Laser melting deposition porosity online monitoring method based on deep transfer learning

Abstract

The invention discloses a laser melting deposition porosity online monitoring method based on deep transfer learning. The laser melting deposition porosity online monitoring method comprises the steps that 1, a source domain sample set and a target domain sample set are obtained; 2, respectively extracting pore data of the molten pool images in the source domain sample set and the target sample set, and carrying out label classification on pores according to pore areas; step 3, establishing an online monitoring model of the porosity of the source domain by using ShuffleNetV2 with packet convolution and channel shuffling; 4, by means of the source domain porosity online monitoring model, constructing a target domain porosity online monitoring model based on deep sub-domain adaptive network transfer learning; according to the method, the porosity characteristics are extracted by means of the lightweight characteristic of the ShuffleNetV2 network and the channel shuffling technology, the porosity monitoring model of the source domain is established, the porosity monitoring of the target domain is realized through the depth sub-domain adaptation method, and the porosity monitoring of different LMD technological processes is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser additive manufacturing, and in particular relates to an online monitoring method for the porosity of laser melting deposition based on deep transfer learning. Background Art

[0002] Laser melting deposition (LMD) is a process in which a laser beam acts on metal powder delivered by a powder sprayer, causing it to melt and deposit along a predetermined route. After solidification, a layer of deposited metal is formed. This reciprocating process eventually completes the manufacturing of the geometric shape of the part. It has a faster deposition rate and higher manufacturing freedom, and has advantages in the rapid manufacture of large parts with good mechanical properties.

[0003] High quality and repeatability of parts have always been important challenges for laser metal additive manufacturing technology to face high-end application scenarios. In LMD manufacturing, the melting and solidification process of materials may introduce various defects, such as cracks, spheroidization, delamination, porosity and poor melting. These defects will reduce the mechanical properties and quality of parts and hinder their use in critical physical occasions. Among them, porosity is generally considered to be one of the important defects that affect the fatigue performance and crack propagation characteristics of parts. In order to reduce the porosity in products, an online porosity monitoring system is very necessary. Combining sensor data with machine learning algorithms to achieve porosity monitoring of printed parts has received widespread attention. However, machine learning algorithms rely on sufficient labeled data and the assumption of independent and identical distribution of data, but in laser melting deposition, changes in process parameters will have a significant impact on the distribution and correlation of experimental process data. It is very time-consuming and expensive to collect enough labeled data from scratch for a new process, and it is difficult to conduct experiments covering all process conditions to build a sample library. Summary of the invention

[0004] The purpose of the present invention is to provide an online monitoring method for the porosity of laser melting deposition based on deep transfer learning. The method first uses the lightweight characteristics of the ShuffleNetV2 network and the channel shuffling technology to extract pore characteristics, establishes a porosity monitoring model in the source domain, and then realizes the porosity monitoring of the target domain through a deep sub-domain adaptation method; it can effectively utilize the historical knowledge of relevant experimental data to realize the porosity monitoring of different LMD process.

[0005] To achieve the above object, the present invention has the following beneficial effects:

[0006] A method for online monitoring porosity of laser melting deposition based on deep transfer learning comprises the following steps:

[0007] Step 1: Obtain source domain sample sets and target domain sample sets;

[0008] The source domain sample set refers to a melt pool image dataset during the historical deposition process;

[0009] The target domain sample set refers to a molten pool image data set collected during the deposition process of new process parameters;

[0010] Step 2: Extract the pore data of the melt pool images in the source domain sample set and the target sample set respectively, and classify the pores according to the pore area;

[0011] Step 3: Establish a source domain porosity online monitoring model using ShuffleNetV2 with grouped convolution and channel shuffling, and use the source domain sample set and its label classification to train the source domain porosity online monitoring model;

[0012] Step 4: With the help of the source domain porosity online monitoring model, the target domain porosity online monitoring model is constructed based on deep sub-domain adaptive network transfer learning;

[0013] The trained source domain porosity online monitoring model is used as the initial target domain porosity online monitoring model. The source domain sample set and the target domain sample set are respectively extracted from the source domain porosity online monitoring model and the initial target domain porosity online monitoring model. Then, similar samples of the source domain pore characteristics and the target domain pore characteristics are projected into different sub-domains, and the distribution alignment is performed by calculating the LMMD distance between the sub-domains; the differences between label classifications are highlighted in the process of transfer learning.

[0014] Furthermore, the melt pool image dataset is preprocessed before step 2, wherein the preprocessing refers to determining the melt pool image size and cropping it according to the maximum pixel value and the connected domain, scaling it to a fixed size through a bilinear interpolation algorithm, and finally performing grayscale normalization.

[0015] Furthermore, in step 2, X-rays are used to scan the print along the deposition direction to extract a tomographic image, and then the pore data in the tomographic image is extracted through an image processing algorithm. The pore data includes shape information, position and area, and the pores are labeled and classified according to the pore area.

[0016] Furthermore, the loss function Loss of the porosity online monitoring model is:

[0017] Loss = L cross_entropy_target +λ 1 ×L cross_entropy_source +λ 2 ×L LMMD ;

[0018] Among them, L cross_entropy_target is the cross entropy loss function of the target domain, L cross_entropy_sourceis the cross entropy loss function of the source domain; 1 is the loss ratio parameter of the source domain; L LMMD is the local maximum difference loss function, λ 2 is the domain adaptation loss weight; Among them, i represents the current training cycle, and nepoch represents the total training cycle.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] (1) In view of the structural characteristics and dynamic correlation of the samples, ShuffleNetv2 is used as the basic network to extract the features of the melt pools of different layers through grouped convolution, and channel shuffling is used to enhance the information exchange between layers, while maintaining low computational and memory requirements.

[0021] (2) The porosity monitoring method in the LAM process based on deep sub-domain adaptation realizes knowledge transfer by aligning the sub-domains of different porosity levels in the source domain and the target domain, providing an effective solution for porosity monitoring in areas with scarce data and complex process conditions, and significantly reducing the cost and time required for monitoring.

[0022] (3) An experimental dataset was constructed to train the proposed model and verify the method, which was compared with other transfer learning methods and no-transfer models. The experimental results demonstrated the effectiveness and superiority of the proposed method.

[0023] (4) By learning cross-domain features or conversion mechanisms, the model can be adjusted according to the characteristics of the domain, and has better adaptability to the complex and changeable process of LMD. Moreover, adaptation within the subdomain is more targeted and efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the structure of the porosity online monitoring model of the present invention.

[0025] Figure 2 Flow chart of image processing in label classification; (a) actual image of the specimen; (b) melt pool image preprocessing; (c) CT image pore extraction process.

[0026] Figure 3 The pore distribution in the source domain and the target domain; (a) pore area distribution in the source domain, (b) pore area distribution in the target domain, and (c) pore category threshold division.

[0027] Figure 4 Results of porosity monitoring performance of different base networks on the source domain.

[0028] Figure 5This is the training process curve of ShuffleNetV2 under the optimal hyperparameters

[0029] Figure 6 Classification confusion matrix of different methods on the target domain; (a) no transfer, (b) fine-tuning, (c) freezing, (d) DAN, and (e) DSAS.

[0030] Figure 7 TSNE visualization of models trained for different target domain sample sets; (a) 10%, (b) 20%, (c) 40%, (d) 60%, (e) 80%, (f) 100%.

[0031] Figure 8 This is a graph showing the impact of increased data volume on accuracy and training time. DETAILED DESCRIPTION

[0032] This embodiment provides a method for online monitoring porosity of laser melting deposition based on deep transfer learning, comprising the following steps:

[0033] Step 1: Obtain source domain sample sets and target domain sample sets;

[0034] The source domain sample set refers to a molten pool image data set in a historical deposition process; the historical deposition process refers to a deposition process with different process parameters or different deposition equipment.

[0035] The target domain sample set refers to a molten pool image data set collected during the deposition process of new process parameters; the new process parameter deposition process refers to a new process parameter or a new deposition equipment deposition process.

[0036] This embodiment can effectively utilize the historical knowledge of relevant experimental data to realize porosity monitoring of different LMD process; in the case of data scarcity, it significantly improves the model development efficiency and reduces the development cost, which is of great significance to the implementation and promotion of LAM quality monitoring.

[0037] In this embodiment, an industrial camera is coaxially arranged on the laser melting deposition print head to collect the molten pool image of the laser melting deposition (LMD) process; specifically, the CF8 / 5MX industrial camera produced by Kappa optronics is used; during the deposition process, the laser emits a 1030nm high-energy laser beam, which is reflected to the substrate through a beam splitter covered with a 1030nm high-reflectivity transparent film, and the interaction between the laser and the material is completed at the substrate to form an interaction area centered on the molten pool. At the same time, the light emitted from the molten pool passes through the beam splitter, and the light beam is reflected to the coaxially arranged industrial camera through the reflector, so that the industrial camera can collect the image of the interaction area.

[0038] Preprocess the molten pool images in the source domain sample set and the target domain sample set. Determine the approximate position of the molten pool through the maximum pixel value and connected components, create a ROI cropping window of a fixed size, and include the molten pool in the window. To match the input size of the model and ensure the size consistency of the input images, scale the size of the cropped images to 224×224 through the bilinear interpolation algorithm. Then perform gray normalization on the images so that the gray values of each process image are within the range of [0, 255].

[0039] Step 2: Extract the pore data in the molten pool images and classify the pores according to the pore data;

[0040] Cut the printed specimen, and use X-ray to scan the specimen along the stacking direction to obtain the tomographic scan images of the horizontal section. Then use image processing algorithms to extract the pores in the tomographic scan images. Calibrate the position and actual area of the pores through the shape information of the pores and the CT pixel resolution. According to the size of the pore area, set an appropriate pore threshold to divide the pores into small pores, medium pores, and large pores.

[0041] Comparing the source domain sample set and the target domain sample set, they have great similarity in the feature space. The same pore threshold division ensures the same label space. Due to the technological differences between the source domain sample set and the target domain sample set, the printed parts have different characteristics in aspects such as interlayer bonding, heat transfer, and defect formation. The marginal probability distributions of the two are different, and the distribution states of the pores and the conditional distributions of the labels are also different.

[0042] Step 3: Establish an online monitoring model for the source domain porosity using grouped convolution and channel shuffle ShuffleNetV2;

[0043] ShuffleNetV2 is a lightweight and highly efficient deep learning model. It follows the residual connection idea of ResNet and adopts depthwise separable convolution and channel shuffle mechanisms in the residual blocks to achieve the highest possible classification or detection accuracy with as few computational resources and memory occupancy as possible.

[0044] Grouped convolution and channel shuffling are very suitable for the structural characteristics of the LAM process samples. First, the samples contain multiple layers of dynamic melt pool images, and grouped convolution can divide the melt pool images of different layers into multiple groups for feature extraction. Secondly, considering the dynamics and heat transfer of the melt pool, the melt pools of different layers interact and influence each other. Grouped convolution will cause the features of the melt pools of each layer to be independent of each other, making it difficult to extract global feature information. The channel shuffling mechanism will shuffle and rearrange the grouped feature maps. Channel shuffling can strengthen the information exchange between the melt pool images of different layers, which helps the model learn richer and more accurate feature representations. Finally, the samples of the LAM process contain multiple layers of dynamic melt pool images. The lightweight characteristics of ShuffleNetV2 enable it to process such high-dimensional samples efficiently and quickly, reducing the demand for computing resources.

[0045] The source domain sample set and its pore data are used to train the source domain porosity online monitoring model.

[0046] Step 4: Construct a target domain porosity online monitoring model (DSAS) based on deep sub-domain adaptive network transfer learning; realize real-time monitoring of the pores in the LAM process through the target domain porosity online monitoring model.

[0047] like Figure 1 As shown in the figure, the trained source domain porosity online monitoring model is used as the initial target domain porosity online monitoring model, and the source domain porosity online monitoring model and the target domain porosity online monitoring model are used to extract the source domain pore characteristics and the target domain pore characteristics respectively. The similar samples in the source domain pore characteristics and the target domain pore characteristics are projected into different sub-domains through the deep sub-domain adaptive network, and the distribution alignment is performed by calculating the LMMD distance between the sub-domains; the differences between categories are highlighted in the process of transfer learning.

[0048] The LMMD distance is used to align at different layers to achieve the effect of sub-domain alignment. The LMMD distance is:

[0049]

[0050] Where LMMD is the distribution distance, c is the label category, c∈{0,1,2}, C is the number of label categories, the label categories are small pores, medium pores, and large pores, so the number of label categories is 3. is the weight of the i-th sample in the source domain sample set belonging to category c, is the i-th sample in the source domain sample set; is the weight of the jth sample in the target domain sample set belonging to category c, is the jth sample in the target domain sample set.

[0051] and The calculation method is the same, specifically, the weight of the sample belonging to c is the weight of the i-th sample of category c in the domain, is the i-th category value of category c. The label vector is converted into one-hot encoding, that is, only 1 is used to represent the current category, and 0 is used to represent the other categories; this indicates that samples of the same category have the same weight value.

[0052] The source domain sample set and the target domain sample set have corresponding label classifications. The DSAS method proposed in this embodiment uses a supervised method to calculate the LMMD distance to reduce the error caused by the pseudo-labels on the target domain. During back propagation, the LMMD loss is not used alone as the loss function of the network, but together with the cross entropy loss function, it constitutes the final total loss. The total loss function of the network is calculated as follows:

[0053] Loss = L cross_entropy_target +λ 1 ×L cross_entropy_source +λ 2 ×L LMMD ;

[0054] Among them, L cross_entropy_target is the cross entropy loss function of the target domain, L cross_entropy_source is the cross entropy loss function of the source domain. Considering the difference in data distribution between the source domain and the target domain, we use λ 1 Parameter to control the loss ratio of the source domain. LMMD is the local maximum difference loss, also known as domain adaptation loss, and the parameter λ 2 is the domain adaptation loss weight that changes with the training cycle,

[0055]

[0056] Where i represents the current training cycle and nepoch represents the total training cycle. As i increases, λ 2 It gradually increases from almost 0. In the early stage of training, the model may need to pay more attention to classification performance. At the same time, since the model parameters have not converged to a better value, the difference between the source domain and the target domain cannot be distinguished well. Therefore, the weight of the domain adaptation loss is set small. As the performance improves, the model has been able to learn the characteristics of the source domain well, but it is not good at learning inter-domain features. Therefore, in order to improve the performance of the model in the target domain learning task, the proportion of domain loss in the total loss is gradually increased, and then the model parameters are brought closer to the target domain through back propagation.

[0057] Experimental verification of the effectiveness of this embodiment

[0058] The experiment was constructed using 30CrNi2MoVA powder and Q235 steel substrate. Multiple thin-walled structure samples were prepared using laser power and scanning speed as process parameters. The molten pool images collected during the deposition process with fixed process parameters and the molten pool images with different process parameters between different samples were used as the source domain sample set. In order to characterize a more complex process, the molten pool images collected during the deposition process with fixed laser scanning speed and changing laser power were used as the target domain sample set. The source domain sample set was prepared using the process parameters in Table 1, and the target domain sample set was prepared using the process parameters in Table 2. The powder was first placed in a closed space at 120°C for about 120 minutes to completely remove the moisture in and between the powders, and argon was filled to make the oxygen concentration reach less than 0.02%. The powder feeding rate was set to 1.5rpm (about 15g / min), and the carrier gas speed was 6L / min. The laser beam diameter was set to 3mm during printing, and the lifting height of each layer was 0.3mm. The length of all thin-walled parts was set to 40mm, and 30 layers were printed in a single pass. The camera collects melt pool images with a pixel size of 920 × 488 in real time at a frame rate of 50 Hz. 1434 samples are collected in the source domain and 491 samples are collected in the target domain.

[0059] Table 1. Source domain sample set printing experiment parameters

[0060]

[0061] Table 2 Target domain sample set printing experiment parameters

[0062]

[0063]

[0064] After printing, the printed thin-walled parts are cut. After printing, the nano Voxel4000 industrial CT system is used to detect all thin-walled parts through X-ray rays, and the horizontal cross-section tomographic images are obtained along the stacking direction. The image segmentation algorithm is used to locate and quantitatively analyze the micron-level pores in the thin-walled parts. Figure 2 As shown in Figure 4, the regional quality of thin-walled parts is evaluated based on their spatial distribution of pores.

[0065] The label distribution of the source domain sample set and the target domain sample set is statistically analyzed. Figure 3 (a) and (b) show the distribution of pore area of ​​samples in the source domain sample set and the target domain sample set, respectively, and (c) shows the threshold division of pores. The distribution of the two is significantly different. The number of the three types of pores in the source domain sample set is relatively evenly distributed, the number of small pores in the target domain sample set accounts for the largest proportion, and the number of large pores is relatively small. In addition, there is a sample imbalance problem in the target domain sample set, which further increases the difficulty of porosity monitoring in the target domain.

[0066] (1) AlexNet, VGG19, ResNet50, DenseNet121, and ShuffleNetV2 are used as source domain pore online monitoring models. They are the most representative CNN models in the field of deep learning and have excellent performance and representation. The source domain sample set is divided into training set and test set in a ratio of 7:3. All models are trained and tested on four NVIDIA GeforceRTX4090 graphics cards and use the same hyperparameters. The classification accuracy, floating point operation amount, parameter amount, and model inference speed of each model are shown in Figure 2. Figure 4 As shown in the figure, it can be seen that the deeper the network, the better the performance. The deepest network DenseNet121 has the worst effect. The porosity monitoring model based on ShuffleNetV2 has achieved the highest accuracy. This is because the network characteristics of ShuffleNetV2 are more compatible with the high dimensionality and grouping of the samples in this article. In addition, due to its lightweight characteristics, ShuffleNetV2 has the lowest FLOPs and parameters, which reduces the computing resources and memory usage, and can be applied to terminal deployment in more industrial scenarios.

[0067] This example uses ShuffleNetV2 as the basic model for source domain porosity monitoring. In order to further optimize the performance of the model, adjustments are made from aspects such as learning rate, batch size, and optimizer weight decay parameters, and a grid search method is used in a small range to find the optimal hyperparameter combination. When the initial learning rate is 0.001, batch size is 32, and the weight decay parameter of the SGD optimizer is 5e-3, the performance of the model is optimal with an accuracy of 90.5%. The loss value of the model on the training set and the accuracy change on the test set are shown in the following figure. Figure 5 shown.

[0068] (2) To prove the superiority of the DSAS model, non-transfer and other popular transfer learning methods were used to train the model. To ensure fairness, the hardware configuration, environment, and hyperparameters were kept the same. Non-transfer means directly training from scratch using ShuffleNetV2 without using the source domain-related models and parameters on the target domain. Fine-tuning training is to load the pre-trained parameters on the source domain as the initial parameters and use the data on the target domain to train and update the parameters. Freeze training means loading the pre-trained parameters on the source domain as the initial parameters, but freezing the underlying parameters of the model. The frozen parameters remain fixed during training, and only the unfrozen network layers are adjusted. Different stages of ShuffleNetV2 were frozen respectively, and finally, the optimal result was obtained when the first two stages were frozen. DAN is a transfer learning network for domain adaptation on a global scale. To demonstrate the superiority of the DSAS model under fair conditions, the baseline network of DAN was set the same as that of DSAS, and the same source domain pre-trained weights were used. The specific performance indicators of porosity monitoring of different models on the target domain are shown in Table 3, and the classification confusion matrix of each model is as Figure 6 .

[0069] Table 3 Porosity Monitoring Performance of Different Methods on the Target Domain

[0070]

[0071] As can be seen from Table 3, when transfer learning is not used, the accuracy obtained by training is the lowest because the amount of data in the target domain is small and the model is not trained sufficiently. The accuracies of several models trained using transfer learning methods are 3.8%, 5.4%, 3.1%, and 6.2% higher than those of the non-transfer model respectively. This shows that transfer learning is effective for porosity monitoring under variable process parameters. Although there are significant differences between the manufacturing processes under fixed process parameters and variable process parameters, they are similar in the overall pattern and are not two completely unrelated processes. Borrowing the historical knowledge of the model under fixed process parameters can lead to better results in porosity monitoring on variable process parameter data.

[0072] Compared with several transfer learning methods, freezing and fine-tuning are the most commonly used transfer learning methods, and they also have quite good results in porosity monitoring with variable process parameters. The accuracy of the frozen model in the target domain reached 88.2%, and since some model parameters do not participate in the training in the frozen training, the entire training of the model takes less time. The batch time in the table represents the average training time of each batch. The accuracy of the fine-tuned model further reached 89.8%. Fine-tuning training can adaptively adjust the model's weights, biases and other parameters according to the data characteristics of the target domain, so that the model can better adapt to new data. Both DAN and DSAS methods belong to deep adaptation networks. The DAN method performs distributed global adaptation calculations between the source domain and the target domain. DSAS no longer performs global adaptation, but adapts the relevant sub-domains separately. The distance between domains is also measured by the local maximum mean difference. In terms of training time, the deep adaptation method requires longer training time than freezing and fine-tuning. In terms of monitoring accuracy, the accuracy of the DAN method is the lowest among several transfer learning methods, only 87.5%, while the DSAS method achieves the best transfer effect, with a porosity classification accuracy of 90.6% for the target domain.

[0073] (3) In order to study the impact of the amount of data in the source domain and the target domain on the model performance, the model was trained using data of different proportions. First, the pre-training model was constructed using the entire source domain sample set, and the model was trained using 10%, 20%, 40%, 60%, 80% and 100% of the target domain sample set. Increasing the amount of target domain data can improve the monitoring accuracy. When the amount of data is small, the model has poor recognition effect on the three categories of pores, but when the amount of data exceeds 60%, the distribution difference of pores with different labels in the feature space is already quite obvious.

[0074] Secondly, using all the target domain data, 10%, 20%, 40%, 60%, 80% and 100% of the source domain data are used to build the pre-training model and DSAS training, and their accuracy and training time are compared on the target domain test set, as shown in Figure 7 and Figure 8 As shown in Figure 2, it can be seen that increasing the amount of data in the source domain can improve the monitoring accuracy of the model in the target domain. This may be because the source domain data not only determines the quality of the pre-training model, but also participates in the adaptation training between domains.

[0075] When the data in the source domain and the target domain exceed 60%, the influence of the increase in data volume on accuracy gradually decreases, which provides guidance for our future research on the efficient use of small sample data. Another conclusion is that the model training time increases with the increase in the amount of source domain data, while the amount of target domain data has almost no effect on the training time. This is because the amount of source domain data far exceeds that of target domain data and occupies the main resources in the training process.

[0076] The deep transfer learning method proposed in this embodiment can effectively utilize the historical knowledge of relevant experimental data to achieve porosity monitoring of data-scarce and more complex process parameters. Especially in the monitoring of LAM with high experimental and detection costs, using transfer learning to reduce the amount of data required and improve the generalization of the monitoring model can significantly improve the model development efficiency and reduce the monitoring cost, which is of great significance to the implementation and promotion of LAM quality monitoring.

[0077] The above description is only a preferred implementation manner of the present invention, but the protection scope of the present invention is not limited thereto, and any modification and replacement based on the technical solution and inventive concept provided by the present invention should be included in the protection scope of the present invention.

Claims

1. A method for online monitoring porosity of laser melting deposition based on deep transfer learning, characterized in that: The steps include: Step 1: Obtain source domain sample sets and target domain sample sets; The source domain sample set refers to a melt pool image dataset during the historical deposition process; The target domain sample set refers to a molten pool image data set collected during the deposition process of new process parameters; Step 2: Extract the pore data of the melt pool images in the source domain sample set and the target sample set respectively, and classify the pores according to the pore area; Step 3: Establish a source domain porosity online monitoring model using ShuffleNetV2 with grouped convolution and channel shuffling, and use the source domain sample set and its label classification to train the source domain porosity online monitoring model; Step 4: With the help of the source domain porosity online monitoring model, the target domain porosity online monitoring model is constructed based on deep sub-domain adaptive network transfer learning; The trained source domain porosity online monitoring model is used as the initial target domain porosity online monitoring model. The source domain sample set and the target domain sample set are respectively extracted from the source domain porosity online monitoring model and the initial target domain porosity online monitoring model. Then, similar samples of the source domain pore characteristics and the target domain pore characteristics are projected into different sub-domains, and the distribution alignment is performed by calculating the LMMD distance between the sub-domains; the differences between label classifications are highlighted in the process of transfer learning.

2. The method for online monitoring porosity of laser melting deposition based on deep transfer learning according to claim 1 is characterized in that: Before step 2, the melt pool image dataset is preprocessed. The preprocessing refers to determining the size of the melt pool image according to the maximum pixel value and the connected domain and cropping it, then scaling it to a fixed size through a bilinear interpolation algorithm, and finally performing grayscale normalization.

3. The method for online monitoring porosity of laser melting deposition based on deep transfer learning according to claim 1 is characterized in that: In step 2, X-rays are used to scan the printed part along the deposition direction to extract a tomographic image, and then the pore data in the tomographic image is extracted through an image processing algorithm. The pore data includes shape information, position and area, and the pores are labeled and classified according to the pore area.

4. The method for online monitoring porosity of laser melting deposition based on deep transfer learning according to claim 1 is characterized in that: The loss function Loss of the porosity online monitoring model is: Loss=L cross_entropy_target +λ1×L cross_entropy_source +λ2×L LMMD ; Among them, L cross_entropy_target is the cross entropy loss function of the target domain, L cross_entropy_source is the cross entropy loss function of the source domain; λ1 is the loss ratio parameter of the source domain; L LMMD is the local maximum difference loss function, λ2 is the domain adaptation loss weight; Among them, i represents the current training cycle, and nepoch represents the total training cycle.

Images