Laser directional energy deposition internal defect monitoring method based on semi-supervised learning

Abstract

The invention discloses a laser directional energy deposition internal defect monitoring method based on semi-supervised learning, and the method comprises the steps: 1, obtaining coaxial molten pool images, carrying out the preprocessing of the coaxial molten pool images, and stacking the coaxial molten pool images to form a domain-level data set; marking data: acquiring defect information of the printed deposition piece through the CT image; 2, establishing a convolution automatic encoder model, and training the convolution automatic encoder model by adopting the domain-level data set and the marked data; 3, connecting a classifier behind an encoder in the convolutional automatic encoder model, and mapping low-dimensional representation output by the encoder into a category space of a sample by using the classifier to realize classification of defect features; according to the method, a coaxial molten pool image is utilized, powerful understanding is established for molten pool characteristics and implicit knowledge in the LDED process based on convolutional auto-encoder unsupervised training, and finally, a category space of a sample is activated through a small amount of marked data, so that semi-supervised monitoring of laser directional energy deposition internal defects is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser additive manufacturing monitoring, and in particular relates to a method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning. Background Art

[0002] Laser Directed Energy Deposition (LDED) is an advanced additive manufacturing technology that uses a high-energy laser beam to melt metal powder or wire layer by layer to form functional metal parts. Due to its high flexibility and material utilization, as well as its ability to process a variety of metal materials, LDED has been widely used in aerospace, automotive, medical and energy fields. However, despite the many advantages of LDED technology, quality control in the manufacturing process remains a major challenge in this field.

[0003] Currently, quality control of the LDED process mainly relies on extensive labeling and experimental characterization based on professional knowledge. These methods are not only expensive and time-consuming, but also difficult to monitor in real time and comprehensively cover all process parameters. Therefore, it is particularly important to develop an efficient, real-time and economical monitoring method.

[0004] In recent years, data-driven methods have shown great potential in many engineering fields, especially in manufacturing process monitoring and optimization. By utilizing technologies such as big data, artificial intelligence, and machine learning, data-driven methods can mine valuable information from a large amount of historical data, thereby achieving real-time monitoring and quality control of complex processes. Introducing data-driven methods in the LDED process can effectively solve the problems faced by traditional methods, improve manufacturing quality, and reduce production costs.

[0005] Although the data-driven approach shows significant advantages in the LDED process, it also has some limitations. First, the initial deployment and training of the system requires a large amount of high-quality data support, which may be difficult to obtain quickly in actual production. Second, the real-time performance of the system may be limited by hardware performance when processing massive amounts of data. Summary of the invention

[0006] The purpose of the present invention is to provide a method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning. The method utilizes coaxial molten pool images and establishes a strong understanding of the molten pool characteristics and implicit knowledge in the LDED process based on unsupervised training of convolutional autoencoders. Finally, the class space of samples is activated by a small amount of labeled data to achieve semi-supervised monitoring of internal defects of laser directed energy deposition.

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

[0008] A method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning comprises the following steps:

[0009] Step 1: Obtain the coaxial melt pool image, preprocess the melt pool image, use the sliding time window method to monitor the melt pool state within a certain time and range, and stack the melt pool images within and between layers in an orderly manner according to the channel dimension to form a domain-level dataset;

[0010] Acquisition of labeling data: CT images are used to obtain the internal structure and defect information of the printed deposited parts, and the defect information is divided into three categories: negligible, medium-sized, and large;

[0011] Step 2: Build a convolutional autoencoder model and train the convolutional autoencoder model using the domain-level dataset and labeled data built in step 1;

[0012] Wherein, the convolutional autoencoder model includes an L2 normalization layer, an encoder and a decoder;

[0013] The L2 normalization layer is used to adjust data distribution, reduce the impact of abnormal samples on model training, and accelerate the gradient descent process to achieve optimal solution convergence;

[0014] The encoder is used to identify key features and defect features in the molten pool image; it includes four convolution modules, and the four convolution modules use residual connections to form an encoder;

[0015] The decoder is composed of four convolution sampling modules, and the four convolution sampling modules are connected in sequence;

[0016] Step 3: Decompose the encoder and decoder in the convolutional autoencoder model trained in step 2, and connect a classifier after the encoder to form a laser directed energy deposition internal defect monitoring model with semi-supervised learning; use the classifier to map the low-dimensional representation output by the encoder into the category space of the sample to achieve the classification of defect features;

[0017] Among them, when training the laser directed energy deposition internal defect monitoring model, the weight parameters of the encoder are fixed; and a small amount of labeled data is used to train the classifier.

[0018] Furthermore, preprocessing the melt pool image refers to first converting the melt pool image in RGB format into a grayscale image with pixel values ​​between 0 and 255, then performing ROI cropping with the melt pool point as the center, and finally further downsampling through bilinear interpolation, wherein the melt pool point refers to the brightest point in the image.

[0019] Furthermore, the threshold segmentation method is used to extract pores from the CT image and calculate their area; the pores are associated with the molten pool image, and the defect information is divided into three categories according to the size of the pore area.

[0020] Furthermore, the convolution module includes a convolution layer, a batch normalization layer, an activation function layer and a pooling layer connected in sequence.

[0021] Furthermore, the convolution sampling module includes a convolution layer, a batch normalization layer, an activation function layer and an upsampling layer.

[0022] Furthermore, the classifier includes a classification module and an output module, the classification module includes a fully connected layer, a batch normalization layer, an activation function layer and a random inactivation layer connected in sequence, and the output module is implemented using a fully connected layer.

[0023] The present invention has the following beneficial effects:

[0024] (1) Construct a domain-level dataset to characterize the quality attributes of local areas during LDED processing, taking into account spatiotemporal factors such as the dynamic changes of the melt pool over time and the mutual influence between different layers;

[0025] (2) A semi-supervised convolutional autoencoder model is established to achieve real-time monitoring of the quality of local areas during LDED printing with a small amount of labeled data. Unsupervised training plus category space activation is used to establish a strong understanding of the melt pool characteristics of additive data, significantly reducing the need for labeled data and enhancing the accuracy and robustness of monitoring.

[0026] (3) The autoencoder structure is designed for melt pool characteristics and singular samples. The morphological characteristics of the melt pool are enhanced in the deep network of the encoder, and the invariance of the model to pore scale changes is maintained by constraining the samples, making the autoencoder more suitable for the characteristics of additive manufacturing data.

[0027] (4) The semi-supervised convolutional autoencoder model is verified through experimental data; the experimental results prove the effectiveness of local quality monitoring in the LDED process; by comparing with other semi-supervised methods, the reasons why LDED data is incompatible with other semi-supervised methods are deeply analyzed, thereby proving the superiority of the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the overall framework of the model proposed in the present invention.

[0029] Figure 2 Schematic diagram of printing strategy and sample construction; (a) linear reciprocating scanning strategy; (b) construction of domain-level samples.

[0030] Figure 3 Schematic diagram of the ablation experiment training process; (a) loss on the training set, (b) accuracy on the test set.

[0031] Figure 4Classification accuracy of the three region qualities under different annotated data ratios.

[0032] Figure 5 is the confusion matrix of SCAE on the test set.

[0033] Figure 6 ROC curves and AUC scores of different semi-supervised models on the test set; (a) meanteacher, (b) pimodel, (c) FixMatch, (d) SoftMatch, and (e) the SCAE model of the present invention. DETAILED DESCRIPTION

[0034] This embodiment provides a method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning, comprising the following steps:

[0035] Step 1: Obtain melt pool images and construct domain-level datasets using coaxial melt pool images.

[0036] In the process of LDED printing deposition parts, a camera coaxially installed with the laser beam is used to obtain coaxial melt pool images in real time. The melt pool images are preprocessed, and a sliding time window is used to monitor the melt pool state within a certain time and range. The melt pool images within and between layers are stacked in an orderly manner according to the channel dimension to form a domain-level dataset.

[0037] Preprocessing the melt pool image means first converting the melt pool image in RGB format into a grayscale image with pixel values ​​between 0 and 255, then cropping the ROI with the melt pool point as the center, and finally further downsampling through bilinear interpolation to reduce the computational burden of the deep learning process. The melt pool point refers to the brightest point in the image.

[0038] The domain-level dataset can reflect the dynamic melt pool information within and between layers in the local area. This embodiment uses a sampling time window with a length of 0.4s to monitor the melt pool status information within a certain time and range, and then stacks the corresponding melt pool images in the 0.4s sampling time window of each layer in an orderly manner along the channel dimension to form a domain-level dataset, reflecting the quality status of an area.

[0039] Acquisition of marked data: X-ray CT imaging technology is used to obtain the internal structure and defect information of the printed deposited parts, and to quantify the local defect information in the deposited parts.

[0040] This embodiment uses a threshold segmentation method to extract pores from CT images and calculate their areas; and associates the pores with the molten pool image, and divides the defect information into three categories according to the size of the pore area: negligible (marked as 0), medium size (marked as 1), and large size (marked as 2).

[0041] Step 2: A convolutional autoencoder model is established based on the morphological characteristics and variable porosity distribution of the melt pool, and the convolutional autoencoder model is trained using the domain-level dataset constructed in step 1.

[0042] The convolutional autoencoder can reconstruct all melt pool images without the need for data labels, and learn and extract effective defect features from the original data set; Figure 1 As shown, the convolutional autoencoder includes an L2 normalization layer, an encoder and a decoder.

[0043] Based on the visualization results of CT scanning, it can be seen that the defect information contains some large pores, resulting in the presence of singular samples in the data set; during the training stage, singular samples may cause numerical instability, especially for smaller eigenvectors that may be dominated by larger eigenvectors. This embodiment uses the L2 normalization layer to reduce the impact of these singular samples on training, and projects the samples onto the surface of the unit sphere for constraints; ensure that the size of the vector is evenly scaled so that each vector can play an equal role in calculating the distance.

[0044] The main point of L2 normalization is that it can enhance the invariance of data to scale changes. That is, no matter what the size or scale of the original data is, the L2 normalized data can maintain consistent distribution characteristics as much as possible. This feature enhances numerical stability and accelerates convergence during training.

[0045] The encoder is used to identify key features and defect features in the molten pool image; this embodiment uses a residual network structure to construct the encoder, improves feature utilization efficiency, allows information to flow between different layers, and enhances the network's ability to focus on molten pool morphological features (such as edges and contours).

[0046] The encoder is composed of a plurality of convolution modules, each of which has the same structure, and the convolution modules include a convolution layer, a batch normalization layer, an activation function layer, and a pooling layer connected in sequence. In order to more conveniently describe the structure of the encoder, the present embodiment adds the first, second, third, and fourth before the convolution module to distinguish them, which has no substantial meaning. Specifically, the encoder includes a first convolution module, a second convolution module, a third convolution module, and a fourth convolution module. The first convolution module, the second convolution module, and the third convolution module are connected in sequence, and the outputs of the first convolution module and the third convolution module are added and connected to the fourth convolution module.

[0047] In order to enhance the encoder's extraction of melt pool features, residual connections are used to construct the encoder, and the shallow network features are merged with the deep network features to enhance the morphological characteristics of the melt pool; this not only speeds up the training process, but also effectively improves the model's extraction effect on the melt pool features and improves the model's classification performance.

[0048] The output of each convolution module in the encoder is xl =P(Conv l (x l-1 )), where x l-1 Represents the output of the previous convolution module, Conv l (·) is a convolution operation, which includes three consecutive transformations: convolution, batch normalization, and nonlinear activation; P(·) is a maximum pooling operation.

[0049] The decoder is used to restore the image resolution by upsampling. The purpose of using the decoder in this embodiment is to complete the training of the encoding part in the automatic encoder so that the encoding part can accurately identify the defect features in the molten pool image. The decoder is composed of multiple convolution sampling modules, each of which has the same convolution module structure, including a convolution layer, a batch normalization layer, an activation function layer and an upsampling layer. The four convolution modules are connected in sequence to output the restored image.

[0050] The output of each convolution sampling module of the decoder is x′ l =U(Conv l (x′ l-1 )), where x′ l-1 Represents the output of the previous convolution sampling module, Conv l (·) is a convolution operation, which includes three consecutive transformations: convolution, batch normalization, and nonlinear activation; U(·) is an upsampling operation.

[0051] The domain-level data in step 1 is used to train the convolutional autoencoder model. The Adam optimization algorithm and the error back propagation algorithm are used to iteratively optimize the weight parameters and bias parameters of the model to minimize the reconstruction loss function and obtain the optimized network parameters.

[0052] Step 3: Decompose the encoder and decoder in the convolutional autoencoder model trained in step 2, and connect a classifier after the encoder to form a laser directed energy deposition internal defect monitoring model with semi-supervised learning; use the classifier to map the low-dimensional representation of the encoder output into the category space of the sample to realize the classification of defect features.

[0053] The classifier includes a classification module and an output module. The classification module includes a fully connected layer, a batch normalization layer, an activation function layer and a random inactivation layer connected in sequence. The output module is implemented using a fully connected layer.

[0054] When training the laser directed energy deposition internal defect monitoring model, the weight parameters of the encoder are fixed; a small amount of labeled data is used to train the classifier, and the Adam optimization algorithm and the error back propagation algorithm are used to iteratively optimize the weight parameters and bias parameters of the model to reconstruct the loss function and obtain the optimized network parameters.

[0055] This embodiment uses the capabilities of convolutional neural networks and autoencoders to extract melt pool features in an unsupervised manner; then a small number of labeled samples and classifiers are used to introduce category information to the encoder. This method can successfully implement semi-supervised learning in the LDED monitoring process.

[0056] Experimental verification of the effectiveness of this embodiment

[0057] Experimental equipment: The fiber laser (TurDisk 4006) and laser head (YC52) were installed on a six-axis robotic arm (RX-160), and the powder feeder (RC-PGF-D) was filled with 30CrNi2MoVA metal powder; the CCD camera (CF8 / 5MX) was installed coaxially with the laser beam, and a beam splitter and filter were configured in the optical path.

[0058] The substrate was polished before the experiment, and the real-time molten pool image (768×494 pixels) was captured at a frequency of 50fps; the laser emitted a high-energy laser beam with a wavelength of 1030nm, which was reflected to the substrate through a beam splitter covered with a 1030nm high-reflectivity transparent film; the interaction between the laser and the material occurred on the substrate, forming an interaction area centered on the molten pool. The light emitted by the molten pool was captured as a molten pool image by a coaxial camera after passing through a spectrometer and a reflector.

[0059] The deposition parts were printed using different process parameters in Table 1. Laser power and scanning speed play a crucial role in determining the quality of the deposition parts. Therefore, various combinations of these two parameters were used to enrich the experimental data. The deposition parts were printed using a linear reciprocating scanning strategy. Each deposition part was 40 mm long and consisted of 30 layers, such as Figure 2 As shown, the reciprocating printing strategy helps to evenly distribute heat while reducing heat accumulation and thermal stress within the printed parts, ultimately improving printing quality and efficiency.

[0060] Table 1 Process parameters

[0061]

[0062] The domain-level dataset was constructed and trained and tested on four NVIDIA GeForce RTX 2080Ti GPUs. A total of 2866 samples were collected and then divided into training and testing datasets in a ratio of 8:2. Table 2 lists some of the hyperparameters used in the model architecture. The conv2d layer and linear layer in the model are initialized with kaiming_normal and a normal distribution with a mean of zero and a standard deviation of 0.01, respectively. We use the Adam optimizer with a learning rate of 0.01, and use the MultiStepLR strategy with a gamma value of 0.1 to adjust the learning rate.

[0063]

[0064] Ablation experiment

[0065] In order to intuitively demonstrate the effectiveness of residual connections and L2 normalization, a series of experiments were conducted using different model configurations: baseline models lacking residual connections and L2 normalization, CAE-L2 models lacking residual connections, CAE-Res models lacking L2 normalization, and CAE-Res-L2 models that combine residual connections and L2 normalization. 20% of the labeled samples were randomly selected for training, and the performance of the model was evaluated on the test dataset. Figure 3 The training loss and test accuracy of the four models are shown. Table 3 summarizes the classification results of these models.

[0066] Table 3 Ablation experiment study of L2 normalization and residual connection

[0067]

[0068] As can be seen from Table 3, the use of residual connection and L2 normalization techniques can improve the prediction accuracy by 2.3% and 1.8%, respectively. Using these two techniques together can improve the accuracy by 3.7%, which predicts the negligible state with 85% accuracy, the medium defect state with 77% accuracy, and the large defect state with 90% accuracy. These experiments show that L2 normalization is able to adjust the data distribution, ensure that the weight of each feature dimension on the objective function is consistent, and mitigate the impact of outliers. The residual connection module enhances the representation and feature utilization of the model by reusing shallow features, thereby better learning the morphological characteristics of the melt pool.

[0069] To further demonstrate the effectiveness and efficiency of L2 normalization, the effects of different normalization methods on model performance were studied. As shown in Table 4, Min-max, Z-score, and L2 normalization methods all produced positive results, among which L2 normalization had the best enhancement effect. In contrast, L1 normalization led to a decrease in model performance. The rationale behind these results can be attributed to the characteristics of each normalization method. Min-max normalization has limited effect on outliers because it simply scales the elements according to their maximum and minimum values, which may ignore extreme values. L1 normalization divides each element by the sum of its absolute values, which may cause some elements to approach zero when there are large outliers in the vector, affecting the entire training process. Z-score is essentially a data normalization method that does not change the distribution of the original data. The squared effect of outliers significantly increases the vector size, making L2 normalization particularly effective in addressing the impact of outliers. This sensitivity to outliers allows L2 normalization to improve model performance by better adjusting the data distribution and mitigating the impact of outliers.

[0070] Table 4 Comparison of classification performance using different normalization methods

[0071]

[0072] Semi-supervised performance with different proportions of labeled data

[0073] In order to demonstrate the superiority of the semi-supervised learning method proposed in this embodiment, semi-supervised and fully supervised learning models with the same network structure, depth and training hyperparameters are constructed. 458 (20%), 687 (30%), 917 (40%), 1146 (50%), 1375 (60%), 1604 (70%) and 2292 (100%) labeled samples are randomly selected for training, and the performance of the model on the test data set is evaluated. Table 5 shows the performance indicators of the proposed semi-supervised method and the fully supervised method when trained with different proportions of labeled data. Figure 4 The prediction accuracy of three quality states under the same conditions was compared.

[0074] Table 5. Semi-supervised and fully supervised learning performance under different labeled data ratios

[0075]

[0076] It can be seen from Table 5 that the semi-supervised learning method proposed in this embodiment significantly improves the model performance, especially when the labeled data is limited. Specifically, when only 20% of the labeled data is used, the prediction accuracy of the supervised learning method is 70.2%, while the semi-supervised learning method achieves an accuracy of 83.8%, which improves the prediction accuracy by 13.6%. In terms of the efficiency of labeled data utilization, although the labeled data is reduced by 80%, the prediction accuracy of the semi-supervised method only decreases by 6.1%. It is worth noting that the semi-supervised method achieves a performance level close to that of the fully supervised method using only 60% of the labeled data. Moreover, when all the labeled data is used, the accuracy of the semi-supervised method reaches a peak of 90.8%, exceeding the fully supervised method.

[0077] like Figure 4 As shown, the accuracy of the semi-supervised and fully supervised methods increases with the increase in the number of labeled data. Comparing the results between the semi-supervised and fully supervised methods, the proposed semi-supervised method is always better than the fully supervised learning method in all data scenarios, which proves the feasibility and effectiveness of the semi-supervised monitoring method proposed in this embodiment in the LDED process.

[0078] like Figure 5As shown, the classification confusion matrix of the SCAE model on the test set shows that its prediction accuracy for the negligible state is 93, for the medium-quality state is 86, and for the large-defect quality state is 94. The model has a relatively high prediction accuracy for large-scale quality states, which is conducive to realizing the optimal control of quality. In supervised learning, the model uses the label as a guide to enforce training and learning. However, the label usually includes noise or incorrect labels, resulting in a decline in the learning result. On the contrary, semi-supervised learning methods utilize a large amount of unlabeled data, which may reduce the impact of noisy labels. The autoencoder in this embodiment extracts richer and more accurate feature representations from the unlabeled raw data, effectively capturing the internal structure and patterns in the data, thereby enhancing the model's robustness to noise and outliers.

[0079] The semi-supervised learning method proposed in this embodiment effectively mines the potential features of unlabeled data and effectively utilizes the labeled data information. This method significantly improves the model performance in the case of limited labeled data, solves the challenges brought by the scarcity of labels in LDED process monitoring, and reduces the dependence on labeled data. Obtaining labeled data in LDED monitoring is time-consuming, difficult, and expensive. The semi-supervised method proposed in this paper can effectively reduce the time and cost associated with LDED process monitoring and provide more accurate and reliable predictions.

[0080] Comparison with Other Advanced Methods

[0081] The model (SCAE) of this embodiment is compared with the leading SOTA models in the field of semi-supervised learning, including MeanTacher, Pimodel, FixMatch, and SoftMatch; the same hardware and hyperparameter settings are used during training.

[0082] Table 6 shows the performance of each model. The prediction accuracy, computational complexity, number of parameters of each model are calculated, and for the time consumption, the average inference time of the model for 100 input data is calculated.

[0083] Table 6 Performance Comparison of Different Semi-Supervised Models

[0084]

[0085] Figure 6 The ROC curve and AUC area of each model are given. The ROC curve evaluates the performance of the model for classification problems at different threshold levels, and the curve closer to the upper left corner indicates better performance. The AUC score reflects the ability of the classifier to rank samples, and the higher the AUC, the better the discrimination between true positives and true negatives.

[0086] The results show that the SCAE model proposed in this embodiment achieves the highest accuracy and has significant advantages in terms of parameter quantity and model inference speed (FPS). Specifically, the average AUC score of SCAE is 0.90, which is significantly higher than other models. As shown in Table 6, the classification prediction accuracy of MeanTeacher and Pimodel is about 60% on average, and they use data augmentation and consistency regularization to capture the key features of the target. However, the LDED sample consists of a series of dynamic images; data augmentation may change the size, brightness and tail direction of the melt pool, invalidate the sample, and produce incorrect sample label pairs. When data augmentation techniques such as random cropping and flipping are applied under full supervision, the accuracy drops by 3.7%, as shown in Table 7; this shows that data augmentation is not effective in the mechanism additive manufacturing dataset. FixMatch and SoftMatch are hybrid models that combine consistency regularization and pseudo-labeling, which generate pseudo-labels for unlabeled samples with high confidence. However, overconfident pseudo-labels may cause the model to overfit. This is a big problem in the presence of noisy labels in the dataset, because incorrect pseudo-labels introduce too much noise and mislead the learning process. FixMatch’s prediction accuracy is 60.1%, while SoftMatch’s is 65.1%, which is slightly higher due to its dynamic threshold mechanism that may filter out some incorrect pseudo-labels. During training, the accuracy of labeled data quickly reaches 99%, while the accuracy of unlabeled data is only 55%.

[0087] Table 7 Performance of data augmentation in fully supervised models

[0088]

[0089] The autoencoder used in this embodiment reconstructs data directly on the original data, and establishes a deep understanding of the characteristics and defect information of the molten pool without the need for additional image preprocessing and data enhancement. Pseudo-labeling technology is not used in the overall semi-supervised learning framework, and only a small amount of labeled data is used for category space activation, which effectively reduces the impact of noise labels.

[0090] 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 monitoring internal defects of laser directed energy deposition based on semi-supervised learning, characterized in that: The steps include: Step 1: Obtain the coaxial melt pool image, preprocess the melt pool image, use the sliding time window method to monitor the melt pool state within a certain time and range, and stack the melt pool images within and between layers in an orderly manner according to the channel dimension to form a domain-level dataset; Acquisition of labeling data: CT images are used to obtain the internal structure and defect information of the printed deposited parts, and the defect information is divided into three categories: negligible, medium-sized, and large; Step 2: Build a convolutional autoencoder model and train the convolutional autoencoder model using the domain-level dataset and labeled data built in step 1; Wherein, the convolutional autoencoder model includes an L2 normalization layer, an encoder and a decoder; The L2 normalization layer is used to adjust data distribution, reduce the impact of abnormal samples on model training, and accelerate the gradient descent process to achieve optimal solution convergence; The encoder is used to identify key features and defect features in the molten pool image; it includes four convolution modules, and the four convolution modules use residual connections to form an encoder; The decoder is composed of four convolution sampling modules, and the four convolution sampling modules are connected in sequence; Step 3: Decompose the encoder and decoder in the convolutional autoencoder model trained in step 2, and connect a classifier after the encoder to form a laser directed energy deposition internal defect monitoring model with semi-supervised learning; use the classifier to map the low-dimensional representation output by the encoder into the category space of the sample to achieve the classification of defect features; Among them, when training the laser directed energy deposition internal defect monitoring model, the weight parameters of the encoder are fixed; and a small amount of labeled data is used to train the classifier.

2. The method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning according to claim 1, characterized in that: The preprocessing of the melt pool image refers to first converting the melt pool image in RGB format into a grayscale image with pixel values ​​between 0 and 255, then performing ROI cropping with the melt pool point as the center, and finally further downsampling through bilinear interpolation, wherein the melt pool point refers to the brightest point in the image.

3. The method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning according to claim 1, characterized in that: Threshold segmentation method is used to extract pores from CT images and calculate their areas; The pores are associated with the molten pool image, and the defect information is divided into three categories according to the size of the pore area.

4. The method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning according to claim 1, characterized in that: The convolution module includes a convolution layer, a batch normalization layer, an activation function layer and a pooling layer which are connected in sequence.

5. The method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning according to claim 1, characterized in that: The convolution sampling module includes a convolution layer, a batch normalization layer, an activation function layer and an upsampling layer.

6. The method for monitoring internal defects of laser directed energy deposition based on semi-supervised learning according to claim 1, characterized in that: The classifier includes a classification module and an output module. The classification module includes a fully connected layer, a batch normalization layer, an activation function layer and a random inactivation layer connected in sequence. The output module is implemented using a fully connected layer.

Images