Metal additive manufacturing part defect identification method and device and electronic equipment

Abstract

The invention relates to the technical field of data processing, in particular to a metal additive manufacturing part defect recognition method which comprises the steps that spectral data collection is conducted on a to-be-detected target metal part, and target spectral data are obtained; determining a plurality of target characteristic spectral lines based on the importance of each characteristic spectral line in the target spectral data; based on the target characteristic spectral lines, a target classification model is obtained through pre-training, defect recognition is conducted on the target metal part, and a defect recognition result is obtained. According to the defect identification method, the laser-induced breakdown spectroscopy (LIBS) technology is used for locally and instantly heating the surface of a material to high temperature through laser pulses, so that substances are broken down and plasmas are generated. By analyzing the spectrum emitted by the plasma, qualitative and quantitative information of various elements in the material can be obtained. Compared with other traditional detection methods, the LIBS has the advantages of high detection speed, no need of sample preparation, capability of simultaneously detecting multiple elements and the like.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, device and electronic equipment for identifying defects in metal additive manufacturing parts. Background Art

[0002] In the past few decades, with the rapid development of metal additive manufacturing (AM) technology, component defect detection methods have also made significant progress. Most of these advances rely on the application of traditional metal defect detection technologies, such as computed tomography (CT), eddy current testing, and ultrasonic testing. However, due to the particularity of materials and the complexity of processes in the metal AM process, the problem of defects in AM parts is still one of the main challenges restricting its widespread application and development. AM is different from traditional manufacturing methods. It manufactures parts by stacking materials layer by layer. This unique manufacturing process may produce defects such as microcracks, holes, and unfused areas locally, thereby affecting the performance and quality of the final product. The quality defects of metal AM components may lead to a decrease in the strength of the components, a decrease in corrosion resistance, and even structural failures during use, thereby causing safety accidents. Therefore, effective defect detection and evaluation is a key link to ensure the reliable application of metal AM technology.

[0003] In order to ensure the quality and performance of metal AM parts, they must be accurately inspected for defects during the production process. At present, common metal additive manufacturing defect detection methods can be divided into two categories: non-destructive testing technology and destructive testing technology. Non-destructive testing technologies include ultrasonic testing, X-ray testing, eddy current testing, etc. These technologies can detect possible internal defects or surface defects without destroying the parts. Ultrasonic testing uses high-frequency sound wave transmission to identify cracks and pores inside the material; X-ray testing uses the transmission characteristics of X-rays to reveal structural problems inside the material; eddy current testing is mainly used to identify surface defects and is suitable for defect detection of conductive materials. However, these traditional methods usually have long detection times, complex operating procedures, and strict environmental requirements, which limits their efficiency in practical applications. Summary of the invention

[0004] To solve the above problems, the present invention proposes a defect identification method for metal additive manufacturing parts, which uses LIBS technology to heat the surface of the material locally to a high temperature through laser pulses, so that the material breaks down and generates plasma. By analyzing the spectrum emitted by the plasma, qualitative and quantitative information of various elements in the material can be obtained. Compared with other traditional detection methods, LIBS has the advantages of fast detection speed, no need to prepare samples, and simultaneous detection of multiple elements. These advantages make LIBS have broad application prospects in metallurgy, environmental monitoring, medicine and other fields.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] In the first technical solution, a method for identifying defects in metal additive manufacturing parts includes:

[0007] Collecting spectral data of the target metal parts to be detected to obtain target spectral data;

[0008] Determining a plurality of target characteristic spectral lines based on the importance of each characteristic spectral line in the target spectral data;

[0009] Based on each of the target characteristic spectral lines, a target classification model is obtained by using pre-training, and defects of the target metal component are identified to obtain a defect identification result.

[0010] Preferably, the method further comprises: determining the importance of each characteristic spectral line, specifically comprising:

[0011] Collecting spectrum data for each sample metal component to obtain a number of sample spectrum data;

[0012] Based on each decision tree in the random forest model, the first error corresponding to each characteristic spectral line is calculated using the corresponding out-of-bag data;

[0013] Randomly add noise to the characteristic spectral lines of all out-of-bag data, and recalculate the second error corresponding to each characteristic spectral line;

[0014] The importance of each characteristic spectral line is determined based on the first error and the second error corresponding to the same characteristic spectral line.

[0015] Preferably, the target classification model is a k-nearest neighbor model; the training process of the k-nearest neighbor model is:

[0016] Dividing the spectral data of each sample into a training set and a test set according to a predetermined ratio;

[0017] Based on the predetermined target metal element, the sample characteristic spectral lines of each sample spectral data in the training set are screened to obtain the initial sample characteristic spectral lines;

[0018] Based on the importance of each initial sample characteristic spectral line, determining a number of target sample characteristic spectral lines;

[0019] Based on the characteristic spectral lines of each target sample and the defect labels corresponding to the characteristic spectral lines of each target sample, the initial k-nearest neighbor model is trained to obtain the current k-nearest neighbor model;

[0020] The current k-nearest neighbor model is tested based on each sample spectral data in the test set to obtain a target k-nearest neighbor model.

[0021] Preferably, the target classification model is a back propagation neural network model; the training process of the back propagation neural network model is:

[0022] Dividing the spectral data of each sample into a training set and a test set according to a predetermined ratio;

[0023] Based on the predetermined target metal element, the sample characteristic spectral lines of each sample spectral data in the training set are screened to obtain the initial sample characteristic spectral lines;

[0024] Based on the importance of each initial sample characteristic spectral line, determining a number of target sample characteristic spectral lines;

[0025] Based on the characteristic spectral lines of each target sample and the defect labels corresponding to the characteristic spectral lines of each target sample, the initial back-propagation neural network model is trained to obtain the current back-propagation neural network model;

[0026] The current back-propagation neural network model is tested based on each sample spectral data in the test set to obtain a target back-propagation neural network model.

[0027] Preferably, before screening the sample characteristic spectral lines of each sample spectral data in the training set based on the predetermined target metal element, the method further comprises:

[0028] The spectrum data of each sample is normalized to obtain the normalized spectrum data of each sample.

[0029] Preferably, the target back-propagation neural network model includes an input layer, a hidden layer, and an output layer.

[0030] Preferably, the target metal element includes any one or more of the following: iron (Fe), chromium (Cr), manganese (Mn) and titanium (Ti).

[0031] In the second technical solution, a defect identification device for metal additive manufacturing parts is characterized by comprising:

[0032] An acquisition module is used to acquire spectral data of a target metal part to be detected to obtain target spectral data;

[0033] A determination module, used for determining a plurality of target characteristic spectral lines based on the importance of each characteristic spectral line in the target spectral data;

[0034] The recognition module is used to perform defect recognition on the target metal component based on each of the target characteristic spectral lines and obtain a defect recognition result by using a target classification model obtained by pre-training.

[0035] In a third technical solution, a storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the metal additive manufacturing component defect identification method described in the first technical solution are implemented.

[0036] In a fourth technical solution, an electronic device includes at least a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program on the memory, the steps of the metal additive manufacturing component defect identification method described in the first technical solution are implemented.

[0037] The beneficial effects of using the present invention are:

[0038] This technical solution uses LIBS combined with the random forest (RF) method for feature extraction, and uses the k-nearest neighbor (KNN) and BP neural network (BPNN) methods for classification to distinguish between defects and defect-free control of metal AM components. Metal AM component samples are placed on a three-dimensional mobile platform for LIBS spectrum acquisition. According to the results of the RF model, appropriate characteristic important spectral lines are selected as input variables for the KNN and BPNN classification models. The recognition performance of the model is evaluated by the accuracy, sensitivity and specificity of the test set. The generalization ability of the model is evaluated by the accuracy, sensitivity and specificity of the validation set. At the same time, this method has the potential to improve the accuracy and efficiency of component quality inspection, in order to provide strong support for the further promotion and application of metal AM technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the LIBS metal AM defect detection experimental platform.

[0040] Figure 2 This is the main algorithm flow of the present invention.

[0041] Figure 3 Figure 2. Difference between defect-free and defective spectra of metal AM parts.

[0042] Figure 4 Confusion matrix of the KNN classification model: (a) training set; (b) test set and (c) validation set.

[0043] Figure 5 Schematic diagram of the characteristic importance of spectral lines.

[0044] Figure 6 Confusion matrix of the RF-KNN classification model: (a) training set, (b) test set, and (c) validation set.

[0045] Figure 7 .Confusion matrix of RF-BPNN classification model: (a) training set, (b) test set, (c) validation set. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the technical solution more clear, the technical solution is further described in detail below in conjunction with specific implementation methods. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the technical solution.

[0047] The main instruments used in this example are Nd:YAG laser (wavelength 1064nm, energy 60mJ, pulse width 8ns, repetition rate 1Hz, wavelength diameter 6mm, beam mode TEM00), ICCD detector, spectrometer (Andor ME5000), digital delay generator (BNC-575) and 3D moving platform. Figure 1 As shown in the figure, in the experiment, the laser is controlled by a digital delay generator to emit a laser signal. The signal is reflected and focused by the mirror and lens, so that the pulsed laser acts on the surface of the metal sample, ablating the sample surface to generate plasma. Figure 1 As shown in the figure, in the experiment, the laser is controlled by a digital delay generator to emit a laser signal. The signal is reflected and focused by the mirror and lens, so that the pulsed laser acts on the surface of the metal sample, ablating the sample surface to generate plasma. During the acquisition delay time, the photon signal radiated by the plasma is collected by the spectrometer, the optical signal is converted into an electrical signal by the ICCD detector, and the spectral data is displayed and stored in the host. In order to avoid the repeatability of the acquisition, a three-dimensional mobile platform is used to control the sample to move at a constant speed. The acquisition delay time is set to 1s and the integration time is set to 1ms, so that a spectrum signal with a high signal-to-noise ratio can be obtained. The experiment was carried out in an air environment with a temperature of 23°C and a relative humidity of 25%.

[0048] The metal AM samples used in the experiment were prepared by directed energy deposition (DED) method, and their elemental composition is shown in Table 1. A total of 100 samples were selected, including 50 defective and non-defective samples. 40 defective and 40 substandard samples were selected, recorded as samples 1#-80#. Three spectra were collected for each sample, and the total number of spectra was 240. The 240 spectra were randomly divided into training set and test set in a ratio of 7:3. At the same time, the spectra of the remaining 20 samples were collected by the same method, recorded as samples 81#-100#, and 60 spectra were obtained as validation sets to verify the generalization ability of the model. At the same time, in order to eliminate the influence of interference factors, the control variable method was used for the experiment. The experimental instrument parameters remained unchanged, and the metal AM samples used the same process flow and the same raw materials. Therefore, the variables in this experiment were only defective and non-defective samples. In order to reduce the interference of external factors such as laser energy fluctuations and experimental environment, when selecting defective areas and non-defective areas for spectrum collection, each spectrum was obtained by accumulating 5 pulses of laser.

[0049] Table 1. Elements and composition of ferrochrome alloys (wt.%)

[0050]

[0051] Related algorithms designed by the present invention. First, random forest (RF) is an integrated learning method that can be used to evaluate the importance of features. In this experiment, RF is used to select the spectral data of metal AM, and select the corresponding characteristic spectral lines according to the importance of the features. The metal AM sample data set is set to T, where each spectrum has N characteristic spectral lines, and a random forest model of M decision trees is established. For each decision tree in the random forest, the error is calculated using the corresponding out-of-bag data (OOB), denoted as OOBer1i. Noise interference is randomly added to the feature X of all OOB samples, and the OOB error is recalculated, denoted as OOBer2i. Assuming that the feature importance of each characteristic spectral line is FIn, the following formula is available:

[0052]

[0053] k-Nearest Neighbor (KNN) is a relatively simple machine learning algorithm. It has a wide range of applications and has received extensive attention and research over the years. The classification idea of ​​this algorithm is similar to that of metal AM defect identification. With defect-free spectral data as the center, the fluctuation range is defined to distinguish metal AM defects from defect-free samples.

[0054] Back Propagation Neural Network (BPNN) is a multi-layer feedforward network trained using the error back propagation algorithm. It is one of the most widely used neural network models. BPNN uses the steepest descent learning rule to continuously adjust the network's weights and thresholds through back propagation to reduce the network's classification error rate.

[0055] In this embodiment, the KNN, RF-KNN and RF-BPNN models are used to identify defects and non-defects in metal AM parts, considering the high similarity, complexity and high dimensionality of LIBS spectral data. The main process of the classification algorithm is as follows: Figure 2 shown.

[0056] Finally, this example uses accuracy, sensitivity and specificity to evaluate the performance of the classification model. The accuracy, sensitivity and specificity of the model are calculated as follows:

[0057]

[0058]

[0059]

[0060] Among them, TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

[0061] The present invention uses the best spectral acquisition parameters and experimental settings to detect defects and non-defects in metal AM parts. Figure 3 As shown, the normalized spectral difference map of defective and non-defective metal AM components is plotted. The spectral wavelength range of metal AM components is 300-600nm. The normalization method is used to keep the two spectra in the same dimension to better compare the difference between component defect and non-defective spectra. At the same time, each type of spectral data is obtained by averaging 168 spectral data of the same type in the training set to minimize the error caused by chance.

[0062] Since this experiment was conducted in an air environment, it is necessary to exclude elements such as C, H, and O as characteristic spectral lines. According to the elemental composition of metal AM parts, four elements were selected to distinguish defective and non-defective samples of metal AM parts. As shown in Table 2, the four selected metal elements are composed of Fe, Cr, Mn, and Ti, respectively, and 29 spectral lines were selected in the range of 300 to 600 nm. Among them, cyan, red, purple, and green represent the characteristic spectral lines of Fe, Cr, Mn, and Ti elements, respectively. The basic principles for spectral line selection are as follows: (1) According to the data from the National Institute of Standards and Technology (NIST) database, the selected spectral line drift must be within 0.1 nm. (2) Self-absorption and self-inversion spectral lines should not be selected.

[0063] Table 2. Classifier input variable analysis rows

[0064]

[0065] In the present invention, when the parameters such as pulse laser energy and defocus amount are constant, the defect-free spectral data of metal AM parts has some very obvious characteristics. Compared with defective samples, the surface of samples without defective components is more regular. When the pulse laser acts on the surface of defect-free samples, the laser-induced plasma is more stable, thereby obtaining more stable spectral data. At this time, according to the classification principle of the KNN model, the KNN model sets the K value so that the distance between the calculated defect-free spectral data points is kept within a certain range. In this case, there is almost no problem of identifying the defect-free spectral data points of the component as defective data points. Therefore, according to the characteristics of the defect-free spectral data of metal AM parts, the KNN model is used to distinguish the defective and defect-free spectra of metal parts, and better classification performance can be obtained.

[0066] The KNN model was established, and 240 spectra were randomly divided into training set and test set in a ratio of 7:3. 60 spectra in the validation set were used for validation. By optimizing the K value, when K = 3, the accuracy was the highest and the best performance was obtained. The confusion matrix of the training set, test set and validation set of the KNN model was plotted, as shown in Figure 4 As shown. The classification results of the training set, test set and validation set of the KNN model are compared, as shown in Table 3. The results show that the classification accuracy of the training set, test set and validation set are 95.833%, 91.667% and 91.667%, respectively, all higher than 90%. Compared with the test set and validation set, the accuracy of the training set is increased by nearly 4.2%. This is because the KNN model is affected by spectral noise when calculating the distance of sample points, and the defective sample points are divided into the category of non-defective samples. At the same time, it can be found from the classification results that the sensitivity of the training set, test set and validation set is 100%, and all non-defective samples can be identified. The specificity of the training set, test set and validation set of the KNN model is 91.765%, 82.857% and 83.333%, respectively. The specificity is low, and defective samples are easily misjudged as non-defective samples. In order to improve the specificity of the model, the spectral line selection method is used to remove the spectral lines that have little effect on the classification, avoid the interference of irrelevant spectral lines on the classification of the KNN model, and improve the classification efficiency.

[0067] Table 3. Comparison of accuracy, sensitivity and specificity of KNN models

[0068]

[0069] In order to further optimize the selection of spectral lines, the RF model is used to reduce the number of spectral lines to 8, thereby reducing the complexity of the classification model. RF evaluates the importance of features by adding noise to the selected features, and selects characteristic spectral lines with importance greater than the average as input variables of the KNN model. According to the importance of the characteristic spectral lines, 8 spectral line features are selected as input variables of the KNN classifier. The variable importance of 29 spectral lines is as follows: Figure 5 Shown

[0070] The RF-KNN model is established, and the features of RF spectrum selection are used as input variables. The data set is divided in the same way as the KNN model. When K = 3, the performance of this model is the best. The confusion matrix of the training set, test set, and validation set of the RF-KNN model is plotted, as shown in Figure 6As shown in Table 4, the classification results of the training set, test set and validation set of the RF-KNN model are compared. The results show that the classification accuracy of the training set, test set and validation set of the RF-KNN model is 97.023%, 95.833% and 95%, respectively. Compared with the KNN model, the accuracy of the test set of the RF-KNN model is improved by 1.19%, and the accuracy of the test set and validation set is improved by 3% to 4%, indicating that the RF-KNN model has better generalization ability and robustness. Although the specificity and AUC of the RF-KNN model have also improved to varying degrees, the specificity of the model is still low compared with the accuracy and sensitivity, making the overall performance of the model poor. At the same time, due to the classification of the two defect-free spectra as defective spectra, the sensitivity of the training set in the RF-KNN model is lower than that of the KNN model. This shows that as the dimension decreases, the fluctuation range of the RF-KNN division becomes more and more accurate, and the distance between two points of different categories becomes smaller and smaller, resulting in misjudgment.

[0071] Table 4. Comparison of accuracy, sensitivity and specificity of RF-KNN models

[0072]

[0073] Due to the limitations of the RF-KNN model, this embodiment uses the RF-BPNN model as a classifier. Compared with RF-KNN, the RF-BPNN model is less affected by the dimension and is more in line with the actual situation that defective products are rare categories in production. At the same time, the BPNN model has a strong nonlinear mapping capability and a multi-layer structure, which enables it to effectively capture complex data features and optimize classification results.

[0074] For the RF-BPNN model, the 8 spectral line features selected by the RF spectrum line are used as input variables. In this model, the input layer consists of 1 input layer, which contains 8 nodes, and each node corresponds to a spectral feature. Next, the model has 1 hidden layer, which contains 8 nodes, which is used to learn the relationship between the input features. Finally, the output layer consists of 1 output layer, which contains 1 node, is responsible for outputting the classification results, and uses the Sigmoid activation function to meet the needs of the binary classification problem. The data set is divided in the same way as the KNN model. The accuracy, sensitivity and specificity are used as the final classification results of the RF-BPNN model. The confusion matrix of the RF-BPNN model is as follows: Figure 7The classification results of the training set, test set and validation set of the RF-BPNN model are compared, as shown in Table 5. The results show that the classification accuracy of the training set, test set and validation set of the RF-BPNN model is 99.4%, 97.2% and 96.67%, respectively. Compared with the KNN model and the RF-KNN model, the accuracy of the RF-BPNN model is greatly improved, and the sensitivity, specificity and AUC are also maintained at a high level. The evaluation indicators of the RF-BPNN model are relatively balanced, and the spectral data of the training set, test set and validation set are less misjudged, and the comprehensive ability is good.

[0075] Table 5. Comparison of accuracy, sensitivity and specificity of RF-BPNN models

[0076]

[0077] In order to better evaluate the quality of the model, the receiver operating characteristic (ROC) curve was used to evaluate the model. The performance of the model gradually improved in the order of KNN, RF-KNN, and RF-BPNN models on the training set, test set, and validation set. The RF-BPNN model had the best effect, with the AUC area of ​​0.9938, 0.9719, and 0.9667 on the training set, test set, and validation set, respectively.

[0078] In this study, RF spectrum line selection and BPNN classifier can accurately and effectively distinguish defective and non-defective samples of metal AM parts. First, the non-defective and defective spectral data of metal AM parts are obtained, and the main elements and corresponding spectral lines are determined with the help of NIST database; then, the KNN model similar to the classification idea of ​​non-defective and defective metal AM parts is used to distinguish the two types of spectral data. The results show that the specificity performance of the KNN model on the three data sets is low. In order to improve the specificity of the KNN model, the RF method is used to select 8 spectral lines with feature importance greater than the average value to avoid the interference of irrelevant spectral lines in the classification of the KNN model. The results show that the specificity and AUC of the three data sets are improved to varying degrees. However, compared with the accuracy and sensitivity of the respective data sets, the specificity of the model is still low, indicating that the comprehensive performance of the model is poor. In order to improve the specificity of defect data recognition, RF-BPNN is used to classify and identify metal non-defective and defective data. The results show that the classification accuracy of the training set, test set and validation set of the RF-BPNN model is 99.4%, 97.2% and 96.67%, respectively. The accuracy was greatly improved, and the sensitivity, specificity and AUC were also maintained at a high level. The evaluation indicators were relatively balanced and the comprehensive ability was good. At the same time, the ROC curve was used to verify the three models.

[0079] Compared with traditional nondestructive testing methods, Laser-Induced Breakdown Spectroscopy (LIBS), as an emerging nondestructive multi-element analysis technology, has shown great application potential. LIBS technology uses laser pulses to instantly heat the surface of a material to a high temperature, causing the material to break down and generate plasma. By analyzing the spectrum emitted by the plasma, qualitative and quantitative information about various elements in the material can be obtained. Compared with other traditional detection methods, LIBS has the advantages of fast detection speed, no need to prepare samples, and the ability to detect multiple elements at the same time. These advantages make LIBS have broad application prospects in metallurgy, environmental monitoring, medicine and other fields.

[0080] In this study, LIBS combined with random forest (RF) method was used for feature extraction, and k-nearest neighbor (KNN) and BP neural network (BPNN) methods were used for classification to distinguish defects from defect-free control of metal AM components. Metal AM component samples were placed on a three-dimensional mobile platform for LIBS spectrum acquisition. According to the results of the RF model, appropriate characteristic important spectral lines were selected as input variables for the KNN and BPNN classification models. The recognition performance of the model was evaluated by the accuracy, sensitivity and specificity of the test set. The generalization ability of the model was evaluated by the accuracy, sensitivity and specificity of the validation set. At the same time, the potential of this method to improve the accuracy and efficiency of component quality inspection was explored, in order to provide strong support for the further promotion and application of metal AM technology.

[0081] In addition, this embodiment also proposes a defect recognition device for metal additive manufacturing parts, including: an acquisition module, used to collect spectral data for the target metal part to be detected to obtain target spectral data; a determination module, used to determine a number of target characteristic spectral lines based on the importance of each characteristic spectral line in the target spectral data; an identification module, used to obtain a target classification model based on each of the target characteristic spectral lines by pre-training, and perform defect recognition on the target metal part to obtain a defect recognition result. The defect recognition device for metal additive manufacturing parts can execute the above method.

[0082] In addition, this embodiment further proposes a storage medium, characterized in that the storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned metal additive manufacturing component defect identification method are implemented. An electronic device includes at least a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program on the memory, the steps of the above-mentioned metal additive manufacturing component defect identification method are implemented.

[0083] The above contents are only preferred embodiments of the present invention. For ordinary technicians in this field, many changes can be made in the specific implementation methods and application scopes based on the ideas of the technical content. As long as these changes do not deviate from the concept of the present invention, they all fall within the scope of protection of this patent.

Claims

1. A method for identifying defects in metal additive manufacturing parts, characterized in that: include: Collecting spectral data of the target metal parts to be detected to obtain target spectral data; Determining a plurality of target characteristic spectral lines based on the importance of each characteristic spectral line in the target spectral data; Based on each of the target characteristic spectral lines, a target classification model is obtained by using pre-training, and defects of the target metal component are identified to obtain a defect identification result.

2. The metal additive manufacturing component defect identification method according to claim 1, characterized in that: The method further includes: determining the importance of each characteristic spectral line, specifically including: Collecting spectrum data for each sample metal component to obtain a number of sample spectrum data; Based on each decision tree in the random forest model, the first error corresponding to each characteristic spectral line is calculated using the corresponding out-of-bag data; Randomly add noise to the characteristic spectral lines of all out-of-bag data, and recalculate the second error corresponding to each characteristic spectral line; The importance of each characteristic spectral line is determined based on the first error and the second error corresponding to the same characteristic spectral line.

3. The metal additive manufacturing component defect identification method according to claim 2, characterized in that: The target classification model is a k-nearest neighbor model; the training process of the k-nearest neighbor model is: Dividing the spectral data of each sample into a training set and a test set according to a predetermined ratio; Based on the predetermined target metal element, the sample characteristic spectral lines of each sample spectral data in the training set are screened to obtain the initial sample characteristic spectral lines; Based on the importance of each initial sample characteristic spectral line, determining a number of target sample characteristic spectral lines; Based on the characteristic spectral lines of each target sample and the defect labels corresponding to the characteristic spectral lines of each target sample, the initial k-nearest neighbor model is trained to obtain the current k-nearest neighbor model; The current k-nearest neighbor model is tested based on each sample spectral data in the test set to obtain a target k-nearest neighbor model.

4. The metal additive manufacturing component defect identification method according to claim 2, characterized in that: The target classification model is a back propagation neural network model; the training process of the back propagation neural network model is: Dividing the spectral data of each sample into a training set and a test set according to a predetermined ratio; Based on the predetermined target metal element, the sample characteristic spectral lines of each sample spectral data in the training set are screened to obtain the initial sample characteristic spectral lines; Based on the importance of each initial sample characteristic spectral line, determining a number of target sample characteristic spectral lines; Based on the characteristic spectral lines of each target sample and the defect labels corresponding to the characteristic spectral lines of each target sample, the initial back-propagation neural network model is trained to obtain the current back-propagation neural network model; The current back-propagation neural network model is tested based on each sample spectral data in the test set to obtain a target back-propagation neural network model.

5. The metal additive manufacturing component defect identification method according to claim 3 or 4, characterized in that: Before screening the sample characteristic spectral lines of each sample spectral data in the training set based on the predetermined target metal element, the method further includes: The spectrum data of each sample is normalized to obtain the normalized spectrum data of each sample.

6. The metal additive manufacturing component defect identification method according to claim 4, characterized in that: The target back propagation neural network model includes an input layer, a hidden layer, and an output layer.

7. The metal additive manufacturing component defect identification method according to claim 3 or 4, characterized in that: The target metal elements include any one or more of the following: iron (Fe), chromium (Cr), manganese (Mn) and titanium (Ti).

8. A defect identification device for metal additive manufacturing parts, characterized in that: include: An acquisition module is used to acquire spectral data of a target metal part to be detected to obtain target spectral data; A determination module, used for determining a plurality of target characteristic spectral lines based on the importance of each characteristic spectral line in the target spectral data; The recognition module is used to perform defect recognition on the target metal component based on each of the target characteristic spectral lines and obtain a defect recognition result by using a target classification model obtained by pre-training.

9. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the metal additive manufacturing component defect identification method described in any one of claims 1 to 7 are implemented.

10. An electronic device, characterized in that: The method comprises at least a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program in the memory, the method implements the steps of the metal additive manufacturing component defect identification method according to any one of claims 1 to 7.

Images