A method, apparatus, device, and medium for real-time monitoring of cracks in an additive manufacturing process

Abstract

The present application relates to a kind of additive manufacturing process crack real-time monitoring method, device, equipment and medium, comprising: S1.in the additive manufacturing process, the laser radiation signal for forming piece is collected in real time and conversion, and the converted electric radiation characteristic signal is handled, and the radiation signal thermogram of forming piece is obtained;S2.to the forming piece simultaneously carries out CT scanning and obtains image data, and after processing, crack characteristic distribution chart is obtained;S3.radiation signal thermogram and crack characteristic distribution chart correspond with the layer number of forming piece, and the radiation signal crack characteristic data set corresponding to layer by layer point by point is obtained;S4.data set is used to train machine learning or deep learning model, and crack determination model is obtained;S5.crack determination model is used for online monitoring to carry out crack identification.The method of the present application can form the radiation signal crack database corresponding to layer by layer point by point in three-dimensional space, and the data consistency and migratability are improved.

Description

Technical Field

[0001] This invention belongs to the field of additive manufacturing technology, specifically relating to a method, device, equipment, and medium for real-time monitoring of cracks in the additive manufacturing process. Background Technology

[0002] Laser additive manufacturing, such as laser powder bed fusion (L-PBF) and laser direct energy deposition (L-DED), offers significant advantages in the integrated and rapid prototyping of complex parts, making it crucial for high-end manufacturing fields like aerospace. However, the high solidification-cooling rate and large temperature gradient of laser additive manufacturing processes, coupled with the repeated thermal cycles of rapid heating and cooling during the forming process, easily lead to the accumulation of significant residual stress within the formed part. This can induce crack formation, reduce the mechanical properties of the part, and even cause printing failures, posing a critical challenge for the large-scale industrial production of laser additive manufactured parts. Cracks formed during additive manufacturing are typically on the micrometer to millimeter scale, making real-time monitoring difficult. Analysis usually relies on offline characterization methods (such as scanning electron microscopy). Offline characterization methods typically require destructive testing of the part, resulting in low efficiency, high cost, and difficulty in providing comprehensive analysis of all locations for each part in a batch. Non-destructive testing methods, such as computed tomography (CT), can only penetrate centimeter-sized metal samples and cannot perform detailed three-dimensional scanning of large-sized parts. The existing technical problems are as follows: existing online monitoring schemes usually only focus on sensor acquisition and model training, lacking (1) processing rules for generating a uniform resolution, locatable two-dimensional radiation signal heat map from the discrete and time-series data obtained from the original monitoring; and (2) a process for strictly corresponding the online monitoring data with the physical characteristics of the formed parts (such as crack distribution data obtained from CT scans) in terms of geometric position and layer number, and establishing a layer-by-layer point-by-point mapping database, resulting in unstable data correspondence and poor model transferability. Therefore, developing an online crack monitoring method for additive manufacturing processes, capturing crack formation in real time, and accurately locating crack positions is a prerequisite for realizing rapid online control of laser scanning process parameters, improving the quality and manufacturing efficiency of formed parts, and is also an important foundation for realizing digital / intelligent additive manufacturing. Summary of the Invention

[0003] To overcome the aforementioned problems in the existing technology, this invention provides a method, device, equipment, and medium for real-time crack monitoring in the additive manufacturing process. It collects radiation signals from the molten pool online, forms a corresponding database, and enables real-time and accurate prediction of internal cracks in the formed parts during the additive manufacturing process.

[0004] A method for real-time crack monitoring in an additive manufacturing process, the method comprising: S1. Real-time acquisition of laser radiation signals for forming parts during the additive manufacturing process and conversion into electrical radiation characteristic signals, and processing of the electrical radiation characteristic signals to obtain a radiation signal thermogram of the forming parts; S2. Simultaneously perform CT scanning on the formed part to obtain image data, and process the image data to obtain a crack feature distribution map; S3. Correspond the radiation signal thermogram and crack feature distribution map to the number of layers of the formed part to obtain a radiation signal crack feature dataset corresponding to each layer and each point. S4. Train a machine learning or deep learning model using a dataset to obtain a crack detection model; S5. Use a crack detection model for online monitoring and crack identification.

[0005] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the acquisition in S1 is achieved using a coaxial radiation monitoring system. The system includes a laser emitter, a lens, a beam splitter, a molten pool monitoring module, an XY scanning unit, and a powder bed. The laser emitter is connected to the lens, and the lens and the molten pool monitoring module are simultaneously connected to the beam splitter. The beam splitter is connected to the XY scanning unit, which is positioned above the powder bed. The molten pool monitoring module is equipped with a photodiode, which converts the acquired laser radiation signal for forming the part into an electrical radiation characteristic signal.

[0006] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the formed part is a metal formed part, and the metal material used includes nickel-based superalloys, titanium alloys, aluminum alloys, steel or copper alloys.

[0007] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the electroradiation characteristic signal is processed in S1 to obtain a radiation signal thermogram of the formed part, specifically including: S11. Store the monitoring results of each layer of the formed part for the monitoring points in the form of offline data; S12. The offline data is segmented according to the absolute coordinate range of the formed part read in the visualization interface to obtain segmented data; S13. Filter, transform, rasterize and interpolate the segmented data to obtain the radiation signal heat map.

[0008] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the image data processing in S2 includes: angle correction, position correction, threshold binarization extraction, connected component denoising, and cropping and color inversion.

[0009] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S3 specifically includes: S31. Using the special geometric features on the formed part as the reference layer for CT scanning, and the junction of different scanning sections of the formed part as the reference surface, a set of corresponding radiation signal thermograms and crack distribution maps are obtained. S32. Based on the spacing between adjacent CT slices and the thickness of the additive manufacturing powder layer, establish a layer-by-layer correspondence from the reference layer upwards and downwards, determine the crack feature distribution map corresponding to each forming layer, and the radiation signal-crack feature dataset corresponding to each layer and point of the formed part.

[0010] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the special geometric features include bosses, through holes, or slots.

[0011] The present invention also provides a real-time crack monitoring device for additive manufacturing process. The device is used to implement the method, including: a first establishment module, used to collect laser radiation signals for forming parts in real time during additive manufacturing process and convert them into electrical radiation characteristic signals, and process the electrical radiation characteristic signals to obtain a radiation signal thermogram of the forming parts. The second module is used to simultaneously perform CT scanning on the molded part to obtain image data, and to process the image data to obtain a crack feature distribution map. The third module is used to correlate the radiation signal thermogram and crack feature distribution map with the number of layers of the formed part, so as to obtain the radiation signal-crack feature dataset corresponding to each layer and each point of the formed part. The training module is used to train a machine learning or deep learning model for crack binary classification using the dataset, so as to obtain a crack classification model. The online monitoring module is used to identify cracks online using a crack detection model.

[0012] The present invention also provides a computer storage medium storing a computer program, the computer program being executed by a processor to implement the method described.

[0013] The present invention also provides an electronic device, the electronic device comprising: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the method.

[0014] Beneficial effects of the present invention The present invention discloses a method, apparatus, equipment, and medium for real-time crack monitoring in additive manufacturing processes, comprising: segmenting, filtering, coordinate transformation, and interpolating rasterization of online monitoring offline data to obtain a radiation signal heatmap for each layer; performing angle / position correction, threshold extraction, connected component denoising, cropping, and color inversion on CT layer-by-layer scan images to obtain a crack feature distribution map corresponding to each point of the radiation signal heatmap; establishing a Z-axis layer-by-layer correspondence by designing reference geometric features and determining reference surfaces on the formed part, and combining the spacing between adjacent CT slices and the powder layer thickness to form a layer-by-layer-point corresponding dataset; on this basis, any applicable machine learning and / or deep learning discrimination model can be trained using this dataset, and the trained model can be used to achieve online automated processing, judgment, and visualization output through an online monitoring script.

[0015] Compared with existing technologies, the advantages of this invention include: ① achieving a unified conversion from radiation time-series signals to spatial radiation signal heatmaps, enabling monitoring data to be used for spatial positioning and machine learning modeling; ② achieving pixel-level point-to-point correspondence between crack labels and radiation signal heatmaps through CT crack feature extraction and geometric correction; ③ establishing a Z-axis layer-by-layer aligned radiation signal crack database with layer-by-layer point-to-point correspondence in the three-dimensional space of the formed part through reference features / reference surfaces, improving data consistency and transferability; ④ the online monitoring script can automatically complete data processing, crack judgment, and result mapping display, supporting post-processing filtering of isolated misjudged areas, improving visualization and positioning reliability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the online radiation signal monitoring system in this invention; Figure 2 The special shape of the forming part used to acquire training data in this invention, as well as the shape and size of its lower cross section; Figure 3 This is a schematic diagram of the deep learning model architecture in this invention; Figure 4 This is a flowchart of the online crack monitoring script in this invention; Figure 5 This is a schematic diagram illustrating the results of the online monitoring script calling a deep learning model to identify and predict six different sets of input data in this invention. Figure 6 This is a schematic diagram of the cross-sectional shape of the molded part used in Example 2. Figure 7 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To better understand the technical solution of this invention, the content of this invention includes, but is not limited to, the specific embodiments described below. Similar technologies and methods should be considered within the scope of protection of this invention. To make the technical problems to be solved, the technical solutions, and advantages of this invention clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0018] It should be understood that the embodiments described in this invention are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0019] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0020] like Figure 7 As shown, the present invention provides a method for real-time monitoring of cracks in the additive manufacturing process. The method includes: S1. Real-time acquisition of laser radiation signals for forming parts during the additive manufacturing process and conversion into electrical radiation characteristic signals, and processing of the electrical radiation characteristic signals to obtain a radiation signal thermogram of the forming parts; S2. Simultaneously perform CT scanning on the formed part to obtain image data, and process the image data to obtain a crack feature distribution map; S3. Correspond the radiation signal heat map and crack feature distribution map to the number of layers of the formed part to obtain the radiation signal crack feature dataset and mapping relationship database corresponding to each layer and point of the formed part, and divide the database into training set and validation set; S4. Use the training set to train a machine learning or deep learning model for crack binary classification to obtain a crack determination model; S5. Use the validation set for online monitoring to identify cracks.

[0021] The design principle of this invention is as follows: According to Planck's blackbody radiation law, the radiation signal of a material is related to its temperature. Therefore, collecting the radiation signal of the molten pool during additive manufacturing can reflect the characteristics of molten pool temperature changes. Based on the correlation between solidification principles and cracking behavior in additive manufacturing, the generation of cracks in additively manufactured parts is directly related to the thermal history of the molten pool: when the molten pool temperature is too low, the part is prone to developing unfused holes, whose rough inner surfaces and irregular shapes lead to local stress concentration, inducing crack formation; when the molten pool temperature is too high, keyhole defects can form, and during rapid solidification, the instability at the tail of the keyhole also easily leads to crack formation. In addition, an increased temperature gradient within the molten pool can lead to increased thermal stress. Under multiple cycles, thermal stress gradually accumulates until the part cracks. The above mechanisms indicate that the thermal history of the molten pool is an important inducing factor for crack formation in additively manufactured parts, and the determination of crack formation through analysis of molten pool radiation signals has a theoretical basis.

[0022] Specifically, the process of this invention is as follows: 1. Data Acquisition: This invention employs a coaxial radiation monitoring system, such as... Figure 1 As shown, the system includes a laser emitter 1, an f-theta lens 2, a beam splitter 3, a molten pool monitoring module 4, an XY scanning unit 5 composed of galvanometers, and a powder bed 6. The laser emitter 1 and the f-theta lens 2 are connected, and the f-theta lens 2 and the molten pool monitoring module 4 are connected to the beam splitter 3. The XY scanning unit 5 composed of galvanometers is positioned above the powder bed 6. After a high-power laser beam is emitted from the laser emitter 1, it passes through the f-theta lens 2 and then through the beam splitter 3, which has high transmittance (>95%) for laser wavelength. The laser beam is then focused onto the processing surface of the powder bed 6 by the XY scanning unit 5. The broadband radiation light (including visible and near-infrared light) generated by the molten pool returns along the original optical path and reaches the beam splitter 3 through the XY scanning unit 5. The beam splitter 3 is designed to have high reflectivity (>90%) for the main wavelength range of the molten pool radiation, such as 700-1100 nm, thus reflecting this portion of the light to a branch optical path molten pool monitoring module 4 at a 90-degree angle to the main optical path. After entering the molten pool monitoring module 4, the branch optical path is optimized and filtered by the focusing lens and narrowband filter set in the monitoring module to optimize the light intensity and select a specific wavelength of light signal. At the focal point of the branch optical path, two photodiodes 7 are set to collect the light signal and convert it into an electrical signal to obtain the characteristic signal of the molten pool radiation.

[0023] 2. The laser additive manufacturing equipment used includes, but is not limited to, laser powder bed melting (L-PBF) equipment or laser directional energy deposition (L-DED) equipment. The equipment is used to form metal parts, wherein the metal materials include, but are not limited to, nickel-based superalloys, titanium alloys, aluminum alloys, steel, copper alloys, etc. These materials are commonly used alloys in metal additive manufacturing. In the following embodiments, the present invention uses nickel-based superalloys.

[0024] 3. Data Acquisition and Storage: The coaxial radiation monitoring system is used to collect real-time molten pool radiation signals from the laser and powder bed interaction points of all formed parts in each forming layer during the forming process. The monitoring results for each layer are stored as offline data. The offline data includes at least the absolute coordinates of the monitoring points, the radiation signal value, and the laser's on / off status. Since the laser may be off during the complete scanning path, the two photodiodes still record the radiation signal at that time, thus recording information such as the laser's on / off status. The offline data represents all data for the current layer of all formed parts in the current forming task. That is, if there is only one formed part in the forming task, the offline data for each layer is all the monitoring data for that current layer of that formed part. If there are multiple formed parts, then it is the monitoring data for the current layer of all formed parts. In this case, a data segmentation step is required later.

[0025] 4. Online monitoring data processing (radiation thermogram generation): The following operation is performed on the offline monitoring data to process the monitoring data of all formed parts in each layer into a radiation signal thermogram for each formed part in each layer: (1) Data segmentation: Based on the absolute coordinate range of each molded part on the molding substrate, the offline data is segmented by the molded part and stored separately. The offline data is all the data of all molded parts in the current layer in the current molding task. That is, if there is only one molded part in the molding task, then the offline data of each layer is all the monitoring data of the molded part in that layer. In this case, this step is not required. If there are multiple molded parts, then this part of the operation is to save the monitoring data of each molded part separately. (2) Data filtering: Remove data points in the data segmented in step (1) where the laser switch state is off (Power on is 0 or its equivalent state) to obtain monitoring point data for each layer of each formed part that only contains the laser light emission state; (3) Coordinate transformation: Establish a relative coordinate system with the center of the smallest outer rectangle of the forming section of the forming part as the origin, calculate the translation parameters, and translate the absolute coordinates of each monitoring point of each layer of each forming part obtained in step (3) to the relative coordinate system, transform the absolute coordinates of each monitoring point into the relative coordinates of each monitoring point, so as to obtain the coordinate transformation data of each layer of each forming part. (4) Rasterization and interpolation: The coordinate transformation data obtained in step (3) is mapped to the relative coordinate system established in step (3) according to its relative coordinates. Each relative coordinate corresponds to the radiation signal monitoring value obtained by photodiode monitoring. The range of radiation signal values ​​is statistically analyzed as a legend. An interpolation method is used to generate a radiation signal heat map with the same resolution as the forming section of the formed part and the same resolution as the CT scan. The monitoring data of all layers of all formed parts are batch plotted to obtain radiation signal heat maps of all layers of all formed parts. These heat maps contain the relative coordinates and radiation signal monitoring values ​​of the monitoring points of all lasers in the corresponding layer of the formed part under the on state.

[0026] 5. CT Crack Feature Extraction and Alignment (Crack Feature Distribution Map Generation): The XY cross-sectional image obtained by CT scanning layer by layer along the Z-axis of the formed part is processed. The XY plane represents information such as the forming surface and cross-section of the formed part during the forming process, and the Z-axis represents the forming direction from bottom to top. The layer interval of the CT scan along the Z-axis here is different from that mentioned in step 4 above. The layer interval of the CT scan along the Z-axis is the CT scan resolution. The layer mentioned in step 4 above represents each forming layer in the forming process, and the layer interval is the powder layer thickness. The following steps are performed on the obtained XY cross-sectional image to obtain the crack feature distribution map corresponding to each point of the radiation thermogram: (1) Angle correction: In order to eliminate the field deviation caused by manual loading of the molded parts before CT scanning, an example CT scan image is taken for each molded part, and the deviation angle α between the cross section and the horizontal plane of the molded part in the example scan image is calculated. All CT scan images of each molded part are rotated by (180°-α) to obtain the CT scan images of all molded parts after angle correction. (2) Position correction: In order to eliminate the position deviation caused by manual loading of the formed part before CT scanning, the position correction is performed on the CT image after angle correction. For this purpose, the theoretical cross-sectional contour is used as the target position, the formed cross-section in the CT result is manually aligned with the cross-sectional contour, and the corresponding displacement L is recorded. The entire set of angle-corrected CT images is then translated in batches to obtain CT images after angle and position correction. (3) Threshold / Binarization Extraction: The corrected CT image is binarized using an optimized threshold (at which all crack pixels are kept as 1 and background pixels are set to 0 as much as possible). Pixels with gray values ​​below the threshold are set to white and pixels with gray values ​​above the threshold are set to black, thereby realizing the binarization extraction of crack features and obtaining a preliminary crack feature distribution map. (4) Connected component denoising: In order to remove these noise points, the area of ​​all black connected components in these preliminary crack feature distribution maps is counted and the noise areas smaller than the threshold are removed. Only crack features with an area greater than the set threshold are retained to obtain the denoised crack feature distribution map. (5) Cropping and Color Inversion: The denoised crack feature distribution map is cropped along the minimum bounding rectangle of the theoretical contour, and the foreground / background is inverted so that the crack pixels are 1 and the background is 0. At this point, the image size and radiation signal thermal properties are obtained. Figure 1 The crack feature distribution map is obtained. Since the resolution of the radiation signal thermogram and the CT scan is the same, the crack feature distribution map and the radiation thermogram correspond point by point on the XY plane.

[0027] 6. Establishment of Layer-by-Layer Correspondence and Database Construction in the Z-Axis: After the above processing, a radiation signal thermogram and crack feature distribution map with point-by-point correspondence can be obtained on the XY plane. However, due to the different interlayer intervals of CT scans and forming tasks, it is impossible to find the crack feature distribution map obtained from the CT scan corresponding to each layer of radiation signal thermogram. It is necessary to further establish the interlayer correspondence between the two types of maps in the Z-axis direction. By designing special geometric features on the formed part, such as bosses, through holes, and slots, these features are obtained by additional forming on the top after the forming task of the formed part is completed. These features are used as the CT scan reference layer, and the intersection of different scanning sections of the formed part is used as the reference plane to obtain a set of corresponding radiation signal thermograms and crack distribution maps. Combining the relationship between the interlayer spacing between adjacent CT scan images (this spacing is the CT scan resolution) and the thickness of the additive manufacturing powder layer, a hierarchical correspondence between the radiation signal thermogram and the CT scan image can be established. The specific method is as follows: First, align the two types of data at the same physical height using a CT scan reference layer. Then, determine the layer number correspondence between the two based on the height conversion relationship TX*△CT=tx*△H. Here, TX is the CT scan layer number, i.e., the layer number corresponding to each CT scan, with a layer spacing △CT equal to the CT scan resolution; tx is the radiation signal thermal image layer number, i.e., the layer number corresponding to each forming layer, with a layer spacing △H equal to the powder layer thickness of that layer. After determining the reference layer correspondence, set the actual physical height of the reference plane to 0. Using the reference layer as a reference, the actual height of the CT scan image of layer Tx downwards is TX*△CT, while the actual height of the radiation signal thermal image of layer tx is tx*△H. When the two heights are equal, i.e., TX*△CT=tx*△H, the correspondence can be determined. For example, for the radiation thermal image of layer 10, its actual height is 10*△H. The CT scan layer number TX corresponding to that layer can be calculated using the above formula, thus determining the CT scan image corresponding to the radiation signal thermal image of that layer. Using the same method, the correspondence between radiation signal thermograms and CT scan images can be established layer by layer, thereby realizing layer-by-layer mapping based on the reference layer, extending upwards and downwards.

[0028] Determine the crack feature distribution map corresponding to each forming layer (i.e., powder layer), and construct a radiation signal-crack feature dataset corresponding to each layer and each point of the formed part. That is, after completing the above operations, all monitoring points on each forming layer of each formed part can be correlated with the crack feature distribution. Figure 1 One-to-one correspondence.

[0029] 7. Crack Detection Model: Based on the above dataset, it is divided into training and validation sets in an 8:2 ratio. The training set is used for learning model parameters during training, while the validation set is used for evaluating model performance and fine-tuning parameters during training. This section uses the training and validation sets for a machine learning or deep learning model for crack binary classification. Taking a deep learning model as an example, its architecture is as follows... Figure 3As shown, the model consists of three modules used to calculate the "crack / no crack" probability for local regions: The temporal fusion module (3D Conv + temporal pooling) is composed of multiple 3D convolutions connected in series. Each 3D convolution contains Conv3D, BatchNorm3D, and ReLU activation functions, used to extract local spatial and cross-layer features from the radiation intensity heatmap input data 8. The number of convolution stacks (conv_stacks) in this module is set to 1, 2, or 3 layers, aiming to gradually expand the model's receptive field by stacking small convolution kernels layer by layer, thereby improving its ability to express cross-layer dimensional features. The initial input size of the model's input unit 9 is [B, 1, L, 20, 20], where B is the batch size: the number of data inputs to the model in batches; L is the number of consecutive frames, used for stacking local regions of multiple radiation signal heatmaps centered on the target layer in the dataset, with cracks as the label; 20×20 is the resolution. After stacking 10 3D convolutions, the output size becomes [B, 32, L, 20, 20]. Subsequently, temporal average pooling is performed on the inter-layer dimensions using module 11, resulting in the output size [B, 32, 20, 20]. The spatial feature extraction module (2DResNet+SE attention) 12 employs a ResNet structure with residual connections, combined with the Squeeze-and-Excitation (SE) attention mechanism, to enhance the model's responsiveness to key spatial regions. This module comprises a stack of three residual modules (configured as [2, 2, 2]). The second and third stages achieve spatial downsampling through stride=2, reducing the input feature resolution from [20×20] to [10×10] and [5×5], respectively, resulting in feature data of size [B, 128, 5, 5]. The global feature pooling and classification module 13 uses an adaptive global average pooling mechanism to spatially compress the input data [B, 128, 5, 5] to [B, 128, 1, 1], then flattens it and connects it to two fully connected layers (Linear→ReLU→Dropout→Linear) 14, mapping it to an output 15 of [B, 1], representing the predicted probability that the corresponding region is a crack. The following rule is used in the above process: local regions of the radiation signal heatmaps of L consecutive layers before and after the target layer in the training set are stacked, with cracks used as labels. The training set data is used as input to train the model according to the above rules, and the performance of the trained model is evaluated and the parameters are tuned using the validation set during the training process. This results in a deep learning model that predicts cracks based on the input data according to the above rules, which serves as the crack judgment model.

[0030] 8. Online monitoring scripts: such as Figure 4 As shown, an online monitoring script is written to apply the trained crack detection model to the actual online forming process. After forming begins, a photodiode collects the radiation signal from the molten pool and saves the data offline to the device's local path. The monitoring script scans the saved local path at preset time intervals to verify whether the current data quantity is a multiple of L. Since the input data is a continuous L-layer radiation signal heatmap, it only meets the input requirements when it is a multiple of L. If it is not a multiple of L, no processing is performed, and the process waits for the condition to be met; if it is a multiple of L, the data processing in step 4 is automatically executed, and a model input unit is constructed. The model trained in step 7 is called to perform crack detection, and an identification output indicating whether a crack exists is obtained. After the input in the set of radiation signal heatmaps is identified, the model output is mapped to the corresponding position of each input unit in the radiation signal heatmap with different colors, obtaining the model's visual identification result of the radiation signal heatmap. Figure 5 As shown, the red pixel area represents the area identified as a crack by the model, and the gray pixel area represents the area identified as a crack-free area by the model. At the same time, isolated region filtering can be further performed on the units identified as cracks. Specifically, if the unit in the region is identified as a crack, its 8 neighboring regions are further checked. If all 8 neighboring regions are crack-free, then the unit in this region is a misjudgment, and the unit is corrected to be crack-free, thereby eliminating isolated misjudged regions and improving the reliability of the results.

[0031] This invention segments, filters, transforms, and interpolates offline monitoring data to obtain a radiation signal heatmap for each layer. It performs angle / position correction, threshold extraction, connected component denoising, cropping, and color inversion on CT layer-by-layer scan images to obtain a crack feature distribution map corresponding to each point in the radiation signal heatmap. By designing reference geometric features and determining reference surfaces on the formed part, and combining the spacing between adjacent CT slices and the powder layer thickness, a Z-axis layer-by-layer correspondence is established, forming a layer-by-layer, point-by-point corresponding dataset and database. Based on this, any applicable machine learning / deep learning discrimination model can be trained, and automated processing, judgment, and visualization output can be achieved through online monitoring scripts.

[0032] As an embodiment of the present invention, the present invention also discloses a real-time crack monitoring device for additive manufacturing process. The device is used to implement the method, including: a first establishment module, used to collect laser radiation signals for forming parts in real time during additive manufacturing process and convert them into electrical radiation characteristic signals, and process the electrical radiation characteristic signals to obtain a radiation signal thermogram of the forming parts. The second module is used to simultaneously perform CT scanning on the molded part to obtain image data, and to process the image data to obtain a crack feature distribution map. The third module is used to correspond the radiation signal heat map and crack feature distribution map with the number of layers of the formed part, so as to obtain the radiation signal crack feature dataset and mapping relationship database corresponding to each layer and point of the formed part, and divide the database into training set and validation set. The training module is used to train a machine learning or deep learning model for crack binary classification using the training set, so as to obtain a crack determination model. The online monitoring module is used to identify cracks online using a crack detection model.

[0033] As an embodiment of the present invention, the present invention also discloses a computer storage medium storing a computer program, the computer program being executed by a processor to implement the method described.

[0034] As an embodiment of the present invention, the present invention also discloses an electronic device, the electronic device comprising: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the method.

[0035] The following specific examples will be used for illustration.

[0036] Example 1: This embodiment takes the L-PBF forming of the easily crackable nickel-based superalloy CM247LC as an example to illustrate the complete implementation process of "online monitoring data post-processing - CT crack characterization extraction - layer-by-layer point-by-point correspondence library construction", and provides an optional online application method for crack identification.

[0037] (1) Online acquisition of radiation signals: The L-PBF equipment equipped with the coaxial radiation monitoring system described above is used to form the easily cracked nickel-based high-temperature alloy CM247LC; the radiation time sequence signal of the molten pool is collected synchronously with each layer scan, and the absolute coordinates (X,Y) of the monitoring point on the current forming surface, the layer number k of the current forming layer, and the original data such as the laser switch mark Power_on are recorded, with k being a natural number.

[0038] (2) Monitoring data preprocessing and radiation signal heat map generation: The original monitoring data collected by photodiodes is sorted by the part number (multiple parts are sorted in the manner of 1, 2, 3...) and the layer number k. The valid data points with Power_on=1 in the original monitoring data are filtered. The minimum bounding rectangle of the effective scanning point set of the monitoring points in the actual light output state of the laser in the kth layer is calculated and its center is taken as the reference origin. The absolute coordinates (X,Y) are translated to the relative coordinate system of the part. Finally, all monitoring points of each layer are transformed to the relative coordinate system and the radiation distribution heat map is drawn using the interpolation method. At the same time, the one-to-one correspondence between pixels and actual spatial coordinates is saved.

[0039] (3) CT crack distribution map extraction: CT scans of the same formed part are performed to obtain XY cross-sectional image data; angle correction (rotation angle calculated based on the cross-sectional chord and rotated as a whole) and position correction (aligning with the theoretical contour and recording displacement parameters) are performed on the image data; threshold segmentation is performed on the corrected CT slices to obtain the crack binary map, and noise is removed by connecting region area thresholding; according to radiation heat... Figure 1 The ROIs are cropped and the resolution is unified to obtain the crack distribution map (crack pixels = 1, non-crack pixels = 0).

[0040] (4) Layer-by-layer alignment and sample library construction: Select the reference plane at the junction of the top boss of the molded part and the main body of the molded part, such as Figure 2 As shown, a pair of corresponding CT slice layers and forming layers are identified; combining the CT slice spacing (Δz-CT) and the powder layer thickness (Δz-layer), a mapping relationship is established between the CT scan results and the radiation signal layer numbers, achieving a layer-by-layer correspondence between the radiation signal thermogram and the crack distribution map; based on this, a database of "radiation signal thermogram-crack distribution map" sample pairs for each layer is established. In this embodiment, 220 sample pairs are obtained from 5 formed parts, and the training set and validation set are divided in an 8:2 ratio.

[0041] (5) Crack identification model training: Divide the radiation signal heatmap of each layer into 5×20×20 input units according to the preset window, and input... Figure 3 The deep learning model with the shown architecture was trained. Focal loss was used as the loss function during training. Based on previous validation work and existing literature, the parameters α and γ in the Focal loss were set to 0.6 and 2, respectively, with an initial learning rate of 5 × 10⁻⁶. -4 The weight decay coefficient is 1×10 -5 The training run consisted of 50 epochs with a batch size of 256. The trained deep learning model achieved a 97.8% accuracy rate in identifying cracks in the CM247LC molded part in this example on the validation set.

[0042] (6) Online monitoring and visualization: To simulate the online monitoring conditions in the actual forming process, the verification data is input into the monitoring catalog at equal time intervals, and the online monitoring script is called to identify cracks. The time for the online monitoring script to call the model to process and predict the input data of each layer is about 8-9 seconds. In the L-PBF forming process, the time required from the completion of the current layer scanning to the start of the next layer is about 10 seconds (for a 280 mm substrate, the interval is longer for larger substrates). This model can identify cracks in the current layer before the next layer is formed, guiding operators to take timely crack suppression measures, such as optimizing process parameters or pausing printing, thereby effectively reducing printing losses and improving the yield.

[0043] Example 2: This embodiment is used to verify the applicability of the "monitoring data heat map generation - CT crack map extraction - layer-by-layer point-by-point registration and library construction" process described in this invention to different geometric shapes of formed parts and different process parameter conditions.

[0044] (1) Forming objects and parameters: Select two different cross-sectional shapes (trapezoidal interface and arched interface) for forming parts, such as Figure 6 As shown, different forming parameters such as laser power, scanning speed, scanning spacing, layer thickness and interlayer rotation angle are set compared to Example 1.

[0045] (2) Generation of radiation signal heatmaps: For the original radiation data of the two types of molded parts during the forming process, the grouping, Power_on filtering, coordinate translation based on the center of the minimum bounding rectangle and grid interpolation are performed according to step (2) of Example 1 to obtain the radiation signal heatmaps of each layer. Since this coordinate normalization strategy takes the geometric center of the scanning point of each layer as a reference, it can reduce the coordinate differences caused by different cross-sectional shapes, thereby ensuring the comparability of the heatmaps in space.

[0046] (3) CT crack distribution map and registration library: CT scans were performed on the two types of molded parts and angle / position correction, threshold segmentation, connected domain noise reduction, clipping and resolution unification were completed according to step (3) of Example 1; then, layer number mapping was established based on the reference plane according to step (4) of Example 1, and the sample database of "radiation signal thermogram-crack distribution map" under the corresponding geometry of the molded parts was established.

[0047] (4) Online monitoring and identification application: In this embodiment, the crack identification model trained in Embodiment 1 can be directly called to perform online prediction of the data in this embodiment; the identification results of the crack monitoring deep learning model show that the crack identification accuracy rate for trapezoidal cross-section forming parts is 98.3%; the crack identification accuracy rate for arched cross-section forming parts is 98.6%. The time for the online monitoring script to call the model to process and predict the input data of each layer is about 8-9 seconds.

[0048] Example 3: This embodiment is used to verify the applicability of the data processing and registration library construction process described in this invention under different material systems, and to explain the layer number mapping setting method when Z-axis parameters such as layer thickness change.

[0049] (1) Molding objects and parameters: Two different grades of materials were selected for L-PBF molding (IN738LC and IN939), and different process parameters were set from those in Example 1: the laser power of IN738LC molding parts was 170 W, the scanning speed was 2000 mm / s, the scanning spacing was 40 μm, the layer thickness was 40 μm, and the interlayer rotation angle was 67°; the laser power of IN939 molding parts was 100 W, the scanning speed was 1000 mm / s, the scanning spacing was 80 μm, the layer thickness was 30 μm, and the interlayer rotation angle was 67°.

[0050] (2) Generation of radiation signal heatmaps: For the raw radiation data collected during the forming process of different materials, Power_on filtering, coordinate translation and grid interpolation are performed according to step (2) of Example 1 to obtain the radiation signal heatmaps of each layer. This processing flow depends on the spatial coordinates and radiation intensity of the monitoring data, and does not depend on the material grade, so it can be directly reused.

[0051] (3) Extraction of CT crack distribution map: CT scans are performed on the molded parts made of different materials. Angle / position correction, threshold segmentation, connected domain denoising and ROI clipping are performed according to step (3) of Example 1 to obtain the crack distribution map of each slice, and the same resolution and coordinate reference are maintained with the radiation signal thermogram.

[0052] (4) Z-direction layer-by-layer correspondence: In this embodiment, the powder layer thickness Δz-layer of IN738LC is different from that in embodiment 1. After determining the reference corresponding layer through the reference plane, the layer number mapping relationship is recalculated according to Δz-CT / Δz-layer to achieve layer-by-layer alignment of the radiation signal thermogram and the CT crack distribution map, and the formed part sample is matched with the database accordingly.

[0053] (5) Online monitoring and identification application: In this embodiment, the crack identification model trained in Embodiment 1 is directly called to perform online prediction of the data in this embodiment; the identification results of the crack monitoring deep learning model show that the crack identification accuracy rate for IN738LC formed parts is 99.2%, and the crack identification accuracy rate for IN939 formed parts is 94.9%. The time for the online monitoring script to call the model to process and predict the input data of each layer is approximately 8-9 seconds.

[0054] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for real-time crack monitoring in an additive manufacturing process, characterized in that, The method includes: S1. Real-time acquisition of laser radiation signals for forming parts during additive manufacturing and conversion into electrical radiation characteristic signals, and processing of electrical radiation characteristic signals to obtain a radiation signal thermogram of the forming parts; S2. Simultaneously perform CT scanning on the formed part to obtain image data, and process the image data to obtain a crack feature distribution map; S3. Correspond the radiation signal thermogram and crack feature distribution map to the number of layers of the formed part to obtain a radiation signal crack feature dataset corresponding to each layer and each point. S4. Train a machine learning or deep learning model using a dataset to obtain a crack detection model; S5. Use the crack detection model to monitor online and identify cracks.

2. The method according to claim 1, characterized in that, The acquisition in S1 is achieved using a coaxial radiation monitoring system, which includes a laser emitter, a lens, a beam splitter, a molten pool monitoring module, an XY scanning unit, and a powder bed. The laser emitter is connected to the lens, and the lens and the molten pool monitoring module are both connected to the beam splitter. The beam splitter is connected to the XY scanning unit, which is positioned above the powder bed. The molten pool monitoring module is equipped with a photodiode, which converts the acquired laser radiation signal for forming the part into an electrical radiation characteristic signal.

3. The method according to claim 1, characterized in that, The formed part is a metal formed part, and the metal materials used include nickel-based high-temperature alloys, titanium alloys, aluminum alloys, steel or copper alloys.

4. The method according to claim 1, characterized in that, In S1, the characteristic signal of electric radiation is processed to obtain the radiation signal thermogram of the formed part, specifically including: S11. Store the monitoring results of each layer of the formed part for the monitoring points in the form of offline data; S12. The offline data is segmented according to the absolute coordinate range of the formed part read in the visualization interface to obtain segmented data; S13. Filter, transform, rasterize and interpolate the segmented data to obtain the radiation signal heat map.

5. The method according to claim 1, characterized in that, The image data processing in S2 includes: angle correction, position correction, threshold binarization extraction, connected component denoising, cropping, and color inversion.

6. The method according to claim 1, characterized in that, S3 specifically includes: S31. Using the special geometric features on the formed part as the reference layer for CT scanning, and the junction of different scanning sections of the formed part as the reference surface, a set of corresponding radiation signal thermograms and crack distribution maps are obtained. S32. Based on the spacing between adjacent CT slices and the thickness of the additive manufacturing powder layer, establish a layer-by-layer correspondence from the reference layer upwards and downwards, determine the crack feature distribution map corresponding to each forming layer, and the radiation signal-crack feature dataset corresponding to each layer and point of the formed part.

7. The method according to claim 6, characterized in that, The special geometric features include bosses, through holes, or slots.

8. A real-time crack monitoring device for additive manufacturing processes, characterized in that, The device is used to implement the method according to any one of claims 1-7, comprising: a first establishment module, used to acquire laser radiation signals for forming parts in real time during additive manufacturing and convert them into electrical radiation characteristic signals, and to process the electrical radiation characteristic signals to obtain a radiation signal thermogram of the forming parts; The second module is used to simultaneously perform CT scanning on the molded part to obtain image data, and to process the image data to obtain a crack feature distribution map. The third module is used to correlate the radiation signal thermogram and crack feature distribution map with the number of layers of the formed part, so as to obtain the radiation signal crack feature dataset corresponding to each layer and point of the formed part. The training module is used to train machine learning or deep learning models using datasets to obtain crack detection models. The online monitoring module is used to identify cracks online using a crack detection model.

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

10. An electronic device, characterized in that, The electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the method of any one of claims 1-7.

Images