Method for identifying fusion defect based on thermal scanning tomographic image

Abstract

The invention discloses a method for identifying a fusion defect based on a thermal scanning tomography image, and belongs to the technical field of additive manufacturing, and the method comprises the following steps: S1, carrying out an image abnormal point detection algorithm; the specific steps in the step S1 are as follows; s11, performing thermal scanning chromatography in-situ monitoring and data collection; s12, data preprocessing and first feature extraction; s13, performing statistical analysis or feature matching; s14, evaluating a result and setting a threshold value; s15, deploying an application; according to the method, the thermal historical information of the formed section is observed on line in situ, so that in-situ identification of the fusion defect is realized, the problem of lack of on-line monitoring quality control technical means in LPBF industrial application can be solved, and macroscopic external quality defects and internal quality defects can be effectively monitored at the same time; and moreover, the use frequency of high-cost nondestructive testing can be reduced, and loss stopping or feedback intervention repairing can be carried out in time when manufacturing abnormity occurs.

Description

Technical Field

[0001] The present invention belongs to the technical field of additive manufacturing, and particularly relates to a method for identifying melting defects based on thermal scanning tomography images. Background Art

[0002] Laser powder bed fusion (LPBF) has been widely used in the manufacturing process of complex components of various types of equipment due to its obvious advantages in the fine manufacturing of complex structures. However, the thermophysical and non-equilibrium metallurgical processes of materials during its process are very complex, and it is easy to form other internal quality defects such as pores, lack of fusion, and microcracks. Due to the possible existence of these internal quality defects, non-destructive testing such as X-ray or industrial CT is often required before the delivery of aviation parts. However, industrial CT has extremely high detection costs and can only be detected offline afterwards; for the LPBF manufacturing system, due to the complex control system and extreme dependence on process plans, some accidental manufacturing anomalies will cause very serious losses due to the lack of equipment stability and process personnel experience, such as insufficient powder supply caused by abnormal powder supply systems and external quality problems such as part warping and deformation caused by inappropriate process plans;

[0003] In the current existing technical means, for manufacturing anomalies such as powder supply problems, CCD or CMOS cameras are usually used to capture the powder bed powder laying integrity information, and then identify whether the powder completely covers the part. However, this method cannot effectively identify warping and deformation; structured light imaging is used, and based on stripe deformation, the Z-direction cross-section undulation can be detected, but there are efficiency problems in forming a high-resolution height map and it cannot meet the generation requirements; for melting defects such as pores, lack of fusion, cracks, etc., synchrotron X-rays can be used to accurately identify them, but this method can only be carried out on experimental instruments and usually single-channel or multi-channel experiments are carried out and it is not applicable to industrial scenarios. Internationally, coaxial photodiodes are also used to monitor the radiation light signals generated by the molten pool, spatter, metal vapor, etc., and analyze the melting defects through the radiation light signals of the molten pool. However, this method is extremely difficult to be associated with the process. Decoupling the one-dimensional signal corresponding to dozens of influencing factors is difficult and it has almost no engineering applicability. In addition, its intrusion into the optical path will also affect the beam quality;

[0004] Therefore, a method for identifying melting defects based on thermal scanning tomography images is needed to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for identifying melting defects based on thermal scanning tomography images to solve the problems raised in the above background art.

[0006] To achieve the above purpose, the present invention provides the following technical solution: A method for identifying melting defects based on thermal scanning tomography images, comprising the following steps:

[0007] S1. Image Abnormal Point Detection Algorithm;

[0008] The specific steps in "Step S1" are as follows;

[0009] S11. Thermal Scanning Tomography In-situ Monitoring and Data Collection;

[0010] S12. Data Preprocessing and Feature Extraction I;

[0011] S13. Statistical Analysis or Feature Matching;

[0012] S14. Result Evaluation and Threshold Setting;

[0013] S15. Deployment and Application;

[0014] S2. Supervised Learning Model;

[0015] The specific steps in "Step S2" are as follows;

[0016] S21. Thermal Scanning Tomography In-situ Monitoring and Data Collection;

[0017] S22. X-ray Computed Tomography or Metallographic Analysis;

[0018] S23. Data Preprocessing and Data Alignment;

[0019] S24. Tomographic Image and Fused Defect Dataset Construction;

[0020] S25. Model Construction and Training Verification;

[0021] S26. Model Deployment and Application;

[0022] S3. Unsupervised Learning Model;

[0023] The specific steps in "Step S3" are as follows;

[0024] S31. Thermal Scanning Tomography In-situ Monitoring and Data Collection: The same as this step above;

[0025] S32. X-ray Computed Tomography or Metallographic Analysis: The same as this step above;

[0026] S33. Data Preprocessing and Feature Extraction II, and the specific features in feature extraction are: texture features, statistical features, shape features, and frequency domain features;

[0027] S34. Clustering and Model Training;

[0028] S35. Model Deployment and Application.

[0029] This method: First, after collecting the in-situ monitoring data of thermal scanning tomography, traditional image anomaly detection algorithms (such as statistical methods and feature matching algorithms) are used to identify the anomaly points, and then compared with the X-ray or metallographic analysis results for evaluation. After determining the detection threshold, a defect identification method is deployed;

[0030] Second, a supervised learning method is adopted, that is, by simultaneously collecting the in-situ monitoring data of thermal scanning tomography and X-ray or metallographic analysis image data, and then constructing a machine learning model for deployment. After that, the defects are predicted online through the supervised learning model;

[0031] Third, after collecting the in-situ monitoring data of thermal scanning tomography, direct clustering analysis is performed through an unsupervised learning model to determine whether there are melting defects.

[0032] As a preferred solution, in "Step S12", the data preprocessing and feature extraction include the following steps:

[0033] S121. Grayscale processing: Convert the thermal scanning tomography image into a grayscale image to reduce the data volume and computational complexity;

[0034] S122. Noise removal;

[0035] S123. Image enhancement and size normalization;

[0036] S124. Feature extraction.

[0037] As a preferred solution, in "Step S13", the statistical analysis or feature matching includes the following steps:

[0038] S131. Statistical analysis;

[0039] S132. Feature matching.

[0040] As a preferred solution, in "Step S22", the X-ray computed tomography or metallographic analysis includes the following steps: S221. X-ray computed tomography;

[0041] S222. Metallographic analysis.

[0042] As a preferred solution, in "Step S25", the model construction and training verification include the following steps:

[0043] S251. Model architecture design: Select a deep learning model architecture suitable for image data, such as a convolutional neural network (CNN), and design the input layer, convolutional layer, pooling layer, fully connected layer, and output layer to ensure that the model can effectively extract image features and perform defect classification;

[0044] S252. Dataset preparation;

[0045] S253. Model training;

[0046] S254. Model verification and evaluation;

[0047] S255. Cross-validation.

[0048] As a preferred solution, in "Step S34", the clustering and model training include the following steps:

[0049] S341. Selection of the model;

[0050] S342. Clustering model training.

[0051] As a preferred solution, in "Step S11", the thermal scanning tomography in-situ monitoring and data collection adopt a thermal scanning tomography on-line monitoring device, and the thermal scanning tomography on-line monitoring device includes a laser, a laser beam, a scanning device, a part forming surface, a forming platform, layer cross-section radiation, an observation window, an optical component, a thermal scanning tomography sensor, and an upper top plate of the forming cavity;

[0052] During the laser powder bed fusion processing, the laser emits a laser beam, and the laser beam heats and melts the metal powder on the part forming surface through the scanning device. During the heating, melting, and cooling and solidification processes, layer cross-section radiation will be formed on the layer cross-section. The layer cross-section radiation is conducted to the observation window located on the upper top plate of the forming cavity through spatial light, and then conducted to the tomography detection sensor through the optical component, and collected by the tomography detection sensor to form a thermal scanning tomography image layer by layer. During the process of printing and processing parts, the thermal scanning tomography image data is collected layer by layer.

[0053] Compared with the prior art, the beneficial effects of the present invention are:

[0054] In the present invention, by on-line in-situ observing the thermal history information of the forming cross-section, and then realizing the in-situ identification of fusion defects, it can not only solve the problem of the lack of on-line monitoring quality control technical means in LPBF industrial applications, but also effectively monitor macroscopic external quality defects and internal quality defects at the same time; and can reduce the use frequency of high-cost non-destructive testing, and can stop losses in time or feedback and intervene for repair when manufacturing anomalies occur. Description of the Drawings

[0055] Figure 1 It is a flow chart of the present invention;

[0056] Figure 2 It is a schematic diagram of data preprocessing and feature extraction I of the present invention;

[0057] Figure 3 It is a schematic diagram of statistical analysis or feature matching of the present invention;

[0058] Figure 4Schematic diagram of X-ray computed tomography or metallographic analysis of the present invention;

[0059] Figure 5 Schematic diagram of model construction and training verification of the present invention;

[0060] Figure 6 Schematic diagram of data preprocessing and feature extraction II of the present invention;

[0061] Figure 7 Schematic diagram of clustering and model training of the present invention;

[0062] Figure 8 Schematic diagram of the structure of the thermal scanning tomography online monitoring device of the present invention;

[0063] Figure 9 Thermal scanning tomography image data diagram of the present invention.

[0064] In the figure: 1. Laser; 2. Laser beam; 3. Scanning device; 4. Part forming surface; 5. Forming platform; 6. Layer cross-section radiation; 7. Observation window; 8. Optical component; 9. Thermal scanning tomography sensor; 10. Upper roof plate of the forming cavity. Detailed implementation manners

[0065] The present invention will be further described below in conjunction with embodiments.

[0066] The following embodiments are used to illustrate the present invention, but cannot be used to limit the protection scope of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions. Any simple improvement of the method of the present invention under the premise of the concept of the present invention belongs to the scope required to be protected by the present invention.

[0067] Please refer to Figures 1-9 , the present invention provides a method for identifying fusion defects based on thermal scanning tomography images, including the following steps:

[0068] S1. Image anomaly point detection algorithm, which is a method for identifying fusion defects based on statistical analysis and feature matching. It is easy to implement and debug, has high computing efficiency, low consumption of computing resources, and low requirements for data volume; when the LPBF equipment system does not have redundant computing resources and has high real-time requirements, this method can be used for defect identification;

[0069] The specific steps in "Step S1" are as follows;

[0070] S11. Thermal scanning tomography in-situ monitoring and data collection;

[0071] The thermal scanning tomography in-situ monitoring and data collection uses a thermal scanning tomography online monitoring device, which includes a laser 1, a laser beam 2, a scanning device 3, a part forming surface 4, a forming platform 5, a layer cross-section radiation 6, an observation window 7, an optical component 8, a thermal scanning tomography sensor 9, and an upper roof plate 10 of the forming cavity;

[0072] During the laser powder bed melting process, the laser 1 emits a laser beam 2, and the laser beam 2 heats and melts the metal powder on the part forming surface 4 through the scanning device 3. During the heating, melting, and cooling and solidification processes, a layer cross-section radiation 6 will be formed on the layer cross-section. The layer cross-section radiation 6 is conducted to the observation window 7 located on the upper roof plate 10 of the forming cavity through spatial light, and then conducted to the tomography detection sensor 9 through the optical component 8, and collected by the tomography detection sensor 9 to form a thermal scanning tomography image layer by layer. During the process of printing and processing parts, the thermal scanning tomography image data is collected layer by layer;

[0073] S12. Data preprocessing and feature extraction I; The data preprocessing and feature extraction of the thermal scanning tomography image altogether include the following steps: S121. Grayscale processing: Convert the thermal scanning tomography image into a grayscale image to reduce the data volume and computational complexity;

[0074] Among them, grayscale processing can be carried out by methods such as weighted average method, maximum value method, and average value method;

[0075] The weighted average method is to perform weighted averaging on the three RGB components of the color image to obtain a grayscale image, and its calculation method is: Gray(i,j) = 0.299 * R(i,j) + 0.578 * G(i,j) + 0.114 * B(i,j);

[0076] Maximum value method: Take the value of the component with the largest value among the three RGB components, that is, Gray(i,j) = Max(R(i,j) + G(i,j) + B(i,j));

[0077] Average value method: Take the average value of the values of the three RGB components, that is, Gray(i,j) = (R(i,j) + G(i,j) + B(i,j)) / 3;

[0078] S122. Noise removal: Use a bilateral filter for noise reduction processing. On the one hand, it can combine spatial proximity and pixel value similarity, and at the same time can retain edge information while reducing noise in the image;

[0079] The bilateral filter can be expressed as:

[0080] I′ x，y =Wp1Σx i ,y i Gσs(‖x_x i‖·Gσ r (|Ix,y_Ix i ,y i |)·Ix i ,y i where:

[0081] I′x,y is the pixel value of the output image at the position (x,y);

[0082] Ix,y is the pixel value of the input image at the position;

[0083] Gσs is a spatial Gaussian function used to consider the spatial distance of pixels;

[0084] Gσr is a range Gaussian function used to consider the difference in pixel values;

[0085] Wp is a normalization factor to ensure that the sum of the output pixel values is 1;

[0086] S123, Image enhancement and size normalization;

[0087] Perform enhancement processing on the image to highlight important features in the image;

[0088] The enhancement methods include histogram equalization and contrast enhancement;

[0089] Histogram equalization makes the contrast of the image more uniform by adjusting the histogram distribution of the image; Contrast enhancement can be achieved through linear transformation or non-linear transformation (such as gamma correction) to improve the contrast of the image;

[0090] Size normalization is to scale the image to a unified size for subsequent processing;

[0091] Size normalization can be achieved through interpolation algorithms such as bilinear interpolation and bicubic interpolation to ensure consistent feature extraction effects of the image at different scales.

[0092] S124, Feature extraction;

[0093] If statistical analysis is adopted, it is necessary to extract global features or local features of the image;

[0094] Such as the average gray value, variance, and standard deviation statistics of the image. These global features can reflect the overall characteristics of the image and provide basic data for subsequent Gaussian model modeling;

[0095] Extracting local features mainly includes edge features and texture features. Edge features can be extracted through algorithms such as the sobel operator and the canny operator, and texture features can be extracted through the gray-level co-occurrence matrix to extract uniformity and contrast features;

[0096] If the feature matching method is adopted, a template is generated according to the image of the normal printing sample (positive sample); the template can be the key areas and feature points in the image. Feature points in the image are extracted through feature detection algorithms (such as SIFT, Harris), and templates are generated centered on these feature points;

[0097] After generating the template, the local binary pattern (LBP) feature descriptor is used to describe the image region. LBP generates a binary code and then converts it into a feature vector by calculating the gray value relationship between each pixel point in the image region and its surrounding pixel points. Furthermore, the feature of each region is represented as a feature vector, forming a set of feature vectors;

[0098] S13. Statistical analysis or feature matching;

[0099] S131. Statistical analysis: Assume that the features of normal images follow a Gaussian distribution, and detect anomalies by calculating the deviation between the features of the image to be detected and the Gaussian model; or use the Mahalanobis distance to measure the difference between the features of the image to be detected and the features of normal samples. If it exceeds a certain threshold, it is considered an anomaly;

[0100] S132. Feature matching: Calculate the distance between the feature vectors of the image to be detected and the template feature vectors. Commonly used distance measurement methods include the Euclidean distance;

[0101] According to the distance magnitude, find the regions in the image to be detected that are most similar to the template features. These regions are the matching regions. Through the matching results, determine whether there are anomalies in the image to be detected. If there are many regions with low matching degrees or no matches, it may be an anomaly;

[0102] S14. Result evaluation and threshold setting;

[0103] Compare the abnormal regions detected by the above methods with the actual abnormal regions obtained by X-ray or metallographic inspection, evaluate the accuracy and robustness of the above detection algorithms, and obtain appropriate threshold setting values through the evaluation results.

[0104] S15. Deployment and application;

[0105] After completing the above steps, integrate the thermal scanning tomography online monitoring device into the LPBF printer system, deploy the algorithm developed in the above steps, and perform detection and analysis on the thermally scanned tomography images generated layer by layer, then printing defects can be identified.

[0106] S2. Supervised learning model and a method for identifying fusion defects in thermal scanning tomography images based on the supervised learning model. This method can make full use of data, combining in-situ monitoring data of thermal scanning tomography and X-ray or metallographic analysis image data, and can obtain characteristic information of defects from different angles and dimensions. The thermal scanning tomography data provides changes in the thermal characteristics inside the material, while the X-ray or metallographic images show the morphology and structure of the defects. This multi-source data fusion can more comprehensively describe the defect characteristics, thereby improving the prediction accuracy of the model. Moreover, this method has strong generalization ability. Supervised learning usually uses deep learning models, which have strong non-linear mapping ability and can learn complex characteristic patterns and relationships between data from a large amount of training data. Even in the case of complex and variable defect characteristics, it can accurately make predictions, has good generalization ability, and can adapt to different types of defects and different monitoring environments. For a printer system with sufficient computing resources, this method can be adopted.

[0107] The specific steps in "Step S2" are as follows:

[0108] S21. In-situ monitoring and data collection of thermal scanning tomography, which is the same as the previous step;

[0109] S22. X-ray computed tomography or metallographic analysis;

[0110] S221. X-ray computed tomography: Place the printed sample that has collected thermal scanning tomography data during the printing process in the scanning area of the X-ray computed tomography device. The tomography device emits X-rays and receives the X-ray signals passing through the target object to obtain two-dimensional projection images of the target object. By rotating the target object or the device, multiple projection images are obtained from different angles, and finally, slice images of the target object are generated through an image reconstruction algorithm.

[0111] S222. Metallographic analysis: Cut, grind, and polish the printed sample that has collected thermal scanning tomography data during the printing process to prepare a metallographic sample with a smooth and flat surface. Then, observe the microstructure image under a metallurgical microscope after grinding and polishing layer by layer.

[0112] S23. Data preprocessing and data alignment: First, perform data cleaning, that is, clean the collected thermal image data and X-ray or metallographic image data. The main purpose of this step is to remove invalid data and interference data to ensure the accuracy and reliability of the data. Then, perform data alignment to align the thermal scanning tomography image data with the X-ray or metallographic image data to make them consistent in spatial position. Adopt the feature point matching method, identify the feature points in the images, calculate the transformation parameters (such as translation, rotation, and scaling) between the images, and achieve precise alignment of the images to provide a consistent data basis for subsequent feature extraction and model training.

[0113] For the convenience of subsequent processing and analysis, the thermal image data and X-ray or metallographic image data are further converted into the same pixel format and data type to ensure data consistency and operability;

[0114] S24. Construction of tomography images and solidification defect datasets: For the correlation annotation of thermal scanning tomography data and solidification defects, first, clarify the categories of solidification defects to be annotated, including insufficient powder supply, excessive warping, pores in microdefects, unfused pores, and microcracks in macroscopic manufacturing defects; for the above defects found in X-ray or metallographic image data, use annotation tools such as labelme and VGG Image Annotator to perform data annotation at the corresponding positions in the thermal scanning tomography image data; the specific operations are as follows: Import the X-ray tomography images or metallographic image data into the annotation tool, and the annotator, based on the image features and professional knowledge, annotate the discovered solidification defects, record the type, location, and size information of the defects, and then, in the thermal scanning tomography image data, find the position corresponding to the defect in the X-ray or metallographic image and use the annotation tool to perform annotation at this position to ensure the accuracy and consistency of the annotation;

[0115] S25. Model construction, training, and validation: After the dataset is prepared, carry out the construction, training, and validation of the supervised learning model; it includes a total of 5 steps: model architecture design, dataset preparation, model training, model validation and evaluation, and cross-validation:

[0116] S251. Model architecture design: Select a deep learning model architecture suitable for image data, such as a convolutional neural network (CNN), and design the input layer, convolutional layer, pooling layer, fully connected layer, and output layer to ensure that the model can effectively extract image features and perform defect classification;

[0117] Input layer: Design the parameters of the input layer according to the size and number of channels of the thermal scanning tomography image (the number of channels of a grayscale image is 1, and the number of channels of a color image is 3);

[0118] For example, if the image size is 2560×2560 pixels, the parameters of the input layer can be set to (2560, 2560, 3) to receive the input image data.

[0119] Convolutional layer: Design multiple convolutional layers for extracting local features of the image. Each convolutional layer contains several convolutional kernels, and the size (such as 3×3, 5×5) and number of convolutional kernels are set according to requirements; the convolution operation is performed by sliding the convolutional kernel on the image to extract edge and texture features in the image; add an activation function, such as ReLU (Rectified Linear Unit), after the convolutional layer to introduce non-linearity and enable the model to learn more complex feature representations;

[0120] The ReLU function sets values less than 0 to 0 and keeps values greater than 0 unchanged, having the advantages of simple calculation and fast convergence speed;

[0121] Pooling layer: A pooling layer is added after the convolutional layer to reduce the spatial dimension of features, reduce the amount of computation and the number of parameters, while retaining important feature information; Common pooling operations include max pooling and average pooling; Max pooling takes the maximum value within the pooling area, which can highlight significant features; Average pooling takes the average value within the pooling area, which can smooth features. The window size and stride of the pooling layer are set according to requirements. For example, a window size of 2×2 and a stride of 2 can halve the size of the feature map;

[0122] Fully connected layer: After passing through multiple convolutional layers and pooling layers, the feature map is flattened into a one-dimensional vector and input into the fully connected layer; The neurons in the fully connected layer are connected to each neuron in the previous layer to integrate global feature information; Multiple fully connected layers are designed, and the number of neurons in each layer is set according to requirements;

[0123] For example, a relatively large number of neurons (such as 512) can be set first, and then gradually reduced until the output layer.

[0124] Output layer: The number of neurons in the output layer is the same as the number of categories of defect classification;

[0125] The categories of fused defects include insufficient powder supply, excessive warping in macroscopic manufacturing defects, and pores, unfused pores, and microcracks in microscopic defects, a total of 5 types. Therefore, the number of neurons in the output layer is 5. If there are new defect classification objects subsequently, the corresponding number of neurons will be added accordingly. The softmax activation function is used in the output layer to convert the original values output by the model into a probability distribution;

[0126] The softmax function can normalize the output values to between 0 and 1 and make the sum of the probabilities of all categories equal to 1, thereby obtaining the prediction probability of each category;

[0127] S252. Dataset preparation;

[0128] First is data division and formatting: The labeled thermal scanning tomography image dataset is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1; The training set is used for model training, the validation set is used for hyperparameter tuning and performance evaluation of the model, and the test set is used for the final performance test of the model; To ensure that the image formats in the dataset are unified, all images are converted to the same size and number of channels (such as 2560×2560) so that the model can process them correctly;

[0129] Since collecting abnormal defect data in the online monitoring dataset usually requires a large number of experiments, and defects are sporadic problems, it is not easy to collect the dataset. Therefore, data augmentation operations can be performed on the training set data to increase data diversity and the generalization ability of the model;

[0130] Data augmentation methods include image rotation (such as random rotation from 0 to 360 degrees), scaling (such as random scaling from 0.8 to 1.2 times), flipping (such as horizontal flipping and vertical flipping), translation (such as randomly translating a certain proportion of the image), adding noise (such as Gaussian noise and salt-and-pepper noise), and color transformation (such as adjusting brightness, contrast, and saturation) to simulate the noise and illumination changes that may occur in the actual monitoring process.

[0131] S253. Model training;

[0132] For multi-classification tasks, select the cross-entropy loss function as the loss function of the model. The cross-entropy loss function can measure the difference between the predicted probability distribution of the model and the true label probability distribution, prompting the model to improve prediction accuracy;

[0133] Use the Adam optimization algorithm to update the model parameters;

[0134] The Adam algorithm combines the advantages of the Momentum and RMSProp optimization algorithms, has an adaptive learning rate and good convergence performance. Set the initial learning rate (such as 0.001), learning rate decay strategy (such as decaying to 0.1 times the original every certain number of rounds) parameters to control the learning process of model training. Divide the training set data into multiple batches for training, with each batch containing a certain number of images (such as 32 images); Batch training can improve memory utilization and training efficiency, and at the same time, through the average calculation of gradients, make the update of the model more stable;

[0135] During the training process, monitor the loss value, accuracy metric, training speed, and memory usage of the model in real time. You can use visualization tools (such as TensorBoard) to display the training curve and related metrics in real time, and promptly discover possible problems in the training process, such as overfitting, underfitting, or unstable training;

[0136] S254. Model validation and evaluation;

[0137] Use the validation set data to evaluate the model and calculate the accuracy, recall, and F1-score metrics. Accuracy represents the proportion of the number of samples correctly predicted by the model to the total number of samples;

[0138] Recall represents the proportion of the number of samples correctly predicted as positive by the model to the number of actual positive samples;

[0139] The F1 score is the harmonic mean of precision and recall, comprehensively considering the accuracy and integrity of the model.

[0140] Draw a confusion matrix to show the prediction results of the model on different classes. The rows of the confusion matrix represent the true classes, and the columns represent the predicted classes. Through the confusion matrix, it can be intuitively seen which classes the model performs well on and which classes have misclassification situations.

[0141] According to the validation results, adjust the hyperparameters of the model to improve the performance of the model. For example, the learning rate (such as increasing or decreasing), batch size (such as increasing or decreasing), and regularization parameters (such as L1 regularization and L2 regularization) can be adjusted to find a better combination of hyperparameters. During the validation process, when the model reaches better performance, save the parameters and structure of the model for subsequent testing and application. When needed, the saved model can be loaded to continue training or make predictions.

[0142] S255, Cross-validation;

[0143] Adopt the method of K-fold cross-validation to divide the dataset into K subsets (such as K = 5 or K = 10), and each subset tries to maintain the consistency of data distribution. Conduct K times of training and validation. Each time, select one subset as the validation set, and the remaining K - 1 subsets as the training set. Repeat K times so that each subset has the opportunity to be the validation set. Summarize the results of K times of training and validation, and calculate the average accuracy, recall, F1 score metrics, and standard deviation statistics to obtain a more reliable model performance evaluation result.

[0144] Through cross-validation, the stability and generalization ability of the model on different data subsets can be evaluated, avoiding evaluation biases caused by the contingency of data division.

[0145] S26, Model deployment and application;

[0146] To achieve the purpose of online identification of fused defects, the trained model needs to be optimized before deployment, such as weight quantization and pruning, to reduce the number of model parameters and computational complexity and improve the running efficiency of the model. Weight quantization is to convert the weight parameters in the model from floating-point numbers to fixed-point numbers with low bit widths, reducing the storage space and computational complexity of the model. Pruning is to use structured pruning or unstructured pruning methods to remove unimportant neuron connections or weights in the model. Structured pruning is carried out in units of channels or neurons, and unstructured pruning is carried out in units of individual weights. The basis for pruning can be the absolute value size of the weights, gradient information, or Hessian matrix.

[0147] After pruning, the model is retrained to recover the performance loss caused by pruning. When retraining, a smaller learning rate and a shorter training cycle can be used, and the focus is on optimizing the remaining neuron connections and weights. In addition, knowledge distillation can be used for model optimization. By training a small model to imitate the output or intermediate feature representation of a large model, the number of model parameters and computational complexity can be reduced while maintaining high performance.

[0148] After obtaining an optimized model, the model can be deployed on a server or the terminal of an LPBF system. The computing configuration requirements for the corresponding server and LPBF system terminal include appropriate CPU, GPU, as well as memory and hard disk requirements. For example, if using an IPC terminal, the CPU should be selected as an Intel 14th generation i5 or higher specification CPU. If using a server system, an Intel Xeon series or AMD EPYC series CPU, or equivalent specifications, should be selected. The GPU uses NVIDIA's RTX series, Tesla series, or Quadro series, with a video memory size of not less than 6GB. The memory should be selected to meet the requirements for data storage and exchange during model operation, choosing a memory of more than 16GB, and preferably 32GB or 64GB of memory. The hard disk selects a high-speed solid-state drive (SSD) as the model storage disk to shorten the model loading and data processing time and improve the execution efficiency of the online monitoring algorithm.

[0149] The above process is based on the deployment requirements of IPC or servers. The fused defect detection neural network model can also be deployed on an embedded system, choosing a low-power and high-performance embedded processor, such as the ARM Cortex series or NVIDIA Jetson series. These processors can provide sufficient computing power while maintaining low power consumption.

[0150] Since the data for defect recognition is image data, an embedded system suitable for image processing and real-time inference needs to be selected. Choose a System on Chip (SoC) processor integrated with a GPU, such as NVIDIA Jetson TX2 or Xavier. The memory should be not less than 8GB, and using an eMMC or microSD card as the storage medium for the embedded system can meet the model loading speed.

[0151] S3. Unsupervised learning model;

[0152] S31. Thermal scanning tomography in-situ monitoring and data collection: The same as this step described above;

[0153] S32. X-ray computed tomography or metallographic analysis: The same as this step described above;

[0154] S33. Data Preprocessing and Feature Extraction II: First, eliminate invalid data, such as invalid data caused by disconnection and monitoring device failures. Then, perform standardization processing on the original image data, and then perform normalization processing on the thermal scanning tomography image data to unify the scales and distributions of different images. Through standardization or zero-mean normalization, map the data ranges of all images to the same numerical range (such as between 0 and 1) to avoid the impact of data differences on subsequent analysis. After completing data preprocessing, extract the thermal distribution features of the thermal scanning tomography images, including temperature gradient: calculate the temperature gradient in the image and identify the temperature change regions around the formed cross-section;

[0155] Texture features: Use texture analysis methods (such as gray-level co-occurrence matrix and LBP) to extract the texture features of the image, reflecting the temperature distribution pattern of the formed cross-section;

[0156] Statistical features: Extract statistical features such as the average temperature, standard deviation, and skewness of the thermal scanning tomography images;

[0157] Shape features: Extract the shape information in the thermal distribution image through edge detection algorithms and further analyze the geometric features of the fused cross-section;

[0158] Frequency domain features: Apply Fourier transform and frequency domain analysis methods to extract the spectral features of the image to help distinguish different thermal distribution patterns.

[0159] After feature extraction, use principal component analysis (PCA) and t-SNE dimensionality reduction methods to map the extracted high-dimensional feature space to a low-dimensional space, reduce the dimensionality of the feature space, and improve the efficiency and effect of subsequent clustering algorithms.

[0160] S34. Clustering and Model Training;

[0161] S341. Model Selection: Select suitable unsupervised learning clustering algorithms; such as K-means, DBSCAN, and Gaussian Mixture Models (GMM). These methods can cluster the thermal scanning data according to the features of the images and automatically identify the patterns in the data;

[0162] K-means: Minimize the distance from the samples to the cluster centers and is suitable for scenarios with relatively uniform data distributions;

[0163] L-DBSCAN: A density-based clustering method that can handle noisy data and is suitable for discovering clusters with irregular shapes;

[0164] M-Gaussian Mixture Models: Assume that the data distribution is a mixture of multiple Gaussian distributions and can handle different cluster morphologies.

[0165] S342. Clustering model training: Use the collected thermal scanning tomography image data and the extracted features to train a clustering model. The model divides the data into multiple categories by analyzing the similarity of image features, and each category represents a different thermal distribution pattern, representing normal thermal distribution and the occurrence patterns of different types of fusion defects respectively.

[0166] Compare the abnormal regions detected by the above method with the actual defect regions obtained by X-ray tomography or metallographic inspection, evaluate the accuracy and robustness of the above detection algorithm, and then determine the clustering parameters of the unsupervised clustering model.

[0167] S35. Model deployment and application: Deploy the trained clustering model to the actual LPBF production system; IPC, servers, and embedded systems can be used for real-time data processing. After each new thermal scanning tomography image data is obtained, data preprocessing, feature extraction are automatically performed and input into the trained clustering model to determine whether there is an abnormality in the current thermal distribution and timely detect fusion defects. The required IPC, server, or embedded computing system is the same as that for supervised learning.

[0168] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying melting defects based on thermal scanning tomography images, characterized in that: The following steps are involved: S1, image outlier detection algorithm; The specific steps in "step S1" are as follows; S11, thermal scanning tomography in situ monitoring and data collection; S12, data preprocessing and feature extraction 1; S13, statistical analysis or feature matching; S14, Result evaluation and threshold setting; S15. Deploy applications; S2, supervised learning model; The specific steps in "step S2" are as follows; S21, thermal scanning tomography in situ monitoring and data collection; S22, X-ray computed tomography or metallographic analysis; S23, data preprocessing and data alignment; S24, construction of tomographic image and melting defect dataset; S25, model construction and training verification; S26, Model deployment and application; S3, unsupervised learning model; The specific steps in "step S3" are as follows; S31, thermal scanning tomography in-situ monitoring and data collection: the same as the above step; S32, X-ray computed tomography or metallographic analysis: same as the above step; S33, data preprocessing and feature extraction 2, the specific features in feature extraction are: texture features, statistical features, shape features and frequency domain features; S34, clustering and model training; S35. Model deployment and application.

2. The method for identifying melting defects based on thermal scanning tomography images according to claim 1, characterized in that: In "step S12", the data preprocessing and feature extraction includes the following steps: S121, grayscale processing: converting the thermal scanning tomography image into a grayscale image to reduce the amount of data and calculation complexity; S122, noise removal; S123, image enhancement and size normalization; S124. Feature extraction.

3. The method for identifying melting defects based on thermal scanning tomography images according to claim 1, characterized in that: In "step S13", the statistical analysis or feature matching includes the following steps: S131, Statistical analysis; S132. Feature matching.

4. The method for identifying melting defects based on thermal scanning tomography images according to claim 1, characterized in that: In "step S22", the X-ray computed tomography or metallographic analysis includes the following steps: S221, X-ray computed tomography; S222. Metallographic analysis.

5. The method for identifying melting defects based on thermal scanning tomography images according to claim 1, characterized in that: In "step S25", the model construction and training verification include the following steps: S251, Model architecture design: Select a deep learning model architecture suitable for image data, such as convolutional neural network (CNN), design the input layer, convolution layer, pooling layer, fully connected layer and output layer to ensure that the model can effectively extract image features and classify defects; S252, data set preparation; S253, model training; S254, Model Validation and Evaluation; S255, cross validation.

6. The method for identifying melting defects based on thermal scanning tomography images according to claim 1, characterized in that: In "step S34", the clustering and model training includes the following steps: S341, Model selection; S342. Clustering model training.

7. The method for identifying melting defects based on thermal scanning tomography images according to claim 1, characterized in that: In "step S11", the thermal scanning tomography in-situ monitoring and data collection adopts a thermal scanning tomography online monitoring device, which includes a laser (1), a laser beam (2), a scanning device (3), a part molding surface (4), a molding platform (5), a layer cross-section radiation (6), an observation window (7), an optical component (8), a thermal scanning tomography sensor (9) and a molding cavity top plate (10); During the laser powder bed melting process, the laser (1) emits a laser beam (2), which heats and melts the metal powder on the part forming surface (4) through a scanning device (3). During the heating, melting and cooling and solidification process, layer cross-section radiation (6) is formed on the layer cross-section. The layer cross-section radiation (6) is transmitted to an observation window (7) located on the upper top plate (10) of the forming cavity through spatial light, and then transmitted to a tomography detection sensor (9) through an optical component (8). The tomography detection sensor (9) collects the radiation and forms a thermal scanning tomography image layer by layer. During the printing and processing of the part, the thermal scanning tomography image data is collected layer by layer.

Images