A method for online monitoring and quality control of surface defects in parts during laser powder bed melting process.

Abstract

This invention discloses a method for monitoring and controlling surface defects in parts during laser powder bed melting. The method includes the following steps: acquiring powder-coated images and laser-scanned images of the part surface layer by layer during the processing; performing image enhancement processing on the powder-coated images to obtain enhanced powder-coated images; inputting the enhanced powder-coated images into a trained Unet semantic segmentation model to segment surface defect regions and generate a mask image; quantifying the surface defects to obtain a surface defect quantification index Ap; inputting the laser-scanned images into a three-class four-layer convolutional neural network to determine the melting state of the part surface to obtain a signal M; inputting the surface defect quantification index Ap and the signal M into a proportional control system to calculate the required process parameters for the next layer, and adjusting the process parameters layer by layer until the final product is formed. This invention can effectively monitor and quantify surface defects in parts, improving part quality and production efficiency.

Description

Technical Field

[0001] This invention relates to the field of monitoring technology in metal additive manufacturing, specifically to a method for monitoring and quality control of surface defect lines in parts during the laser powder bed melting process. Background Technology

[0002] Laser Powder Bed Fusion (LPBF), a leading additive manufacturing technology, is widely used in the manufacturing industry. Based on CAD models, this technology uses lasers to melt and solidify metal powder materials layer by layer, constructing the solid structure of parts. Compared to traditional manufacturing processes, LPBF offers numerous advantages, including high design freedom, short product cycles, and high material utilization, and has been widely applied in fields such as medical implants and aerospace engine design. However, the LPBF forming process is accompanied by a series of defects, such as spheroidization, spatter, swelling, and surface roughness. These defects severely degrade the internal surface quality of the parts, adversely affecting powder distribution in subsequent layers and impairing interlayer fusion. In particular, collisions between the powder spreading carriage and the rough internal surface can damage parts and recoating mechanisms, resulting in significant waste of manpower and materials.

[0003] Online real-time in-situ monitoring can detect defects in a timely manner and issue early warnings on the severity of defects. Feedback control can dynamically adjust the process parameters of subsequent layers based on the monitored signals. This automated process requires no manual intervention, can monitor the printing status of parts in real time, and adjust the parameters of subsequent layers in a timely manner, effectively reducing material waste and improving the surface quality of parts, enhancing the intelligence level of LPBF equipment, and promoting the development of intelligent manufacturing.

[0004] In existing monitoring technologies, such as the 3D printing process fault diagnosis method and device with patent publication number CN109968671B, the method monitors the temperature of the current layer in real time and uses a machine learning model to determine whether there is a fault in the printing of the current layer during the fused deposition modeling process and provides fault type information. However, the metal additive manufacturing process is more complex, especially the LPBF process, which needs to be carried out in a closed chamber and under inert gas protection, making monitoring and control very difficult.

[0005] In addition, patent publication number CN116740001A describes an online monitoring method for laser selective melting defects. This method uses an industrial camera to acquire images after powder spreading and laser scanning, and combines the acquired images with process parameter information to achieve online defect classification. Another example is patent publication number CN117390433A, which describes an online monitoring method, device, equipment, and medium for laser powder bed melting defects. This method converts air-borne acoustic emission signals into a spectrum and characterizes defect information by calculating the ratio of signal frequency amplitude density in the low-frequency, mid-frequency, and high-frequency bands of the spectrum. Yet another example is patent publication number CN115582559A, which describes an online monitoring system and method for defects in powder-spread additive manufacturing. This system uses acoustic emission technology to achieve online monitoring of defects in the LPBF process, using acoustic signals to acquire surface defects of parts in real time. However, none of the above monitoring methods have yet implemented layer-by-layer feedback control based on the monitoring signals to mitigate the degree of defects. Summary of the Invention

[0006] To overcome the defects and shortcomings of existing technologies, this invention provides a method for monitoring and controlling surface defects in parts during the laser powder bed melting process. This invention uses an industrial camera to identify the melting state layer by layer and quantify the degree of defects. Based on the monitored signals, a feedback control method is used to adjust the process parameters of subsequent layers layer by layer, thereby improving part quality and production efficiency.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting, comprising the following steps:

[0009] During the processing, images of powder coating on the surface of the part and images after laser scanning are acquired layer by layer, and image enhancement processing is performed on the powder coating images on the surface of the part to obtain enhanced powder coating images on the surface of the part.

[0010] A Unet semantic segmentation model for surface defect segmentation of a part is constructed and trained. An enhanced powder-coated image of the part surface is input into the trained Unet semantic segmentation model to segment the surface defect region and generate a mask image.

[0011] Surface defects are quantified to obtain the surface defect quantification index Ap;

[0012] The image after laser scanning is input into a three-class four-layer convolutional neural network used to determine the melting state of the part surface to obtain signal M;

[0013] The surface defect quantification index Ap and signal M are input into the proportional control system to calculate the process parameters required for the next layer, and the process parameters are adjusted layer by layer until the final shape is formed.

[0014] As a preferred technical solution, the image enhancement processing employs contrast-limited histogram equalization.

[0015] As a preferred technical solution, the Unet semantic segmentation model includes a feature extraction part and an upsampling part;

[0016] The feature extraction part extracts features from the powder-coated image on the surface of the part through the convolution operation of the convolution module, and reduces the feature map size by max pooling downsampling;

[0017] The upsampling part uses a transposed convolution operation to upsample the feature map, thereby enlarging the size of the feature map.

[0018] The feature map obtained by max pooling downsampling is concatenated with the feature map obtained by upsampling using skip connections;

[0019] The feature map obtained from the last upsampling and the feature map obtained from the first feature extraction are concatenated by channels. The feature map is then processed by the convolution operation of the convolution module. The number of channels is adjusted by 1x1 convolution to obtain a single-channel feature for classification. The sigmoid function is then used to process the feature map to obtain a pixel-by-pixel classification result with the same size as the original input image. The final mask map is then generated to achieve semantic segmentation of the image.

[0020] As a preferred technical solution, the convolutional module includes two convolutional layers: a first convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1, and a second convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1. Each convolutional layer is followed by batch normalization and leaky ReLU activation. The input of the second convolutional layer is the output of the first convolutional layer, and the input of the first convolutional layer is the feature map obtained by max pooling downsampling the output of the previous convolutional module.

[0021] As a preferred technical solution, the feature extraction part extracts features from the powder-coated image on the surface of the part through the convolution operation of the convolution module. The feature extraction is performed on the powder-coated image on the surface of the part through multiple feature extraction operations, and the number of convolution kernels in the convolution module corresponding to each feature extraction operation is different.

[0022] As a preferred technical solution, the upsampling part uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to perform upsampling operation on the feature map. The number of transposed convolutions used in each upsampling operation is different. The new feature map obtained is then concatenated with the feature map of the same size obtained from the feature extraction part through skip connections to obtain a fused feature map.

[0023] The convolution module performs convolution operations on the fused feature maps, and the number of convolution modules used each time is different.

[0024] As a preferred technical solution, the laser-scanned image is input into a three-class, four-layer convolutional neural network used to determine the melting state of the part surface to obtain signal M, specifically including:

[0025] The image after laser scanning undergoes four feature extraction operations in sequence. Each feature extraction operation includes convolution, batch normalization, Leaky ReLU activation, and max pooling downsampling. The extracted features are then processed by global average pooling and three convolutional layers with a kernel size of 1×1. The probability value of the melting state is output through the softmax function.

[0026] As a preferred technical solution, surface defects are quantified to obtain a surface defect quantification index, which is specifically expressed as follows:

[0027]

[0028] Where R and C are the width and length of the original image, respectively, and II[·] is an indicator function used to determine whether the predicted probability of a pixel exceeds a predefined threshold α. If it exceeds the threshold α, then II(P(r,c)) = 1, otherwise it is 0.

[0029] As a preferred technical solution, the probability values ​​of the melting state are output through the softmax function, and the maximum probability value is selected as the final result to obtain the M value. The calculation formula is expressed as follows:

[0030]

[0031] Among them, P 欠熔化 P 正常 P 过熔化 These represent the probability values ​​for under-melting, normal melting, and over-melting, respectively.

[0032] As a preferred technical solution, the surface defect quantification index Ap and signal M are input into the proportional control system to calculate the process parameters required for the next layer, and the process parameters are adjusted layer by layer, specifically including:

[0033] P 下一层 =P 当前层 +M×Kp×Error-(1-M)Kp×Error

[0034] Among them, P 下一层 This is the laser power required for the next layer, P 当前层 is the laser power of the current layer, M is a Boolean value used to indicate the current melting state, that is, if the current melting state is over-melting, M=1, otherwise M=0, Kp is the scaling factor, and Error represents the error between the current monitoring signal Ap and the expected value.

[0035] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0036] (1) Compared with the limitations of existing vision-based online monitoring of laser powder bed defects, which can only classify defects but cannot quantify them layer by layer, this invention uses an industrial camera to collect images of the part surface after powdering during the processing, processes them, and directly inputs them into a semantic segmentation model for defect segmentation to obtain quantitative indicators of surface defects. It can automatically determine the melting state of the part based on the image after laser scanning and adaptively adjust the direction of the laser process parameters according to the melting state, which can effectively alleviate the surface defects of subsequent layers of parts, making the production process more intelligent and adaptive.

[0037] (2) Traditional online monitoring systems may require manual intervention to adjust process parameters. However, this invention utilizes online monitoring information and feedback control theory to realize the automatic adjustment of process parameters during LPBF, thereby reducing the need for manual intervention, improving production efficiency, and ensuring the stability of part quality. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the method for monitoring and controlling surface defect lines and quality control of parts during the laser powder bed melting process of the present invention.

[0039] Figure 2 This is a schematic diagram of the overall device of the present invention, which uses an industrial camera to capture images of the surface of a part after powder coating and laser scanning.

[0040] Figure 3 (a) is a schematic diagram of the original image of powder coating on the surface of the part;

[0041] Figure 3 (b) is a schematic diagram of the powder coating on the surface of the reinforced part;

[0042] Figure 3 (c) is a schematic diagram of the mask image after defect segmentation;

[0043] Figure 4 This is a schematic diagram of the architecture of the Unet semantic segmentation model for segmenting surface defects of parts according to the present invention;

[0044] Figure 5 (a) is a schematic diagram of the architecture of the convolution module of the Unet semantic segmentation model of the present invention;

[0045] Figure 5 (b) is a schematic diagram of the max pooling downsampling architecture of the Unet semantic segmentation model of the present invention;

[0046] Figure 5 (c) is a schematic diagram of the transposed convolutional upsampling architecture of the Unet semantic segmentation model of the present invention;

[0047] Figure 6 (a) is a schematic diagram of the undermelting state of the part surface after laser scanning according to the present invention;

[0048] Figure 6 (b) is a schematic diagram of the normal melting state of the part surface after laser scanning according to the present invention;

[0049] Figure 6 (c) is a schematic diagram of the overmelted state of the part surface after laser scanning according to the present invention;

[0050] Figure 7 This is a schematic diagram of the architecture of the four-layer convolutional neural network used to identify the melting state of a part according to the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] Example

[0053] like Figure 1 As shown, this embodiment provides a method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting, including the following steps:

[0054] S1: During the processing, images of powder coating on the surface of the part and images after laser scanning are acquired layer by layer, and image enhancement processing is performed on the powder coating images on the surface of the part to obtain enhanced powder coating images on the surface of the part.

[0055] like Figure 2 As shown, based on the part location information, industrial cameras are used to take images of the part surface layer by layer after powder coating and after laser scanning during the manufacturing process;

[0056] After applying powder, the original image lacks sufficient clarity in both the background and the target, such as... Figure 3 As shown in (a), to enhance the contrast between the background and the target while mitigating the effects of lighting, the image enhancement process employs Limiting Contrast Histogram Equalization (CLAHE). CLAHE effectively reduces image noise and better distinguishes defect areas from background areas, such as... Figure 3 As shown in (b).

[0057] S2: Input the enhanced powder-coated image of the part surface into the trained online surface defect monitoring neural network model to segment the surface defect region and generate a mask image, such as... Figure 3 As shown in (c);

[0058] In this embodiment, the neural network model for online surface defect monitoring adopts the Unet semantic segmentation model for part surface defect segmentation. This Unet semantic segmentation model includes a feature extraction part and an upsampling part, and uses skip connections to concatenate the feature maps obtained from feature extraction and upsampling. Specifically, the feature extraction part extracts features from the powder-coated image of the part surface through convolution operations, and uses max pooling downsampling to reduce the feature map size and increase the abstraction level of the features. The upsampling part uses transposed convolution operations to upsample the feature map, enlarging its size. Simultaneously, skip connections are used to concatenate the feature map obtained from max pooling downsampling with the feature map obtained from upsampling to retain more detailed information. Finally, the feature map obtained from the last upsampling and the feature map obtained from the first feature extraction are concatenated, the feature map is processed using a convolution module, and then a 1x1 convolution is used to adjust the number of channels to obtain a single-channel feature for classification. This feature is then processed using the sigmoid function to obtain a pixel-by-pixel classification result of the same size as the original input image, generating the final mask map and achieving semantic segmentation of the image.

[0059] The input to the feature extraction part is the original image of the part surface, and the output is a feature map with low resolution and high-level feature representation. The input to the upsampling part is the final feature map output from the feature extraction part, and the output is a feature map with higher resolution and more semantic meaning. The network structure is as follows: Figure 4 As shown, it has the following characteristics:

[0060] (1) The input to the online monitoring neural network model for surface defect online segmentation and quantization is an image of powder coating on the surface of the part. In order to adapt to the semantic segmentation network already trained in this embodiment, the input image needs to be grayscale processed and its size adjusted to 224×224;

[0061] (2) The feature extraction part employs convolutional modules and max pooling operations to extract image features and reduce the dimensionality of the feature map. Specifically, firstly, convolutional modules are used to extract image features, and then max pooling is used to retain important information and reduce the size of the feature map;

[0062] (3) The upsampling part employs transposed convolution upsampling and convolution modules to increase the resolution of features and improve the network's expressive power. A 2×2 kernel size and a stride of 2 transposed convolution operation are used. This enlarges the image by inserting two blank pixels between adjacent elements on the feature map and performing convolution operations, thereby increasing the size of the feature map. The feature map size after transposed convolution is:

[0063] H 输出 =(H 输入 -1) × step size + K - 2 × fill

[0064] Where K is the kernel size, the convolution module further processes the magnified feature map to extract richer and more complex features, thereby enhancing the network's expressive power. This combination enables the network to better capture detailed information in the image and improves the accuracy and performance of surface defect segmentation.

[0065] (4) Both the feature extraction and upsampling parts use the same type of convolutional module, such as Figure 5 As shown in (a), each convolutional module includes two convolutional layers: a first convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1, and a second convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1. Each convolutional layer is followed by batch normalization and leaky ReLU activation. The number of convolutional kernels used in these two convolutional layers is consistent. In this embodiment, the first convolutional layer is used to extract low-level features of the input, such as the edges and surface textures of defects, while the second convolutional layer is used to extract high-level features of the input, such as the shape of defects. The batch normalization layer is used to normalize the input of each channel to avoid some gradient vanishing or gradient exploding. Leaky ReLU activation is used to increase the non-linear expressive power of the neural network, which helps the network learn more complex functional relationships.

[0066] The upsampling part performs a transposed convolutional upsampling operation on the feature map, such as... Figure 5 As shown in (c), the resolution of the feature map is doubled using a 2×2 transposed convolution kernel with a stride of 2 and padding of 0. Then, the upsampled feature map is concatenated with the downsampled feature map from the feature extraction part through skip connections. The concatenated feature map is then used for convolution operations by a convolution module.

[0067] The second convolutional input of each convolutional module is the output of the first convolutional layer. Specifically, the input of the first convolutional module in the feature extraction part is the powder-coated image of the part surface; the features extracted by the first convolutional layer serve as the input of the second convolutional layer, and the input of subsequent convolutional modules are the dimensionality-reduced feature maps obtained by max pooling downsampling of the previous convolutional modules; the input of the upsampling part is the output of the last convolutional module in the feature extraction part, and the input of the convolutional modules in the upsampling part is a new feature map obtained by concatenating the channels of two feature maps, which are the upsampled feature map and the feature output by the convolutional modules in the feature extraction part, respectively.

[0068] In the feature extraction part, max pooling downsampling is used, such as Figure 5 As shown in (b), the parameters are a 2×2 convolutional kernel, a stride of 2, and padding of 0. Dimensionality reduction is performed on the feature map output by each convolutional module. Then, the dimensionality-reduced feature map is input into the next convolutional module.

[0069] (5) The process of feature extraction is as described in (2) and (4). The difference is that the number of convolution kernels in the convolution module used in each feature extraction operation is different. The number of convolution kernels in the convolution module used in the five feature extraction operations are 32, 64, 128, 256 and 512 respectively.

[0070] (6) The processing flow of the upsampling part is as described in (3). The upsampling part uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to perform upsampling operation on the feature map, which doubles the size of the feature map. The difference is that the number of transposed convolutions used in each upsampling operation is different. The number of transposed convolutions used in the upsampling order is 256, 128, 64, and 32 respectively.

[0071] In addition, the new feature map is concatenated with the feature map of the same size obtained from the feature extraction operation through skip connections to obtain a fused feature map.

[0072] The number of convolutional kernels in the convolutional modules used for the convolutional operation on the feature map after channel concatenation is different. The number of convolutional kernels in the convolutional modules used for the four convolutional operations are 256, 128, 64, and 32, respectively.

[0073] (7) Finally, the 32-channel features obtained from the first convolution module of the feature extraction part and the 32-channel features obtained from the last upsampling operation are concatenated. The convolution module containing 32 kernel functions is used to operate on the concatenated feature map. The number of channels is adjusted by 1×1 convolution to obtain a single-channel feature for classification. Then, this feature is processed by the sigmoid function to obtain a pixel-by-pixel classification result with the same size as the original input image, and a mask map is generated.

[0074] S3: Quantify surface defects to obtain the monitoring signal Ap, which is the surface defect quantification index;

[0075] After completing surface defect segmentation, the degree of surface defects needs to be quantified and used as a monitoring signal input to the control system. The number of all segmented pixels recorded in the semantic segmentation mask image is counted, and the proportion of defect points to the entire image is calculated to obtain the surface defect quantification index Ap, which is achieved through the following formula:

[0076]

[0077] Where R and C are the width and length of the original image, respectively, and II[·] is an indicator function used to determine which pixels have a predicted probability that exceeds a predefined threshold α. If the probability exceeds the threshold α, then II(P(r,c)) = 1, otherwise it is 0.

[0078] S4: Input the laser-scanned image into a three-class, four-layer convolutional neural network used to determine the melting state of the part surface to obtain signal M; such as Figure 6 (a) Figure 6 (b) and Figure 6 As shown in (c), the three different melting states after laser scanning are under-melting, normal melting, and over-melting; this three-class four-layer convolutional neural network has the following characteristics:

[0079] (1) The image used for surface melting state identification is the image after laser scanning. In order to adapt to the convolutional neural network that has been trained in this embodiment, the input image after laser scanning must be a grayscale image, and the size is not limited.

[0080] (2) The surface melting state recognition in this embodiment uses a four-layer convolutional neural network. Its input is the image captured after laser scanning, and it goes through four feature extraction operations. The input of the first feature extraction operation is the image captured after laser scanning, and the inputs of the remaining feature extraction operations are the outputs of the previous feature extraction operations. Each feature extraction operation includes convolution, batch normalization, Leaky ReLU activation, and max pooling. The parameters of the convolution operation are: a convolution with a kernel size of 3×3, a stride of 1, and padding of 1; a max pooling operation with a sampling window size of 2×2 and a stride of 2; the number of convolution kernels used in the four feature extraction operations are 32, 64, 128, and 256, respectively. Finally, the extracted features are processed by global average pooling and three convolutional layers with a kernel size of 1×1, and the probability value of each category (normal melting, over-melting, and under-melting) is output through the softmax function.

[0081] like Figure 7 As shown, each convolutional layer has a kernel size of 3×3, a stride of 1, and padding of 1. Convolution, batch normalization, Leaky ReLU activation, and max pooling downsampling with a kernel size of 2×2, a stride of 2, and padding of 0 are performed sequentially. The number of kernels used in these four convolutional layers are 32, 64, 128, and 256, respectively. Global average pooling is performed on the extracted features, and then three 1×1 convolutional layers are used to adjust the number of channels. The probability value of the melting state is calculated using the SoftMax function, and the highest probability value is selected as the final result to obtain the M value. The calculation formula is as follows:

[0082]

[0083] S5: Input the monitoring signals Ap and M into the proportional control system to calculate the process parameters required for the next layer, and adjust the process parameters layer by layer.

[0084] In this embodiment, for the layer-by-layer monitoring and control of part surface defects in LPBF, the controller adopts proportional control. The signal Ap monitored for each layer is used to determine whether the area of the surface without defects has reached the preset value, and proportional control is performed based on the error between it and the preset value; the monitored signal M is used to determine whether the part surface has melted. Part surface defects are usually caused by over-melting or under-melting, and the control directions for over-melting and under-melting are different. Therefore, it is possible to determine whether to increase or decrease the laser power for the subsequent layer based on M. During the processing, first, it is judged whether the surface of the metal part meets the expected requirements. If it meets the requirements, that is, when Ap > E (expected value), the process parameters of the subsequent layer are kept unchanged. If Ap < E, the power of the next layer is updated according to the following formula:

[0085] P 下一层 = P 当前层 + M × Kp × Error - (1 - M)Kp × Error

[0086] Where, P 下一层 is the laser power required for the next layer, P 当前层 is the laser power of the current layer, M is a boolean value used to represent the current melting state, that is, if the current melting state is over-melting, M = 1, otherwise M = 0, Kp is the proportionality coefficient, which can be adjusted according to the material and the LPBF equipment itself. Taking the Di-Metal 300 equipment produced by Guangzhou Leijia Additive Manufacturing Company and using 316L as the processing material as an example, Kp can be set to 180, and Error represents the error between the currently monitored signal Ap and the expected value.

[0087] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting process, characterized in that, Includes the following steps: During the processing, images of powder coating on the surface of the part and images after laser scanning are acquired layer by layer, and image enhancement processing is performed on the powder coating images on the surface of the part to obtain enhanced powder coating images on the surface of the part. A Unet semantic segmentation model for surface defect segmentation of a part is constructed and trained. An enhanced powder-coated image of the part surface is input into the trained Unet semantic segmentation model to segment the surface defect region and generate a mask image. Surface defects are quantified to obtain the surface defect quantification index Ap, which is specifically expressed as: Where R and C are the width and length of the original image, respectively; II[·] is an indicator function used to determine whether the predicted probability of a pixel exceeds a predefined threshold α. If it exceeds the threshold α, then II[(P(r,c)≥α]=1, otherwise it is 0; The laser-scanned image is input into a three-class, four-layer convolutional neural network used to determine the melting state of the part's surface, obtaining signal M, which specifically includes: The image after laser scanning undergoes four feature extraction operations in sequence. Each feature extraction operation includes convolution, batch normalization, Leaky ReLU activation, and max pooling downsampling. The extracted features are then processed by global average pooling and three convolutional layers with a kernel size of 1×1. The probability value of the melting state is output through the softmax function. The softmax function outputs the probability values ​​of the molten state, and the highest probability value is selected as the final result to obtain the M value. The calculation formula is as follows: Among them, P 欠熔化 P 正常 P 过熔化 These represent the probability values ​​for under-melting, normal melting, and over-melting, respectively. The surface defect quantification index Ap and signal M are input into the proportional control system to calculate the process parameters required for the next layer, and the process parameters are adjusted layer by layer until the final shape is formed. The surface defect quantification index Ap and signal M are input into the proportional control system to calculate the required process parameters for the next layer, and the process parameters are adjusted layer by layer, specifically including: P 下一层 =P 当前层 +M×Kp×Error-(1-M)Kp×Error Among them, P 下一层 This is the laser power required for the next layer, P 当前层 is the laser power of the current layer, M is a Boolean value used to indicate the current melting state, i.e., the current melting state is over-melting, M=1, otherwise M=0, Kp is the proportional coefficient, which is adjusted according to the material and the LPBF equipment itself, and Error represents the error between the current monitoring signal Ap and the expected value.

2. The method for monitoring and quality control of surface defect lines in parts during laser powder bed melting process according to claim 1, characterized in that, The image enhancement process employs contrast-limited histogram equalization.

3. The method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting process according to claim 1, characterized in that, The Unet semantic segmentation model includes a feature extraction part and an upsampling part; The feature extraction part extracts features from the powder-coated image on the surface of the part through the convolution operation of the convolution module, and reduces the feature map size by max pooling downsampling; The upsampling part uses a transposed convolution operation to upsample the feature map, thereby enlarging the size of the feature map. The feature map obtained by max pooling downsampling is concatenated with the feature map obtained by upsampling using skip connections; The feature map obtained from the last upsampling and the feature map obtained from the first feature extraction are concatenated by channels. The feature map is then processed by the convolution operation of the convolution module. The number of channels is adjusted by 1x1 convolution to obtain a single-channel feature for classification. The sigmoid function is then used to process the feature map to obtain a pixel-by-pixel classification result with the same size as the original input image. The final mask map is then generated to achieve semantic segmentation of the image.

4. The method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting process according to claim 3, characterized in that, The convolutional module includes two convolutional layers: a first convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1, and a second convolutional layer with a kernel size of 3×3, a stride of 1, and padding of 1. Each convolutional layer is followed by batch normalization and leaky ReLU activation. The input of the second convolutional layer is the output of the first convolutional layer, and the input of the first convolutional layer is the feature map obtained by max pooling downsampling the output of the previous convolutional module.

5. The method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting process according to claim 3, characterized in that, The feature extraction part extracts features from the powder-coated image on the surface of the part through the convolution operation of the convolution module. The feature extraction is performed through multiple feature extraction operations, and the number of convolution kernels in the convolution module is different for each feature extraction operation.

6. The method for monitoring and controlling surface defect lines and quality of parts during laser powder bed melting process according to claim 3, characterized in that, The upsampling part uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to upsample the feature map. The number of transposed convolutions used in each upsampling operation is different. The new feature map is then concatenated with the feature map of the same size obtained from the feature extraction part through skip connections to obtain a fused feature map. The convolution module performs convolution operations on the fused feature maps, and the number of convolution modules used each time is different.

Images