Multi-dimensional data fusion SLM defect type and distribution monitoring method
By integrating multi-dimensional data fusion and the Md-Mt-CNN convolutional neural network model, combined with visible light, laser reflection, and near-infrared signals, the lag problem in SLM forming defect detection was solved, achieving high-precision defect monitoring.
Patent Information
- Application Number
- CN202511593955.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-02-17
AI Technical Summary
Existing technologies for SLM forming defect detection suffer from lag, and single-dimensional monitoring is prone to missing information, making it difficult to accurately monitor defect types and distribution in real time.
A multi-dimensional data fusion method was adopted, which combined the synchronous acquisition of visible light, laser reflection and near-infrared shortwave signals. Defect identification was performed by the convolutional neural network model Md-Mt-CNN, and image alignment and feature extraction were performed by combining CT scan results.
It enables rapid real-time monitoring of defects during the SLM forming process, improves the accuracy and precision of defect type and distribution identification, and reduces the loss of monitoring information.
Smart Images
Figure CN121540735A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of online monitoring in additive manufacturing, and more specifically, relates to a multi-dimensional data fusion method for monitoring the types and distribution of defects in SLM (Soldered Materials Manufacturing). Background Technology
[0002] Selective Laser Melting (SLM), a type of powder bed melting technology, is one of the fastest-growing additive manufacturing technologies. The SLM forming process involves slicing a 3D CAD model into layers and manufacturing each layer sequentially. During the processing of each layer, a scraper spreads metal powder onto the substrate, and a laser beam scans along the slicing path to melt the powder. Through layer-by-layer melting and accumulation, a complete 3D part is finally formed.
[0003] In the SLM process, factors such as laser power, scanning rate, scanning spacing, gas composition in the forming chamber, and temperature have a significant impact on the SLM forming quality. Poor combinations of process parameters can easily lead to defects in the formed parts, resulting in a decline in the performance of the formed parts. In actual production, SLM forming defects mainly include three categories: first, macroscopic defects such as surface depressions, pores, inclusions, and cracks; second, microscopic defects such as internal pores, inclusions, lack of fusion, and cracks; and third, dimensional accuracy defects such as warping and shrinkage caused by stress concentration.
[0004] To detect and control the aforementioned defects, existing technologies typically employ the following methods: after the part is formed, defects are detected using methods such as X-ray, CT, and eddy current testing to identify their type and distribution, and then process parameters such as laser parameters and scanning paths are adjusted to reduce defect generation. However, traditional methods for defect detection and control in additive manufacturing have a significant lag; by this time, defects have already formed, and can only be addressed through post-processing after detection, rather than being detected and removed immediately after forming.
[0005] Furthermore, for some monitoring data commonly used in existing technologies, the monitoring data collected in the visible light dimension only indicates the quality of the material surface forming during the forming process and cannot reflect defects caused by temperature fluctuations in the formed part; the monitoring data in the infrared dimension can reflect temperature fluctuations, but it is difficult to show clear defect types and morphologies, making it easy for single-dimensional defect monitoring to miss monitoring information and lead to missed defects.
[0006] Accordingly, there is an urgent need in this field to study and improve these technical issues in order to better meet the high-quality and efficient online defect monitoring needs of SLM. Summary of the Invention
[0007] To address one or more of the above-mentioned deficiencies or needs in existing technologies, this invention provides a multi-dimensional data fusion method for monitoring the type and distribution of defects in SLM (Self-Made Molded Parts). By fully combining the working characteristics of SLM and the requirements for defect monitoring, multi-dimensional data is selected, collected and fused synchronously, and a monitoring model is designed based on this. Consequently, it can not only quickly and in real-time collect melt pool data during the forming process, but also more completely reflect the changes in the formed part during the SLM forming process, thereby further improving the accuracy of defect monitoring and enabling accurate monitoring of defect types and defect distribution in different regions.
[0008] To achieve the above objectives, according to the present invention, a multi-dimensional data fusion method for monitoring SLM defect types and distribution is provided, characterized in that the method includes the following steps: S1. Design experimental parameters and print SLM samples; S2. Synchronously acquire three-dimensional signals in the visible light band, laser reflection band, and near-infrared shortwave band during the printing process; S3. The three-dimensional signals are superimposed to obtain the corresponding monitoring images, namely, visible light monitoring image, laser reflection monitoring image and near-infrared shortwave monitoring image; S4. Use CT to scan the SLM sample, obtain the corresponding three-dimensional structure and cut it to remove the air phase. Then, divide the three-dimensional structure into multiple slice images along the Z-axis, which correspond to the thickness of the scanned layer. For each slice image, analyze the morphological characteristics and location distribution of unfused porosity defects and keyhole defects, complete the representation of the defect characteristics accordingly, and align the CT scan image with the monitoring image. S5. Take each of the monitored images as input and the defect features of the CT scan images as output to construct a dataset; S6. Based on the constructed dataset, a convolutional neural network is used to identify the defect types and defect distribution inside the SLM specimen.
[0009] As a further preferred embodiment of the present invention, in step S1, the experimental parameters are preferably designed using the orthogonal experimental method.
[0010] As a further preferred embodiment of the present invention, in step S2, the visible light band signal reflects the dynamic changes in the SLM additive manufacturing process, and the corresponding changes in light intensity are used to characterize the forming quality of the material surface; the laser reflection band signal reflects the direct emission and absorption rate of laser energy, and is used to characterize the reflected light characteristics and process energy utilization rate of the material; the near-infrared short-wave band signal reflects the molten pool temperature on the material surface, and is used to characterize the temperature change of the material.
[0011] As a further preferred embodiment of the present invention, step S2 preferably includes an operation of acquiring printing-related data, which includes laser start / stop signals, slicing laser scanning paths, etc.
[0012] As a further preferred embodiment of the present invention, in step S3, the monitoring image is preferably improved by using a spatial domain enhancement algorithm and a frequency domain enhancement algorithm to reduce noise interference during the forming process.
[0013] As a further preferred embodiment of the present invention, in step S4, for the alignment of the CT scan image and the monitoring image, it is preferable to start from the top layer of the block and perform image registration according to the number of layers and thickness.
[0014] As a further preferred embodiment of the present invention, in step S5, the monitoring images corresponding to the current printing layer and its upper and lower splicing layers are preferably combined into a three-channel image as input according to the forming sequence; correspondingly, the defect features of the current printing layer and its upper and lower splicing layers in the CT scan image are extracted and output according to the forming sequence, thereby reflecting the fluctuation between layers in the forming sequence.
[0015] As a further preferred embodiment of the present invention, in step S6, a convolutional neural network model is preferably used as the backbone network, and a multi-dimensional-multi-task cascaded neural network model Md-Mt-CNN is formed on this basis, thereby identifying the defect type, statistical characteristics, shape characteristics and spatial distribution characteristics of the defects inside the SLM sample.
[0016] As a further preferred embodiment of the present invention, for the multi-dimensional-multi-task cascaded neural network model Md-Mt-CNN, a three-dimensional input structure is preferably adopted, and a general feature layer is used to extract the input features of each dimension. Then, through decoupling and synchronous output, defect classification and feature regression prediction are realized simultaneously in one model.
[0017] As a further preferred embodiment of the present invention, the signal acquisition frequency of the monitoring image is related to the defect classification and prediction accuracy, and higher defect classification accuracy can be achieved by increasing the signal acquisition frequency, while the size of the defect features that can be identified is smaller.
[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. This invention fully combines the working conditions and defect monitoring requirements of SLM. By selecting three-dimensional signals—visible light band, laser reflection band, and near-infrared short-wave band—to perform synchronous acquisition, compared with conventional monitoring methods such as single visible light / near-infrared in the prior art, it can not only quickly and in real time acquire molten pool data during the forming process, but also more completely reflect the changes in the formed part during the SLM forming process. 2. The data acquisition process of this invention integrates more comprehensive monitoring information, and further uses CT scan results as a guide and can correspond the number of scan layers, thereby making the included SLM forming process information more accurate and reliable, which is conducive to improving the identification accuracy of defect types and distribution. 3. This invention takes the visible light signal, laser reflection signal and near-infrared shortwave signal collected in real time during the forming process as input, and can reconstruct the monitoring signal collected in each layer through image enhancement to generate a high-quality monitoring image; on this basis, by extracting the defect feature data in CT images and using a convolutional neural network for recognition, the entire defect monitoring process can be completed in a highly efficient and intelligent way, and the prediction function of defect distribution and feature regression can be realized at the same time. 4. The present invention further optimizes the specific types of defect monitoring models. Among them, by proposing a multi-dimensional, multi-task cascaded neural network model Md-Mt-CNN, it can not only extract the features of three types of monitoring data through convolution, but also achieve more refined defect classification and distribution monitoring under grid division by fusing the features of the three types of monitoring information. The defect monitoring accuracy can be significantly improved compared with the single-dimensional CNN model. Attached Figure Description
[0019] Figure 1 This is a basic flowchart of the SLM defect type and distribution monitoring method based on multi-dimensional data fusion according to the present invention; Figure 2 This is a technical schematic diagram illustrating the SLM defect type and distribution monitoring method according to the present invention; Figure 3 This is a schematic diagram used to demonstrate the reconstruction of the original three-dimensional monitoring signal and image; Figure 4 This is a schematic diagram used to demonstrate the alignment of CT scan images with three-dimensional monitoring images; Figure 5 This is a schematic diagram of a preferred embodiment of the present invention, used to display a multi-dimensional, multi-task cascaded neural network model Md-Mt-CNN; Figure 6 It is used for more specific display Figure 5 A schematic diagram of the 3D image input module and feature fusion module of the shared feature convolutional layer. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0021] Figure 1 This is a basic flowchart of the SLM defect type and distribution monitoring method based on multi-dimensional data fusion according to the present invention. Figure 2 This is a schematic diagram illustrating the technical principle of the SLM defect type and distribution monitoring method according to the present invention. Reference will be made below. Figure 1 and Figure 2 To explain the invention in more detail.
[0022] like Figure 1 and Figure 2 As shown, the defect type and distribution monitoring method of the present invention mainly includes the following operation steps.
[0023] Step 1: Design experimental parameters and print SLM samples; More specifically, in this step, for example, orthogonal experimental design can be used to design experimental parameters and print SLM samples.
[0024] Step 2: Simultaneously acquire three-dimensional signals in the visible light band, laser reflection band, and near-infrared shortwave band during the printing process; More specifically, in this step, appropriate sensors can be selected to simultaneously acquire signals in the visible light band, laser reflection band, and near-infrared shortwave band during the printing process. The acquisition frequency directly affects the accuracy of subsequent defect monitoring, so it can be adjusted according to actual needs. In addition, printing-related data such as laser start / stop signals, slicing laser scanning paths, and slicing files can also be acquired.
[0025] The reason for choosing the above three-dimensional signals in this invention is that the visible light band signal reflects the dynamic changes in the SLM additive manufacturing process, and the corresponding changes in light intensity can be used to characterize the forming quality of the material surface; the laser reflection band signal reflects the direct emission and absorption rate of laser energy, and can be used to characterize the reflected light characteristics and process energy utilization of the material; the near-infrared short-wave band signal reflects the molten pool temperature on the material surface, and can be used to characterize the temperature changes of the material.
[0026] Compared with conventional monitoring methods such as single visible light / near infrared in existing technologies, this method can not only quickly collect melt pool data in real time during the forming process, but also more completely reflect the changes in the formed part during the SLM forming process.
[0027] Step 3: The three-dimensional signals are superimposed to obtain the corresponding monitoring images, namely, the visible light monitoring image, the laser reflection monitoring image, and the near-infrared shortwave monitoring image.
[0028] More specifically, based on the laser start / stop signal, the laser scanning path of the slice, the slice file and other printed data, the three-dimensional signals of the visible light band, the laser reflection band and the near-infrared short-wave band can be superimposed to obtain visible light monitoring images, laser monitoring images and near-infrared short-wave monitoring images respectively.
[0029] See Figure 3 According to a preferred embodiment of the present invention, the original monitoring image can preferably be further improved in quality by using spatial domain enhancement algorithm and frequency domain enhancement algorithm to reduce noise interference during the forming process, and then the improved monitoring image is provided for subsequent steps.
[0030] Step 4: Use CT to scan the SLM sample, obtain the corresponding three-dimensional structure and cut it to remove the air phase. Then, divide the three-dimensional structure into multiple slice images along the Z-axis, which correspond to the thickness of the scanned layer. For each slice image, analyze the morphological characteristics and location distribution of unfused porosity defects and keyhole defects, and complete the representation of the defect characteristics accordingly. Align the CT scan image with the monitoring image.
[0031] More specifically, in this step, CT scanning of the SLM specimen is preferably used; for example, a high-precision scanning mode can be employed to acquire the three-dimensional structure of the SLM specimen and perform cutting. See [link / reference] Figure 4 The example demonstrates the operation of aligning a CT scan image with the monitoring image.
[0032] According to another preferred embodiment of the present invention, considering that the molded part is a block with a label on the top, with a clear outline and good molding continuity, and that each block has been monitored and CT scanned layer by layer, accordingly, for the alignment of the CT scan image and the monitoring image, it is preferred to start from the top layer of the block, and perform image registration according to the number of layers and thickness.
[0033] Step 5: Take each of the monitored images as input and the defect features of the CT scan images as output to construct a dataset.
[0034] Considering the forming sequence of upper and lower layers, the monitoring signals of the upper and lower layers can be used to reflect the influence of the upper and lower layers on the current printed layer. Therefore, according to another preferred embodiment of the present invention, in this step, the three-dimensional signals are not simply spliced together, but combined with the forming sequence, the monitoring images corresponding to the current printed layer and its upper and lower splicing layers are combined into a three-channel image as input. This can fully extract the fluctuation between layers in the forming sequence. Correspondingly, also according to the forming sequence, it is preferable to extract the defect features of the current printed layer and its upper and lower splicing layers in the CT scan image and output them.
[0035] Step 6: Based on the constructed dataset, a convolutional neural network is used to identify the defect types and defect distribution inside the SLM specimen.
[0036] More specifically, in this step, a convolutional neural network model is preferably used as the backbone network, and a multi-dimensional, multi-task cascaded neural network model, Md-Mt-CNN, is formed on this basis to identify the defect types, statistical characteristics, shape characteristics, and spatial distribution characteristics of the defects within the SLM sample. The basic principles and structural concepts of this multi-dimensional, multi-task cascaded neural network model, Md-Mt-CNN, are well known to those skilled in the art, and therefore will not be elaborated here.
[0037] According to another preferred embodiment of the present invention, see Figure 5 and Figure 6 For the multi-dimensional, multi-task cascaded neural network model Md-Mt-CNN, a three-dimensional input structure is preferred, and a shared feature layer is used to extract the input features of each dimension. Then, the output is decoupled and synchronized, so that defect classification and feature regression prediction can be achieved simultaneously in one model.
[0038] In summary, the SLM defect type and distribution monitoring method of the present invention performs synchronous acquisition by selecting three-dimensional signals—visible light band, laser reflection band, and near-infrared shortwave band—which, compared with conventional monitoring methods such as single visible light / near-infrared in the prior art, can quickly and in real-time acquire melt pool data during the forming process, more completely reflecting the changes in the formed part during the SLM forming process, and correspondingly improving the accuracy of defect monitoring. It can achieve accurate monitoring of defect types and defect distribution in different regions, and therefore has good practical value and application prospects.
[0039] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A multi-dimensional data fusion method for monitoring SLM defect types and distribution, characterized in that, The method includes the following steps: S1. Design experimental parameters and print SLM samples; S2. Synchronously acquire three-dimensional signals in the visible light band, laser reflection band, and near-infrared shortwave band during the printing process; S3. The three-dimensional signals are superimposed to obtain the corresponding monitoring images, namely, visible light monitoring image, laser reflection monitoring image and near-infrared shortwave monitoring image; S4. Use CT to scan the SLM sample, obtain the corresponding three-dimensional structure and cut it to remove the air phase. Then, divide the three-dimensional structure into multiple slice images along the Z-axis, which correspond to the thickness of the scanned layer. For each slice image, analyze the morphological characteristics and location distribution of unfused porosity defects and keyhole defects, complete the representation of the defect characteristics accordingly, and align the CT scan image with the monitoring image. S5. Take each of the monitored images as input and the defect features of the CT scan images as output to construct a dataset; S6. Based on the constructed dataset, a convolutional neural network is used to identify the defect types and defect distribution inside the SLM specimen.
2. The method as described in claim 1, characterized in that, In step S1, it is preferable to use the orthogonal experimental design method to design experimental parameters.
3. The method as described in claim 1 or 2, characterized in that, In step S2, the visible light band signal reflects the dynamic changes in the SLM additive manufacturing process, and the corresponding changes in light intensity are used to characterize the forming quality of the material surface; the laser reflection band signal reflects the direct emission and absorption rate of laser energy, and is used to characterize the reflected light properties of the material and the process energy utilization rate. The near-infrared short-wave band signal reflects the temperature of the molten pool on the material surface and is used to characterize the temperature change of the material.
4. The method as described in claim 3, characterized in that, In step S2, it is preferable to further include the operation of acquiring printing-related data, which includes laser start / stop signals, slicing laser scanning paths, etc.
5. The method according to any one of claims 1-4, characterized in that, In step S3, the monitoring image is preferably enhanced by spatial domain enhancement algorithm and frequency domain enhancement algorithm to improve image quality and reduce noise interference during the forming process.
6. The method according to any one of claims 1-5, characterized in that, In step S4, for the alignment of the CT scan image and the monitoring image, it is preferable to start from the top layer of the block and perform image registration according to the number of layers and thickness.
7. The method according to any one of claims 1-6, characterized in that, In step S5, preferably, the monitoring images corresponding to the current printed layer and its upper and lower splicing layers are combined into a three-channel image as input according to the forming sequence; correspondingly, the defect features of the current printed layer and its upper and lower splicing layers in the CT scan image are extracted and output according to the forming sequence, thereby reflecting the fluctuation between layers in the forming sequence.
8. The method according to any one of claims 1-7, characterized in that, In step S6, a convolutional neural network model is preferably used as the backbone network, and a multi-dimensional, multi-task cascaded neural network model Md-Mt-CNN is formed on this basis to identify the defect types, statistical characteristics, shape characteristics and spatial distribution characteristics of the defects inside the SLM sample.
9. The method as described in claim 8, characterized in that, For the multi-dimensional, multi-task cascaded neural network model Md-Mt-CNN, a three-dimensional input structure is preferred, and a general feature layer is used to extract the input features of each dimension. Then, through decoupling and synchronous output, defect classification and feature regression prediction can be achieved simultaneously in one model.
10. The method as described in claim 9, characterized in that, The signal acquisition frequency of the monitoring image is related to the defect classification and prediction accuracy. Higher defect classification accuracy can be achieved by increasing the signal acquisition frequency, while the size of the defect features that can be identified is smaller.