IoU scores and preventing the impact of prediction frameworks with high classification scores but inaccurate positions on the detection results. By optimizing object queries, more accurate sedimentary features are selected at the decoding stage, ultimately improving the accuracy and robustness of sedimentary defect detection.

[0074] As Figure 8 shown, in S3, the IoU-aware query selection method is used to constrain the model during training to ensure the consistency between classification and IoU scores. Based on this, object queries contain more encoder features, which are accurately classified and precisely located, thereby improving the accuracy of the detector.

[0075] In addition, precision (P), recall (R), and mean average precision (mAP) are used as evaluation metrics, and the intersection over union (IoU) between the predicted and actual bounding boxes is used as the evaluation threshold. The IoU threshold is divided into intervals from 0.50 to 0.95 with a step size of 0.05, i.e., the average precision (AP) from 0.50 to 0.95 is calculated by sampling every 0.05 within this range, and the final AP is the average of these individual precision calculations. AP50 represents the AP measurement when the IoU threshold is 0.5. In the object detection model, mAP measures the average maximum precision of each category at different recall values R, balancing precision and recall and providing a single numerical performance metric. The calculation formulas for P, R, mAP, and AP are as follows:

[0076]

[0077] In the formulas, TP, FP, and FN represent "true positives", "false positives", and "false negatives" respectively, which are defined according to the IOU between the predicted bounding box and the ground truth. is the precision-recall curve the recall rate at any point on it; k is the interpolation point number; N is the total number of interpolation points; is the precision rate when the recall rate is ; ΔR(k) is the change in recall rate between the (k - 1)-th and k-th interpolation points; n is the total number of categories in multi-class object detection; i is the category number. If the IOU is greater than the threshold, the bounding box is marked as TP, representing the number of correctly recognized objects; otherwise, it is marked as FP, representing the number of misrecognized objects.

[0078] In addition, in the study of detection speed, frames per second (FPS) is used as the main metric to evaluate the detection speed, representing the number of images that the model can process per second. The higher the FPS, the faster the model processing speed and the higher the real-time performance.

[0079] S4. Use a decoder with an auxiliary prediction head to map the optimization object into classification confidence and bounding boxes, and accelerate the training convergence through the denoising module in the decoder.

[0080] In S4, use a decoder with an auxiliary prediction head to improve the classification and bounding box prediction accuracy. Specifically, after optimizing the object query through the Transformer decoder, the auxiliary prediction head maps the optimization object into classification confidence and bounding boxes to accurately detect the deposition defect features; the training convergence is accelerated through the denoising module in the decoder, enabling the training model to learn the positions and classifications of deposition defects more quickly and accurately.

[0081] As Figure 8 shown, in step S4, during the training process, the ground truth box is denoised and used as part of the decoder input; enabling the decoder's query to more accurately predict specific targets and accelerating the convergence rate of the training process;

[0082] As shown in Figure 9, the hyperparameter settings for network training are also crucial; if the learning rate is too high, the model may eventually oscillate around the optimal solution; if too low, more training phases may be required, resulting in inefficient use of resources; to solve this problem, a piecewise decay schedule and linear warm-up are used to set the learning rate; in this embodiment, the base learning rate is set to 0.0001, and the piecewise decay coefficient is 0.9, that is, for every 100 training times, the learning rate is multiplied by 0.9; the linear warm-up learning rate starts from 0.00001 and gradually increases to the base learning rate of 0.001 in the first 2000 steps; the optimizer used is Adam W3d, and the weight decay is 0.05 to prevent overfitting caused by large penalty weight values; during the training process, gradient clipping is used to limit the gradient value and avoid gradient explosion; the maximum gradient norm is set to 0.1; the evolution process of mAP@0.5 (mean average precision within the IoU threshold range of 0.5) ( Figure 9a ) and the evolution process of mAP@0.5:0.95 (mean average precision within the IoU threshold range of 0.5 - 0.95) ( Figure 9b ) are obtained during the training process of the real-time detection method (RT-DETR) of the present invention.

[0083] In addition, the embodiments of the present application also provide real-time detection of deposition defects in the application scenario of wire arc additive manufacturing (WAAM) for the above method.

[0084] During the metal arc additive manufacturing process, the quality and geometry of the deposited part are crucial for the strength and overall quality of the formed structural component. However, manual inspection in a high-temperature environment poses serious safety hazards and is prone to subjective errors. The real-time detection method for deposition defects proposed in this invention aims to achieve end-to-end automatic real-time detection of deposition defects during the arc additive process, significantly improving the detection accuracy and timeliness.

[0085] As shown in Figure 10, the deposition process of metal wire is carried out using an arc additive manufacturing system. The arc additive manufacturing system includes a flexible six-axis robot and a stable fuse power supply. The deposition is carried out using 316L stainless steel wire to obtain the original deposition image.

[0086] As shown in Figure 10, the original deposition image ( Figure 10a ) of the arc additive process is cropped into an image with a resolution of 800×800 pixels and marked according to the requirements in step S1 to obtain the marked deposition image ( Figure 10b ) of the arc additive process. It can be seen that there are significant differences in the deposition quality during the arc additive process, indicating that various defects and inconsistencies may occur during the deposition process. This emphasizes the importance of real-time deposition defect detection, which can be identified and corrected in a timely manner to ensure that the deposition quality meets the standards. Real-time detection also reduces rework and material waste, improving production efficiency and safety.

[0087] The detection details of the real-time detection result image ([[]] Figure 10c [[]]) of the deposition defects during the arc additive process are shown in Table 1. It can be seen that the mAP@0.5:0.95 (mean average precision within the IoU threshold range of 0.5 to 0.95) of the real-time detection method (RT-DETR) of this invention is 0.801, which has the highest accuracy compared with other models. The FPS (frames per second) of the real-time detection method of this invention is 67, which is better than other models, reflecting its advantages in terms of accuracy and timeliness.

[0088] Table 1 Comparison of detection performance of different models

[0089]

[0090] It should be noted that the same or similar parts in this embodiment and Embodiment 1 can be referred to each other and will not be elaborated in this application.

[0091] Embodiment 3:

[0092] Based on Embodiments 1 and 2, Embodiment 3 of this application provides an artificial intelligence-based real-time metal deposition defect detection system, including:

[0093] An acquisition module for acquiring a deposition defect data set and extracting key deposition features using a backbone network; ​​

[0094] A composition module, which is used to combine an attention mechanism and multi-layer convolutional operations to form an efficient hybrid encoder, and capture high-level semantic features in the deposition features through the efficient hybrid encoder and perform multi-scale feature fusion;

[0095] An allocation module, which is used to use the intersection over union (IoU)-aware query selection method to assign corresponding classification scores to deposition features with different IoUs during the training process to constrain the model learning, and select more accurate deposition features during the decoding phase by optimizing object queries;

[0096] A mapping module, which is used to map the optimized object to classification confidence and bounding boxes by using a decoder with an auxiliary prediction head, and accelerate the training convergence through the denoising module in the decoder.

[0097] It should be noted that the system provided in this embodiment is the system corresponding to the methods provided in Embodiments 1 and 2. Therefore, for the parts that are the same or similar in this embodiment and Embodiments 1 and 2, reference can be made to each other and will not be elaborated in this application.

[0098] In summary, the artificial intelligence-based real-time metal deposition defect detection method provided by the present invention extracts key deposition features through a backbone network to improve the deposition defect detection accuracy, improves the deposition feature capture and fusion through an efficient hybrid encoder to improve the accuracy and efficiency of the detection model, improves the prediction framework accuracy through IoU-aware query selection, and improves the classification and bounding box prediction accuracy through a decoder with an auxiliary prediction head; taking deposition defects as the object, designs an artificial intelligence algorithm model for end-to-end real-time object detection (RT-DETR), which has a fast training speed and strong generalization ability, and is applicable to application scenarios of metal wire deposition forming structural parts that require efficient, real-time, and accurate detection such as arc additive manufacturing. And through actual verification, the method of the present invention is effective.

Claims

1. A metal real-time deposition defect detection method based on artificial intelligence, characterized in that: include: S1. Obtain a deposition defect dataset and use the backbone network to extract key deposition features; S2, combining the attention mechanism and multi-layer convolution operations to form an efficient hybrid encoder, through which the high-level semantic features in the deposition features are captured and multi-scale feature fusion is performed; S3. Using the IoU-aware query selection method, the corresponding classification scores are assigned to the deposition features with different IoU ratios during the training process to constrain the model learning, and more accurate deposition features are selected in the decoding stage by optimizing the object query; S4. Mapping the optimized object to the classification confidence and bounding box using a decoder with an auxiliary prediction head, and accelerating training convergence through a denoising module in the decoder.

2. The method for real-time metal deposition defect detection based on artificial intelligence according to claim 1, characterized in that: S1 also includes: expanding the deposition defect data set by using a data enhancement method.

3. The method for real-time metal deposition defect detection based on artificial intelligence according to claim 2, characterized in that: In S1, the deposition defect dataset is marked as "good" or "bad" according to the presence or absence of the defect form, and annotated; the deposition defects include deposition burrs, deposition holes and deposition dents; the deposition defect dataset is divided into a training set and a validation set, which are used for model training and performance evaluation respectively; In S1, the backbone network is an improved PResNet50, in which the initial 7×7 convolutional layer is replaced by three 3×3 convolutional layers, and a 2×2 average pooling layer is introduced in the residual block to replace some 1×1 convolutional layers; the improved PResNet50 is used to enhance the retention of deposition key information and improve the efficiency of deposition feature extraction.

4. The method for real-time metal deposition defect detection based on artificial intelligence according to claim 3 is characterized in that: In S2, the efficient hybrid encoder includes an intra-scale feature interaction module based on an attention mechanism and a cross-scale feature fusion module based on a convolutional neural network; The intra-scale feature interaction module is used to capture high-level semantic features in sedimentary features; the cross-scale feature fusion module is used to fuse multi-scale features to enhance overall sedimentary feature information; In S2, for shallow sedimentary features, convolutional layers are used for extraction; for deep sedimentary features, high-level sedimentary semantic features are included, and sedimentary feature maps output by efficient hybrid encoders are used for cross-scale fusion to extract multi-scale sedimentary features, thereby improving the efficiency of sedimentary feature extraction; In S2, a parallel structure is used to extract sedimentary features of different degrees from the sedimentary feature map, and sedimentary feature maps of multiple different scales are combined to enhance the fusion effect of sedimentary features.

5. The method for real-time metal deposition defect detection based on artificial intelligence according to claim 4 is characterized in that: In S3, precision, recall and average precision of all categories are used as evaluation indicators, and the intersection-over-union ratio between the predicted and actual marked boxes is used as the evaluation threshold; the calculation formulas for precision, recall, average precision and average precision of all categories are: Among them, P represents precision, R represents recall, AP represents average precision, mAP represents average precision of all categories, TP, FP and FN represent "true positive", "false positive" and "false negative" respectively, which are defined based on the intersection-over-union ratio between the predicted bounding box and the reference truth value; It is the precision P-recall R curve The recall rate of any point on ; k is the interpolation point number; N is the total number of interpolation points; The recall rate is ; ΔR(k) is the change in recall between the k-1th and kth interpolation points; n is the total number of categories for multi-category target detection; i is the category number; when the IoU is greater than the threshold, the bounding box is marked as TP, representing the number of correctly identified objects; when the IoU is less than or equal to the threshold, the bounding box is marked as FP, representing the number of incorrectly identified objects; in S3, when evaluating the detection speed, the number of frames per second is used as the main indicator for evaluating the detection speed, expressed in FPS.

6. The method for real-time metal deposition defect detection based on artificial intelligence according to claim 5 is characterized in that: In S4, during the training process, the reference truth boxes are denoised and used as part of the decoder input to speed up the training convergence; A piecewise decay schedule and linear warm-up are used to set the learning rate to solve the training hyperparameter problem.

7. The artificial intelligence-based metal real-time deposition defect detection system is characterized by: Used to perform the method according to any one of claims 1 to 6, comprising: An acquisition module, used to acquire a deposition defect dataset and extract key deposition features using a backbone network; A construction module, used to combine an attention mechanism and multi-layer convolution operations to form an efficient hybrid encoder, through which high-level semantic features in the deposition features are captured and multi-scale feature fusion is performed; An assignment module is used to assign corresponding classification scores to deposition features with different IoU ratios during training using an IoU-aware query selection method to constrain model learning and select more accurate deposition features during decoding by optimizing object queries; A mapping module is used to map the optimized object into a classification confidence and a bounding box using a decoder with an auxiliary prediction head, and accelerate training convergence through a denoising module in the decoder.

8. A computer storage medium, characterized in that: The computer storage medium stores a computer program; when the computer program is executed on a computer, the computer executes any one of the methods described in claims 1 to 6.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 6.

Images