A method, system, medium, and apparatus for real-time detection of defects in laser powder bed fusion additive manufacturing

Abstract

This invention discloses a method, system, medium, and equipment for real-time defect detection in laser powder bed fusion additive manufacturing, belonging to the field of additive manufacturing quality monitoring and image processing. The method includes: acquiring image data and constructing a dataset; replacing the standard convolutional modules in the backbone and neck networks of the YOLOv11 network with frequency-domain dynamic convolutional modules; embedding the detection head of this network into a parallelized patch-aware attention module; training the improved network; performing real-time defect detection and outputting the results. This method forms a "detailed supply-semantic filtering" mechanism through the synergistic effect of the frequency-domain dynamic convolutional module and the parallelized patch-aware attention module. It enriches detailed features at a shallow level and performs intelligent filtering and fusion at a deep level, effectively improving the model's detection accuracy, enhancing its adaptability to varying process environments, and significantly improving recall and localization accuracy. This provides an effective solution for real-time industrial quality control in laser powder bed fusion additive manufacturing processes.

Description

Technical Field

[0001] This application relates to the field of additive manufacturing quality monitoring and image processing technology, and in particular to a method, system, medium and equipment for real-time detection of defects in laser powder bed fusion additive manufacturing. Background Technology

[0002] Laser Powder Bed Fusion (LPBF) technology manufactures complex three-dimensional components by melting metal powder layer by layer, and has become an important additive manufacturing technology in aerospace, biomedicine, and other fields. However, this process involves complex multiphysics interactions between the laser, powder, and molten pool, which easily generates various microscopic defects, including lap defects, spatter defects, and fusion abnormalities such as inadequate or over-fusion. These defects significantly reduce the mechanical properties, fatigue life, and service reliability of the components, becoming a major bottleneck restricting the expansion of this technology into the manufacturing of critical load-bearing components.

[0003] Traditional quality assurance methods, such as industrial computed tomography (CT) and scanning electron microscopy (SEM), are offline, destructive, or costly inspection methods, and cannot be used for real-time diagnosis and closed-loop control of the manufacturing process. Therefore, developing machine vision-based online monitoring and intelligent defect detection technologies has significant engineering value for achieving real-time quality control of the laser powder bed melting process, improving manufacturing yield, and reducing production costs.

[0004] Therefore, how to achieve high-precision real-time detection of minute or low-contrast defects during laser powder bed fusion additive manufacturing, while ensuring that the model has good adaptability to varying process environments and maintains real-time inference speed, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, medium, and equipment for real-time detection of defects in laser powder bed fusion additive manufacturing, addressing the aforementioned technical problems.

[0006] The following technical solution is adopted in this specification: This specification provides a method for real-time defect detection in laser powder bed fusion additive manufacturing, specifically including: Image data of the laser powder bed fusion additive manufacturing process were collected to construct a multi-energy condition defect dataset containing lap defects, spatter defects, and fusion anomalies.

[0007] A YOLOv11 real-time object detection network is introduced, and the C3k2 standard convolutional modules in its backbone and neck networks are replaced with frequency-domain dynamic convolutional modules. At the same time, a parallelized patch-aware attention module is embedded in the detection head of the YOLOv11 real-time object detection network model to obtain an improved YOLOv11 real-time object detection network, which is used as a defect detection network.

[0008] The frequency domain dynamic convolution module maps convolution operations to the frequency domain for frequency band decomposition and adaptive modulation; the parallelized patch-aware attention module extracts features at different scales through parallel local patch branches, global patch branches, and hierarchical cascaded branches, and performs adaptive weighted fusion through a channel-space attention fusion mechanism.

[0009] The defect detection network is trained based on the multi-energy condition defect dataset to obtain a trained defect detection network.

[0010] The real-time monitoring image of the laser powder bed fusion additive manufacturing process to be inspected is input into the trained defect detection network for real-time defect detection, and the defect detection results are output.

[0011] Furthermore, the frequency domain dynamic convolution module includes a frequency domain parameter grouping unit, a kernel spatial modulation unit, and a frequency band modulation unit; the frequency domain parameter grouping unit divides the learnable parameters into multiple discrete frequency groups in the Fourier domain, each group corresponding to a specific frequency range; the kernel spatial modulation unit adaptively adjusts the convolution kernel element-wise based on local features and global context information; the frequency band modulation unit implements spatially variable frequency band weighting, applying different frequency enhancement strategies to different image regions.

[0012] Furthermore, the local patch branch uses a first-size pixel patch for convolution operation to extract the detailed features of the splashing defect; the global patch branch uses a second-size pixel patch for convolution operation to capture the contextual information of the overlapping and fusion defects; the hierarchical cascaded branch, through cascaded convolutional layers, combines the detailed features of the splashing defect with the contextual information of the overlapping and fusion defects to gradually construct multi-scale feature semantic associations.

[0013] Furthermore, the first-sized pixel patch is smaller than the second-sized pixel patch.

[0014] Furthermore, the method for constructing a multi-energy condition defect dataset including lap defects, spatter defects, and fusion anomalies specifically involves: calculating the energy density of the image data of the laser powder bed fusion additive manufacturing process based on the laser power, scanning speed, scanning spacing, and powder layer thickness; and dividing the image data of the laser powder bed fusion additive manufacturing process into a low-energy input group, a medium-energy input group, and a high-energy input group based on the relationship between the bulk energy density and a first threshold and a second threshold.

[0015] Furthermore, the first threshold is less than the second threshold; the volume energy density of the low-energy input group is not greater than the first threshold; the volume energy density of the medium-energy input group is greater than the first threshold and not greater than the second threshold; and the volume energy density of the high-energy input group is greater than the second threshold.

[0016] Furthermore, the defect detection results specifically include defect category, bounding box location, and confidence score.

[0017] This specification provides a real-time defect detection system for laser powder bed fusion additive manufacturing, specifically including: The dataset construction module is used to collect image data of the laser powder bed fusion additive manufacturing process and construct a multi-energy condition defect dataset.

[0018] The YOLOv11 Real-Time Target Detection Network Improvement Module is used to improve the backbone network, neck network, and detection head of the YOLOv11 Real-Time Target Detection Network to obtain a defect detection network.

[0019] The network training module is used to train the defect detection network using a multi-energy conditional defect dataset.

[0020] The defect detection network application module is used to input real-time monitoring images of the laser powder bed fusion additive manufacturing process to be inspected into the trained defect detection network, perform real-time defect detection, and output the defect detection results.

[0021] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0022] This specification provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0023] The beneficial effects of this invention are as follows: Compared with the prior art, this invention replaces the standard convolution modules in the YOLOv11 backbone and neck network with a frequency-domain dynamic convolution module, mapping the convolution operation to the frequency domain for frequency band decomposition and adaptive modulation. This enables the network to dynamically adjust the response characteristics of the convolution kernel and prioritize the extraction of frequency features most relevant to defects under the current process conditions. Simultaneously, by embedding a parallelized patch-aware attention module consisting of local, global, and hierarchical cascaded branches in the detection head, and utilizing a channel-spatial attention fusion mechanism to adaptively weight and fuse semantic information at different scales, the model can accurately focus on small or low-contrast defect regions. Through the synergistic effect of the frequency-domain dynamic convolution module and the parallelized patch-aware attention module, a "detail supply-semantic filtering" mechanism is formed, enriching detailed features at a shallow level and performing intelligent filtering and fusion at a deep level. This effectively improves the model's detection accuracy, enhances its adaptability to varying process environments, and significantly improves recall and localization accuracy, providing an effective solution for real-time industrial quality control in laser powder bed fusion additive manufacturing processes. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the LPBF process parameters and sample configurations at different energy input levels in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of the improved YOLOv11 defect detection network described in this invention; Figure 4 This is a schematic diagram of the frequency domain dynamic convolution module of the present invention; Figure 5 This is a schematic diagram of the parallelized patch-aware attention module of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0027] This invention provides a real-time defect detection method for laser powder bed fusion additive manufacturing, comprising four steps: acquiring image data and constructing a multi-energy condition defect dataset; replacing the standard convolutional modules in the backbone and neck networks of the YOLOv11 real-time target detection network with frequency domain dynamic convolutional modules; embedding a parallelized patch-aware attention module into the network's detection head; training the improved defect detection network; and performing real-time defect detection and outputting the results. This method utilizes the synergistic effect of the frequency domain dynamic convolutional module and the parallelized patch-aware attention module to form a "detailed feature supply-semantic filtering" mechanism. It enriches detailed features at a shallow level and performs intelligent filtering and fusion at a deep level, effectively improving the model's detection accuracy, enhancing its adaptability to varying process environments, and significantly improving recall and localization accuracy. This provides an effective solution for real-time industrial quality control in laser powder bed fusion additive manufacturing.

[0028] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0029] A method for real-time defect detection in laser powder bed fusion additive manufacturing, such as Figure 1 As shown, it specifically includes: Step 1: Collect image data of the laser powder bed fusion additive manufacturing process and construct a multi-energy condition defect dataset containing lap defects, spatter defects, and fusion anomaly defects.

[0030] Preferably, the method for constructing a multi-energy condition defect dataset including lap defects, spatter defects, and fusion anomaly defects is as follows: based on the laser power, scanning speed, scanning spacing, and powder layer thickness, the energy density of the image data of the laser powder bed fusion additive manufacturing process is calculated, and based on the relationship between the volume energy density and the first and second thresholds, the image data of the laser powder bed fusion additive manufacturing process is divided into a low-energy input group, a medium-energy input group, and a high-energy input group.

[0031] Preferably, the first threshold is less than the second threshold; the body energy density of the low energy input group is not greater than the first threshold; the body energy density of the medium energy input group is greater than the first threshold but not greater than the second threshold; and the body energy density of the high energy input group is greater than the second threshold.

[0032] Step 2: Introduce the YOLOv11 real-time object detection network, replacing the standard C3k2 convolutional modules in its backbone and neck networks with frequency-domain dynamic convolutional modules; simultaneously, embed a parallelized patch-aware attention module into the detection head of the YOLOv11 real-time object detection network model, resulting in... Figure 3 The improved YOLOv11 real-time object detection network shown is used as a defect detection network.

[0033] like Figure 4 As shown, the frequency domain dynamic convolution module maps convolution operations to the frequency domain for frequency band decomposition and adaptive modulation; as Figure 5 As shown, the parallelized patch-aware attention module extracts features at different scales through parallel local patch branches, global patch branches, and hierarchical cascade branches, and performs adaptive weighted fusion through a channel-space attention fusion mechanism.

[0034] Preferably, the frequency domain dynamic convolution module includes a frequency domain parameter grouping unit, a kernel spatial modulation unit, and a frequency band modulation unit; the frequency domain parameter grouping unit divides the learnable parameters into multiple discrete frequency groups in the Fourier domain, each group corresponding to a specific frequency range; the kernel spatial modulation unit adaptively adjusts the convolution kernel element by element based on local features and global context information; the frequency band modulation unit implements spatially variable frequency band weighting, applying different frequency enhancement strategies to different image regions.

[0035] Preferably, the local patch branch uses a first-size pixel patch for convolution operation to extract the detailed features of the splashing defect; the global patch branch uses a second-size pixel patch for convolution operation to capture the contextual information of the overlapping and fusion defects; the hierarchical cascaded branch uses cascaded convolutional layers to combine the detailed features of the splashing defect with the contextual information of the overlapping and fusion defects to gradually construct multi-scale feature semantic associations.

[0036] Preferably, the first-size pixel patch is smaller than the second-size pixel patch.

[0037] Step 3: Train the defect detection network based on the multi-energy condition defect dataset to obtain the trained defect detection network.

[0038] Step 4: Input the real-time monitoring image of the laser powder bed fusion additive manufacturing process to be inspected into the trained defect detection network, perform real-time defect detection, and output the defect detection results.

[0039] Preferably, the defect detection results specifically include defect category, bounding box location, and confidence score.

[0040] Example This embodiment provides a real-time defect detection method for additive manufacturing that integrates frequency-domain dynamic convolution and multi-scale attention mechanisms, specifically including: S1: Dataset Construction Acquire in-situ monitoring images of the laser powder bed fusion additive manufacturing process, such as... Figure 2 As shown, different process parameters such as laser power, scanning speed, scanning spacing, and powder layer thickness were set to obtain image samples containing overlapping defects, spatter defects, and fusion abnormalities. Based on volume energy density, the data was divided into low-energy input group, medium-energy input group, and high-energy input group, completing the annotation and dataset partitioning.

[0041] S2: Data Preprocessing and Data Augmentation The original images are dynamically adjusted, contrast enhanced, and noise suppressed to eliminate environmental noise and improve defect identification. Data augmentation strategies such as geometric transformation, photometric adjustment, and random erasure are used to expand sample diversity.

[0042] S3: Construct a defect detection network based on improved YOLOv11 Based on the YOLOv11 framework, a defect detection network consisting of a backbone network, a neck network, and a detection head is constructed. The C3k2 modules in the backbone and neck networks are replaced with frequency-domain dynamic convolutional enhancement modules to perform feature decomposition, kernel spatial modulation, and frequency band modulation in the frequency domain, achieving adaptive feature extraction in the frequency domain. A parallelized patch-aware attention module is embedded in the detection head, extracting multi-scale features in parallel through local patch branches, global patch branches, and hierarchical cascaded branches, and achieving feature weighting enhancement through channel-spatial attention fusion.

[0043] S4: Model Training, Application, and Evaluation The improved YOLOv11 defect detection network was trained using a preprocessed and data-augmented multi-energy conditional defect dataset. Model convergence was achieved through an optimizer and learning rate scheduling strategy to obtain the optimal defect detection model. Real-time monitoring images of the laser powder bed melting process were input into the optimal defect detection model to achieve real-time defect detection and localization. Mean accuracy (mAP), recall, precision, and inference speed (FPS) were used as evaluation metrics. Comparative experiments were conducted to verify the detection performance of the proposed method, and the results are shown in Table 1.

[0044] This specification provides a real-time defect detection system for laser powder bed fusion additive manufacturing, specifically including: The dataset construction module is used to collect image data of the laser powder bed fusion additive manufacturing process and construct a multi-energy condition defect dataset.

[0045] The YOLOv11 Real-Time Target Detection Network Improvement Module is used to improve the backbone network, neck network, and detection head of the YOLOv11 Real-Time Target Detection Network to obtain a defect detection network.

[0046] The network training module is used to train the defect detection network using a multi-energy conditional defect dataset.

[0047] The defect detection network application module is used to input real-time monitoring images of the laser powder bed fusion additive manufacturing process to be inspected into the trained defect detection network, perform real-time defect detection, and output the defect detection results.

[0048] Specific limitations regarding the real-time defect detection system for laser powder bed fused additive manufacturing can be found in the above-described limitations regarding the real-time defect detection method for laser powder bed fused additive manufacturing, and will not be repeated here. Each module in the aforementioned real-time defect detection system for laser powder bed fused additive manufacturing can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0049] This specification also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0050] This specification provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0051] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0052] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for real-time detection of defects in laser powder bed fusion additive manufacturing, characterized in that, include: Image data of the laser powder bed fusion additive manufacturing process were collected to construct a multi-energy condition defect dataset containing lap defects, spatter defects, and fusion anomaly defects. A YOLOv11 real-time object detection network is introduced, and the C3k2 standard convolutional modules in its backbone and neck networks are replaced with frequency-domain dynamic convolutional modules. At the same time, a parallelized patch-aware attention module is embedded in the detection head of the YOLOv11 real-time object detection network model to obtain an improved YOLOv11 real-time object detection network, which is used as a defect detection network. The frequency domain dynamic convolution module maps convolution operations to the frequency domain for frequency band decomposition and adaptive modulation; the parallelized patch-aware attention module extracts features at different scales through parallel local patch branches, global patch branches and hierarchical cascaded branches, and performs adaptive weighted fusion through a channel-space attention fusion mechanism. The defect detection network is trained based on the multi-energy condition defect dataset to obtain a trained defect detection network. The real-time monitoring image of the laser powder bed fusion additive manufacturing process to be inspected is input into the trained defect detection network for real-time defect detection, and the defect detection results are output.

2. A method of real-time defect detection in laser powder bed fusion additive manufacturing as claimed in claim 1, wherein: The frequency domain dynamic convolution module includes a frequency domain parameter grouping unit, a kernel spatial modulation unit, and a frequency band modulation unit. The frequency domain parameter grouping unit divides the learnable parameters into multiple discrete frequency groups in the Fourier domain, each group corresponding to a specific frequency range. The kernel spatial modulation unit adaptively adjusts the convolution kernel element-wise based on local features and global context information. The frequency band modulation unit implements spatially variable frequency band weighting, applying different frequency enhancement strategies to different image regions.

3. A method of real-time defect detection in laser powder bed fusion additive manufacturing as claimed in claim 1, wherein: The local patch branch uses a first-size pixel patch for convolution to extract detailed features of the splashing defect; the global patch branch uses a second-size pixel patch for convolution to capture contextual information of the overlapping and fusion defects; the hierarchical cascaded branch uses cascaded convolutional layers to combine the detailed features of the splashing defect with the contextual information of the overlapping and fusion defects to gradually construct multi-scale feature semantic associations.

4. A method of real-time defect detection in laser powder bed fusion additive manufacturing as claimed in claim 3, wherein: The first pixel patch is smaller than the second pixel patch.

5. A method of real-time defect detection in laser powder bed fusion additive manufacturing as claimed in claim 1, wherein, The method for constructing a multi-energy condition defect dataset including lap defects, spatter defects, and fusion anomalies specifically involves: calculating the energy density of the image data of the laser powder bed fusion additive manufacturing process based on the laser power, scanning speed, scanning spacing, and powder layer thickness; and dividing the image data of the laser powder bed fusion additive manufacturing process into a low-energy input group, a medium-energy input group, and a high-energy input group based on the relationship between the volume energy density and a first threshold and a second threshold.

6. A method of real-time defect detection in laser powder bed fusion additive manufacturing as claimed in claim 5, wherein: The first threshold is less than the second threshold; the volume energy density of the low-energy input group is not greater than the first threshold; the volume energy density of the medium-energy input group is greater than the first threshold and not greater than the second threshold; the volume energy density of the high-energy input group is greater than the second threshold.

7. A method of real-time defect detection in laser powder bed fusion additive manufacturing as claimed in claim 1, wherein: The defect detection results specifically include defect category, bounding box location, and confidence score.

8. A laser powder bed fusion additive manufacturing defect real-time detection system, characterized in that include: The dataset construction module is used to collect image data of the laser powder bed fusion additive manufacturing process and construct a multi-energy condition defect dataset; The YOLOv11 real-time target detection network improvement module is used to improve the backbone network, neck network, and detection head of the YOLOv11 real-time target detection network to obtain a defect detection network. The network training module is used to train the defect detection network by combining a multi-energy conditional defect dataset; The defect detection network application module is used to input real-time monitoring images of the laser powder bed fusion additive manufacturing process to be inspected into the trained defect detection network, perform real-time defect detection, and output the defect detection results.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

Images