Automatic defect identification method and system based on electric arc additive manufacturing mathematical radiographic inspection

Abstract

The invention provides an automatic defect identification method and system based on electric arc additive manufacturing mathematical radiographic inspection. The method belongs to the technical field of industrial nondestructive inspection and artificial intelligence crossing. The method comprises the steps that a dynamic ray absorption model is established based on material microstructure characteristics in the electric arc additive manufacturing process; performing simulation calculation on ray absorption characteristics under different process parameters by using the model to generate theoretical background image data; performing detection area grid division according to the theoretical background image data, and determining a radiographic detection key parameter configuration scheme; by establishing a dynamic ray absorption model and combining real-time ray detection and inversion calculation, key information such as a molten pool temperature field and flow field distribution can be accurately obtained, a reliable basis is provided for defect identification, and the detection accuracy is improved.

Description

Technical Field

[0001] This invention proposes an automatic defect identification method and system based on mathematical X-ray inspection in arc additive manufacturing, which belongs to the interdisciplinary field of industrial non-destructive testing and artificial intelligence. Background Technology

[0002] Wafer Additive Manufacturing (WAAM) is increasingly widely used in high-end fields such as aerospace and energy equipment due to its low cost, high forming efficiency, and ability to manufacture complex components. However, in this process, uneven solidification of the molten pool and fluctuations in process parameters can easily lead to defects such as porosity, lack of fusion, and hot cracks, which directly affect the mechanical properties and service safety of components. Therefore, defect detection is the core link in WAAM quality control.

[0003] Currently, the mainstream detection method is radiographic testing (RT). Traditional RT relies on manual interpretation, which suffers from low efficiency, strong subjectivity, and a high rate of missed detection of microscopic defects. Existing automatic identification technologies also have significant limitations: most models ignore the influence of material microstructure (such as grain size and phase composition) on X-ray absorption characteristics, resulting in poor adaptability of the detection model to actual working conditions; feature extraction uses multi-focus visual images and does not integrate physical behavior data such as molten pool temperature field and flow field, resulting in insufficient defect identification accuracy and cause explanation ability; moreover, the system lacks self-evolution ability, requiring re-collection of samples and retraining of models to deal with new types of defects, resulting in weak adaptability; at the same time, the false alarm rate is high, making it difficult to meet the needs of industrial batch inspection. Therefore, there is an urgent need for an automatic defect identification method based on WAAM that takes into account microscopic characteristic adaptability, multi-modal feature fusion, self-evolution ability, and low false alarm rate. Summary of the Invention

[0004] This invention provides an automatic defect identification method and system based on mathematical X-ray inspection in arc additive manufacturing, to solve the problems mentioned in the background section above:

[0005] The present invention proposes an automatic defect identification method based on mathematical X-ray inspection in arc additive manufacturing, the method comprising:

[0006] S1. Based on the microstructure characteristics of materials in the electric arc additive manufacturing process, a dynamic X-ray absorption model is established; the X-ray absorption characteristics under different process parameters are simulated and calculated using this model to generate theoretical background map data; the detection area is meshed according to the theoretical background map data, and the configuration scheme of key X-ray detection parameters is determined.

[0007] S2. Collect raw X-ray image data of the molten pool area through a real-time X-ray detection system, and perform molten pool behavior inversion calculations in conjunction with a dynamic X-ray absorption model to generate molten pool temperature field and flow field distribution data; calculate the prior probability of defects based on the molten pool behavior data using the Bayesian inference method to obtain the initial occurrence probability distribution data of different types of defects; perform scattering noise suppression processing on the raw X-ray image data to generate preprocessed X-ray image data.

[0008] S3. Construct a coupled inference network architecture that includes physical feature extraction branches and visual feature extraction branches; input preprocessed ray image data into the visual feature branch to extract defect morphology features, and simultaneously input molten pool behavior data and defect prior probability data into the physical feature branch to extract defect causal features; realize spatial alignment and semantic fusion of multimodal features through the feature fusion module to generate fused feature representation data;

[0009] S4. Construct a defect classification decision tree based on the fused feature representation data, and use an interpretable machine learning method to generate a defect type determination rule base; perform recursive matching analysis on the fused features through the rule base to generate preliminary identification result data; construct a false alarm pattern library using historical detection data, and perform false alarm pattern comparison and screening on the preliminary identification results to generate optimized identification result data.

[0010] S5. Establish a self-evolving module that includes a new defect sample learning mechanism to automatically incorporate new defect samples confirmed in the optimized identification results into the training dataset; update the parameters of the coupled inference network through an online incremental learning algorithm to generate an updated intelligent detection model; iteratively process subsequent detection data based on the updated model to form an intelligent detection system for arc additive manufacturing defects with self-evolving capabilities, and output comprehensive detection report data.

[0011] The automatic defect identification system based on mathematical X-ray inspection in arc additive manufacturing proposed in this invention includes:

[0012] One or more processors;

[0013] Memory, used to store one or more programs.

[0014] Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0015] Beneficial effects of this invention:

[0016] By establishing a dynamic X-ray absorption model and combining real-time X-ray detection and inversion calculation, key information such as the temperature field and flow field distribution of the molten pool can be accurately obtained, providing a reliable basis for defect identification and improving detection accuracy. A coupled inference network architecture is constructed, integrating visual and physical features to achieve multimodal data spatial alignment and semantic fusion, comprehensively mining defect features and enhancing defect identification capabilities. An interpretable machine learning method is used to generate a judgment rule base, making the defect identification process transparent, easy to understand and verify, and enhancing the credibility of the results. A false alarm pattern library is constructed using historical detection data to screen preliminary identification results, effectively eliminating false alarms and improving detection reliability. A self-evolving module is established to incorporate new defect samples into the training dataset, updating model parameters through online incremental learning, enabling the system to continuously adapt to new situations and continuously improve detection performance. Finally, a comprehensive inspection report is generated, integrating information in a structured document format for easy user access and analysis, providing strong support for quality control in arc additive manufacturing. Attached Figure Description

[0017] Figure 1 This is a diagram of the method described in this invention;

[0018] Figure 2 This is a schematic diagram of the pores described in this invention;

[0019] Figure 3 This is a diagram showing the location of the pores described in this invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] One embodiment of the present invention, such as Figure 1 As shown, an automatic defect identification method based on mathematical radiography in arc additive manufacturing is described, the method comprising:

[0022] S1. Based on the microstructure characteristics of materials in the arc additive manufacturing process, a dynamic X-ray absorption model considering the influence of microstructure is established; the X-ray absorption characteristics under different process parameters are simulated and calculated using this model to generate theoretical background map data containing material microstructure information; the detection area is meshed according to the theoretical background map data to determine the configuration scheme of key X-ray detection parameters.

[0023] S2. Collect raw X-ray image data of the molten pool area through a real-time X-ray detection system, and perform molten pool behavior inversion calculations in conjunction with a dynamic X-ray absorption model to generate molten pool temperature field and flow field distribution data; calculate the prior probability of defects based on the molten pool behavior data using the Bayesian inference method to obtain the initial occurrence probability distribution data of different types of defects; perform scattering noise suppression processing on the raw X-ray image data to generate preprocessed X-ray image data.

[0024] S3. Construct a coupled inference network architecture that includes physical feature extraction branches and visual feature extraction branches; input preprocessed ray image data into the visual feature branch to extract defect morphology features, and simultaneously input molten pool behavior data and defect prior probability data into the physical feature branch to extract defect causal features; realize spatial alignment and semantic fusion of multimodal features through the feature fusion module to generate fused feature representation data;

[0025] S4. Construct a defect classification decision tree based on the fused feature representation data, and use interpretable machine learning methods to generate a defect type determination rule base; perform recursive matching analysis on the fused features through the rule base to generate preliminary identification result data containing explanations of the physical causes of defects; construct a false alarm pattern library using historical detection data, and perform false alarm pattern comparison and screening on the preliminary identification results to generate optimized identification result data.

[0026] S5. Establish a self-evolving module that includes a new defect sample learning mechanism to automatically incorporate new defect samples confirmed in the optimized identification results into the training dataset; update the parameters of the coupled inference network through an online incremental learning algorithm to generate an updated intelligent detection model; iteratively process subsequent detection data based on the updated model to form an intelligent detection system for arc additive manufacturing defects with self-evolving capabilities, and finally output a comprehensive detection report data including defect location, type, physical cause and confidence level.

[0027] The working principle and effects of the above technical solution are as follows:

[0028] By establishing a dynamic X-ray absorption model that considers the microstructure of materials and combining it with process parameters to simulate and determine key detection parameters, the accuracy of initial parameters for X-ray detection is improved, and detection deviations caused by blind parameter settings are reduced.

[0029] Scattering noise is suppressed on the original X-ray image, and Bayesian inference is used to calculate the prior probability of defects, which reduces the interference of image noise on subsequent identification and improves the accuracy of preliminary defect judgment.

[0030] A dual-branch coupled inference network is used to extract the morphological and causal features of defects respectively. Then, feature fusion is used to align multimodal information, which enhances the comprehensiveness of defect feature extraction and avoids the recognition limitations caused by a single feature dimension.

[0031] Based on decision trees and interpretable machine learning to generate a judgment rule base, and by screening the results through a false alarm pattern library, the accuracy of defect identification has been improved, and false judgments such as molten pool splash and surface scratches have been reduced. At the same time, the interpretability of the physical causes of the identification results has been enhanced, making it easier for staff to understand the root causes of defects.

[0032] By leveraging self-evolutionary modules and online incremental learning to update the model, the detection system's adaptability to novel defects is enhanced, reducing the cost of manually collecting samples and retraining the model. It also improves the efficiency and stability of long-term detection, and the final comprehensive report provides a more comprehensive reference for process optimization.

[0033] In one embodiment of the present invention, S1 includes:

[0034] S11. For the arc additive manufacturing material to be tested (e.g., titanium alloy, aluminum alloy, etc.), the microstructure parameters of the material are collected using acquisition equipment, including a scanning electron microscope (SEM) and X-ray diffraction (XRD); the microstructure parameters include the grain size, phase composition, and pore distribution of the material; a raw dataset of the material's microstructure is generated; outlier removal and data standardization are performed on the raw dataset to obtain a standardized set of microstructure parameters;

[0035] S12. Based on a standardized set of microstructure parameters, and combined with Beer-Lambert's law of interaction between radiation and matter, a microstructure factor is introduced to modify the traditional radiation absorption model, and a dynamic radiation absorption model is constructed. The microstructure factor includes the grain scattering coefficient and the phase interface absorption coefficient.

[0036] S13. The model parameters are fitted and optimized using the least squares method to ensure that the error between the model output and the experimental data of known material ray absorption is less than 5%, and a verified dynamic ray absorption model is generated; the key process parameter ranges of arc additive manufacturing (e.g., current 100-300A, voltage 15-30V, wire feeding speed 5-15mm / s) are determined, and a process parameter combination matrix is ​​constructed.

[0037] S14. Input the parameter combination matrix into the validated dynamic X-ray absorption model to simulate and calculate the distribution of X-ray absorption intensity of the material under different parameters, and generate a X-ray absorption simulation dataset containing material microstructure information; based on this dataset, draw the X-ray detection theoretical background map of the corresponding process parameters to form a theoretical background map dataset.

[0038] S15. Perform grayscale gradient analysis on the images in the theoretical background image dataset to identify areas with significant differences in ray absorption (i.e., areas sensitive to potential defects) and determine key detection areas. Use an adaptive grid partitioning algorithm (e.g., quadtree grid partitioning) to partition the detection area into a grid, generating a weighted detection area grid map (the grid density of key areas is higher than that of non-key areas). Combine the ray absorption characteristics of the grid map with the detection accuracy requirements to optimize the key parameters of the ray source, including intensity, detection distance, and exposure time, and generate a configuration scheme for key ray detection parameters.

[0039] The working principle and effects of the above technical solution are as follows:

[0040] By collecting material micro-parameters through SEM and XRD, and then removing outliers and standardizing the data, the accuracy of the material microstructure data was improved, and the interference of invalid noise in the original data on subsequent modeling was reduced.

[0041] By incorporating the Beer-Lambert law to introduce grain scattering coefficient and phase interface absorption coefficient to correct the traditional model, the adaptability of the X-ray absorption model to the actual microscopic properties of materials is enhanced, and the calculation deviation of the model caused by neglecting microscopic factors is reduced.

[0042] The model parameters were optimized using the least squares method and the error was controlled to be less than 5%, which improved the accuracy of the dynamic ray absorption model. At the same time, the range of process parameters was determined and a combination matrix was constructed, which reduced the blindness of parameter selection in subsequent simulations.

[0043] After inputting the parameter matrix for verification, the model generates a simulation dataset and a theoretical background diagram, which enhances the ability to predict the X-ray absorption characteristics of materials under different processes and reduces the blind spot of not knowing the material absorption situation before actual testing.

[0044] By using grayscale analysis to locate key areas and adaptive mesh generation, and then optimizing ray parameters based on absorption characteristics, the accuracy of detecting key areas is improved, the consumption of ineffective detection resources in non-key areas is reduced, and the problem of poor detection results caused by improper parameter settings is also reduced.

[0045] In one embodiment of the present invention, S2 includes:

[0046] S21. Based on the generated configuration scheme of key parameters for X-ray detection, parameter calibration is performed on the real-time X-ray detection system (e.g., industrial CT, high-energy X-ray detector). The parameter calibration includes X-ray source stability calibration, detector sensitivity calibration, and image resolution calibration. During the arc additive manufacturing process, continuous X-ray imaging is performed on the molten pool area at a sampling frequency of 20-50 frames / second to acquire the original X-ray image sequence of the molten pool area and generate the original X-ray image dataset.

[0047] S22. Extract the grayscale information of the molten pool region from each frame of the original ray image dataset and input it into the dynamic ray absorption model verified in S12. Calculate the energy absorption distribution inside the molten pool using a model inversion algorithm (e.g., the conjugate gradient method), and further derive the spatiotemporal distribution data of the molten pool temperature field (temperature range 500-2500℃) and flow field (flow velocity 0.1-1m / s) to generate a spatiotemporal dataset of molten pool behavior.

[0048] S23. Obtain common defect types in arc additive manufacturing, including porosity (e.g., Figure 2 As shown, a defect type classification system is established for defects such as non-fusion, cracks, and inclusions. The generated spatiotemporal dataset of molten pool behavior (e.g., local temperature below melting point, abnormal flow velocity region) is used as evidence variables and substituted into the Bayesian inference formula to calculate the initial probability of occurrence of each defect under different molten pool behavior conditions.

[0049] S24. Perform spatial interpolation on the probability results to generate heatmaps of the initial occurrence probability distribution of different types of defects, and obtain a dataset of the prior probability distribution of defects.

[0050] S25. Analyze the noise types (mainly Compton scattering noise) in the original X-ray image dataset, and use a combination of adaptive median filtering and wavelet thresholding to suppress noise in the image; perform grayscale stretching and contrast enhancement on the denoised image to highlight the grayscale difference between the molten pool and the potential defect area; eliminate image artifacts through an edge-preserving smoothing algorithm, and finally generate a preprocessed X-ray image dataset.

[0051] The working principle and effects of the above technical solution are as follows:

[0052] The system was calibrated for stability, sensitivity and resolution according to the key parameters of X-ray inspection. The dynamics of the molten pool were captured by high-frequency sampling at 20-50 frames / second, which not only improved the parameter accuracy and image acquisition reliability of the real-time X-ray inspection system, but also reduced the problem of invalid raw data caused by system parameter deviation or sampling omission.

[0053] By extracting grayscale information from images and substituting it into a dynamic ray absorption model, the spatiotemporal distribution of the molten pool temperature field and flow field is calculated using an inversion algorithm. This transforms surface images into internal physical field data, improving the depth of understanding of the actual behavior inside the molten pool and reducing the one-sidedness of judging the state of the molten pool solely based on visual images.

[0054] After establishing a defect classification system, Bayesian inference was used in conjunction with melt pool behavior data to calculate the initial probability of defects. The distribution was also presented through a heat map, making the risk of defect occurrence more intuitive, improving the pertinence of defect prediction, and reducing the effort wasted on non-high-risk areas during subsequent identification.

[0055] The method employs adaptive median filtering combined with wavelet thresholding for noise reduction, followed by grayscale stretching and edge smoothing. This approach effectively addresses Compton scattering noise and image artifacts, improving the clarity and defect region identification of the preprocessed X-ray image while reducing the interference of noise and artifacts on subsequent feature extraction and defect recognition.

[0056] In one embodiment of the present invention, S3 includes:

[0057] S31. Based on the multimodal data requirements for defect identification (visual images, physical behavior, probability distribution), construct a coupled inference network architecture with a dual-branch + fusion module; determine that the visual feature extraction branch adopts an improved ResNet-50 network (with added attention mechanism) to extract visual features such as morphology, grayscale, and edges of defects; determine that the physical feature extraction branch adopts a fully connected neural network to extract physical features such as the correlation between molten pool temperature anomalies, flow velocity anomalies, and prior probability of defects; clarify that the output dimension of both branches is 256-dimensional to prepare for subsequent fusion.

[0058] S32. Normalize the generated preprocessed ray image dataset (map pixel values ​​to the 0-1 range) and divide it into a training subset and a validation subset in a 7.5:2.5 ratio; input the training subset into the visual feature extraction branch, and gradually extract the local morphological features of defects (such as the roundness of pores, the aspect ratio of cracks, location, etc.) through the convolutional layers, pooling layers, and attention layers of the improved ResNet-50 network. Figure 3 As shown), the model incorporates global grayscale features (e.g., the grayscale difference between the defect and the substrate); the branch model is fine-tuned using a validation subset to ensure feature extraction accuracy is higher than 90%, ultimately generating a defect visual feature dataset;

[0059] S33. Standardize the spatiotemporal dataset of molten pool behavior and the dataset of prior probability distribution of defects (mean is 0, variance is 1) to construct a physical feature input matrix; input the input matrix into the physical feature extraction branch, and learn the physical laws such as the correlation between molten pool temperature field anomaly and porosity formation, and the correlation between flow field anomaly and non-fusion through the hidden layers of a fully connected neural network (3 layers, each with 512, 256, and 256 neurons respectively);

[0060] S34. Optimize network parameters using a loss function (e.g., mean squared error loss) to generate a dataset of physical causes of defects; use a spatial attention alignment algorithm to spatially match the dataset of visual features of defects with the dataset of physical causes of defects (ensuring that the visual features and physical features of the same defect region correspond).

[0061] S35. Input the aligned bimodal features into the feature fusion module, and use a fusion strategy of element-weighted summation + feature concatenation (weights are determined by cross-validation, with visual features having a weight of 0.6 and physical features having a weight of 0.4) to perform semantic-level feature fusion; compress the dimensions of the fused features (from 512 dimensions to 256 dimensions) to generate a fused feature representation dataset.

[0062] The working principle and effects of the above technical solution are as follows:

[0063] To address the needs of multimodal data, a dual-branch coupled architecture was designed. The visual branch uses an improved ResNet-50 with an attention mechanism to capture morphological details, while the physical branch uses a fully connected network to mine the correlation between the melt pool and defects. The output dimension is also unified to facilitate subsequent fusion. This not only enhances the targeting of different types of feature extraction, but also reduces the problem of relying on a single branch to extract features and the incomplete information coverage.

[0064] The preprocessed images are normalized and divided into training and validation sets. Then, features are extracted step by step through network layers and fine-tuned to ensure an accuracy of over 90%, which improves the accuracy of defect visual features and reduces feature extraction bias caused by uneven data distribution or unoptimized models.

[0065] First, the molten pool behavior and defect probability data are standardized to construct an input matrix. Then, a 3-layer fully connected network is used to learn the physical laws, so that the physical features can accurately correspond to the causes of defects (such as temperature anomalies corresponding to porosity). This improves the matching degree between physical features and defects and reduces the situation of insufficient capture of the relationship between molten pool behavior and defects.

[0066] Optimizing network parameters using a loss function makes physical features more reliable. Then, spatial attention alignment ensures that the visual and physical feature positions of the same defect correspond, which improves the consistency of bimodal features and reduces the error caused by feature position misalignment in subsequent fusion.

[0067] By using a weighted splicing strategy to fuse bimodal features (taking into account both visual details and physical causes), and then compressing the dimensions to remove redundancy, the practicality of the fused features is enhanced (it can fully reflect the characteristics of defects), while reducing the subsequent computational burden brought by high-dimensional features.

[0068] In one embodiment of the present invention, S34 includes:

[0069] For the fully connected neural network in the physical feature extraction branch, optimization parameters are set (initial learning rate 0.001, number of iterations 50, batch size 32); the network's predicted output for the physical feature input matrix is ​​obtained and compared with the physical cause labels of known defects (such as "porosity-temperature field inhomogeneity" and "lack of fusion-flow field stagnation"); the mean square error loss function is used to calculate the loss value for each iteration and generate the loss value change curve.

[0070] Backpropagation updates network parameters and verifies convergence. Based on the loss value, backpropagation is performed through the Adam optimizer to update the hidden layer weights and biases of the fully connected network. Every 10 iterations, the network performance is evaluated using a validation set (20% of the physical feature input matrix). When the loss value of the validation set fluctuates less than 0.005 for 3 consecutive iterations, the network is considered to have converged, the iteration is stopped, and the optimized physical feature extraction branch model is obtained.

[0071] The defect physical cause feature dataset is generated by inputting the complete physical feature matrix constructed by S33 into the optimized physical feature extraction branch model, and outputting a 256-dimensional physical cause feature vector for each sample; the feature vectors are validated (abnormal vectors with a magnitude of 0 are removed), sorted by sample ID, and a structured defect physical cause feature dataset is generated.

[0072] Spatial coordinate information of dual-modal features is extracted from the defect visual feature dataset. The spatial coordinates of the ray image corresponding to each visual feature are extracted (pixel coordinates are converted to world coordinates with an accuracy of 0.1 mm). The spatial coordinates of the melt pool region of the corresponding sample are extracted from the defect physical cause feature dataset (derived based on the spatiotemporal data of melt pool behavior). A visual feature-physical feature-spatial coordinate association table is established.

[0073] The spatial attention alignment algorithm is used to achieve position matching based on the spatial coordinates in the association table. The algorithm calculates the coordinate deviation between visual features and physical features, and dynamically adjusts the coordinates of physical features through attention weights (determined by feature importance) to make the coordinate deviation of the bimodal features in the same defect area less than 0.5mm. After alignment, samples with failed coordinate association are deleted, and a spatially aligned bimodal feature dataset (visual features + physical causal features) is generated.

[0074] The working principle and effects of the above technical solution are as follows:

[0075] By setting clear network optimization parameters, calculating iterative loss using mean squared error loss and generating change curves, updating parameters through backpropagation of the Adam optimizer, and judging convergence by validation set loss fluctuations, the parameter accuracy and stability of the physical feature extraction model are improved, and feature extraction bias caused by insufficient network optimization is reduced.

[0076] Substituting the complete physical feature input matrix into the optimized model, the output feature vector is verified and anomalies are removed. The structured dataset is generated by sorting by ID, which improves the reliability and standardization of the physical cause feature data of defects and reduces the interference of invalid or chaotic data on subsequent processing.

[0077] Extracting the spatial coordinates of visual and physical features and establishing an association table, converting pixel coordinates into high-precision world coordinates, enhances the spatial correlation of dual-modal features and reduces the difficulty of feature correspondence caused by missing or inconsistent coordinate information.

[0078] The spatial attention alignment algorithm is used to dynamically adjust the coordinates, control the deviation within 0.5mm, delete failed association samples, improve the spatial consistency of dual-modal features in the same defect area, reduce the error caused by positional misalignment to feature fusion, and make subsequent fusion more accurate.

[0079] In one embodiment of the present invention, S35 includes:

[0080] S351. Perform spatial consistency verification on the aligned bimodal features (visual features and physical causal features) output by S34. By calculating the spatial coordinate deviation value of feature points (with a threshold set to 0.5 pixels), abnormal samples with excessive deviation are removed. The bimodal features that pass the verification are sorted by time series to construct a feature fusion input queue.

[0081] S352. Based on the weight coefficients determined by cross-validation (0.6 for visual features and 0.4 for physical features), perform a weighted summation calculation on the corresponding dimension elements of the bimodal features in the input queue (e.g., the i-th dimension element of visual features × 0.6 + the i-th dimension element of physical features × 0.4) to generate a preliminary weighted fusion feature vector (256 dimensions).

[0082] S353. Retain the independent information of the original bimodal features, and directly concatenate the visual feature vector (256 dimensions) and the physical cause feature vector (256 dimensions) in dimensional order to form a concatenated feature vector (512 dimensions); standardize the concatenated features through the BatchNorm layer;

[0083] S354. Input the weighted fusion features and the spliced ​​features into the gated fusion unit, learn the feature importance mask through the sigmoid activation function, and dynamically filter and reorganize the two fusion results to strengthen key semantic information (such as core defect features) and weaken redundant information (such as background noise features).

[0084] S355. Principal Component Analysis (PCA) algorithm is used to compress the dimensions of the optimized fusion features (512 dimensions), retaining the principal components with a cumulative variance contribution rate of 95%, reducing the feature dimension to 256 dimensions; the compressed features are normalized (mapped to the 0-1 range), and finally the fusion feature representation dataset is generated.

[0085] The working principle and effects of the above technical solution are as follows:

[0086] Spatial consistency verification is performed on the aligned bimodal features to remove abnormal samples with coordinate deviations exceeding 0.5 pixels. An input queue is also constructed by sorting the data by time series, which improves the purity of the feature fusion input data and reduces the interference of abnormal samples to the subsequent fusion process.

[0087] The dual-modal features are weighted and summed according to the weights determined by cross-validation (visual 0.6, physical 0.4), which allows the visual features responsible for capturing the morphology of defects and the physical features responsible for finding the causes to be reasonably matched, thereby enhancing the information balance of the initial fused features and reducing the information bias problem caused by unreasonable weights of single features.

[0088] By concatenating bimodal features while preserving their independent information, and then using the BatchNorm layer for standardization, the integrity of the features is improved (without losing original details), and the bias caused by the difference in the scale of different features is reduced, making subsequent fusion smoother.

[0089] The gated fusion unit learns feature masks through the sigmoid activation function, dynamically filters and recombines weighted and spliced ​​features, which can focus on strengthening the core information of defects and weakening background noise, thereby improving the targeting of fused features and reducing the interference of redundant information on subsequent defect identification.

[0090] When using PCA to compress dimensionality, principal components with 95% cumulative variance are retained, and then the features are normalized. This ensures that the key information of the fused features is not lost, reduces the subsequent computational burden caused by high-dimensional features, and improves the stability of the feature data, making it easier for subsequent model processing.

[0091] In one embodiment of the present invention, S355 includes:

[0092] The optimized fused features (512 dimensions) output are cleaned to remove null and outlier values ​​(using the Z-score method with a threshold of ±3); the processed feature data is converted into the matrix format (number of samples × feature dimension) required by the PCA algorithm to generate the feature matrix to be reduced in dimensionality.

[0093] Perform PCA decomposition on the feature matrix and calculate the variance contribution rate of each principal component; accumulate the variance contribution rates from largest to smallest to select the set of principal components with a cumulative variance contribution rate of 95%, determine the number of principal components to be retained (ensure the final dimension is 256); record the eigenvectors of the corresponding principal components and generate the PCA transformation matrix.

[0094] The feature matrix to be reduced in dimensionality is multiplied by the PCA transformation matrix to map the 512-dimensional features to the selected 256-dimensional principal component space, thus completing the dimensionality compression. The compressed feature vectors are then checked for integrity (to ensure no missing dimensions) to generate the dimensionality-reduced feature dataset.

[0095] Calculate the maximum and minimum values ​​of each dimension in the dimensionality-reduced feature dataset, and use the min-max normalization formula (eivalue = (eivalue - minimum value) / (maximum value - minimum value)) to map all eigenvalues ​​to the range of 0-1; verify the normalization effect through standard deviation test (ensure that the standard deviation of the normalized data is in the range of 0.2-0.3) to generate a normalized feature dataset;

[0096] The normalized feature dataset is indexed and associated by sample ID, and feature source tags (such as the association ID corresponding to the original visual / physical features) are added; the processed data is converted into a standardized data format (such as CSV or Tensor format), and finally a fused feature representation dataset is generated.

[0097] The working principle and effects of the above technical solution are as follows:

[0098] The feature data was cleaned using the Z-score method (threshold ±3) to remove null and outlier values, and then converted into the matrix format required for PCA. This improved the purity of the data to be reduced in dimension reduction and reduced the interference of outliers on the principal component analysis results.

[0099] By calculating the variance contribution rate through PCA decomposition, principal components with a cumulative contribution of 95% were selected and the number of components to be retained was determined. While reducing the dimensionality to 256 dimensions, key feature information was not lost. This not only reduced the computational burden of high-dimensional data, but also reduced the information loss caused by excessive dimensionality reduction.

[0100] After performing dimension mapping using matrix multiplication, an integrity check is performed to ensure that the compressed feature vector has no missing dimensions, which improves the reliability of the dimensionality reduction features and reduces errors caused by incomplete data in subsequent processing.

[0101] The min-max normalization was used to map the feature values ​​to the range of 0-1, and the effect was verified by the standard deviation test (0.2-0.3 interval), which enhanced the stability and consistency of the feature data and reduced the interference of differences in the scale of different dimensions on the subsequent model.

[0102] By associating the sample ID with an index, supplementing the source tags, and converting it into a standardized format, the fused feature dataset is made easier to call in subsequent processes, improving the usability of the data and reducing processing obstacles caused by format confusion or incomplete information.

[0103] In one embodiment of the present invention, step S4 includes:

[0104] S41. Using the Random Forest feature importance evaluation algorithm, the 256-dimensional features in the generated fusion feature representation dataset are ranked by importance, and the top 50 key features (such as defect area, gray-level difference, temperature anomaly value, and prior probability value) are selected. Based on the key features, the C4.5 algorithm is used to construct a defect classification decision tree, setting the maximum depth of the decision tree to 10 and the minimum number of sample splits to 5, to ensure that the decision tree can classify accurately while avoiding overfitting. The decision tree is optimized through 10-fold cross-validation to generate an initial defect classification decision tree model.

[0105] S42. Based on the initial defect classification decision tree model, an interpretable machine learning method combining LIME (Local Interpretability Model - Agnostic Interpretation) and SHAP (Shapley Additive Explanations) is used to extract the judgment logic of each decision node (e.g., defect area > 0.5 mm² and grayscale difference > 50, porosity, temperature anomaly < -200℃ and flow field velocity < 0.2 m / s, non-fusion). The extracted judgment logic is standardized and organized, and redundant and contradictory rules are removed to form a structured defect type judgment rule base (containing 4 types of defects and a total of 32 core rules).

[0106] S43. For each feature data in the fused feature representation dataset, recursively match it according to the rule order of the defect type determination rule base (match high-priority rules first, such as the feature rule for cracks); during the matching process, simultaneously record the physical cause explanation corresponding to each rule (e.g., matching the rule of temperature gradient > 500℃ / mm, determining it as a hot crack, the cause is that the molten pool cools too quickly); calculate the confidence of the matching results (based on the rule matching degree and feature importance), and generate a preliminary identification result dataset containing defect location, type, cause and confidence level.

[0107] S44. Collect historical data of X-ray inspection of arc additive manufacturing for the past N years (where N≥3) (containing 10,000+ valid samples), screen out false alarm cases (e.g., misjudging molten pool spatter as inclusions, misjudging surface scratches as cracks), extract the feature patterns of false alarm cases (e.g., gray scale distribution of spatter, edge continuity of scratches), and construct a false alarm pattern library.

[0108] S45. Compare the preliminary recognition result dataset with the false alarm pattern library using the cosine similarity algorithm, with the similarity threshold set to 0.85. Remove false alarm results with a matching degree higher than the threshold. Perform sampling verification on the remaining results (sampling ratio 10%) to ensure that the false alarm rate is less than 3%. Finally, generate the optimized recognition result dataset.

[0109] The working principle and effects of the above technical solution are as follows:

[0110] The top 50 key features were selected using random forest, and the maximum depth (10) and minimum number of splits (5) of the decision tree were controlled by the C4.5 algorithm. 10-fold cross-validation was also used for optimization, which not only improved the classification accuracy of the initial defect classification decision tree, but also avoided model overfitting and reduced the interference of redundant features on the model's judgment.

[0111] By combining the decision-making logic extracted from LIME and SHAP, a structured rule base containing 4 types of defects and 32 core rules was created, which enhanced the interpretability of defect judgment (the judgment basis can be clearly seen), reduced the difficulty of understanding caused by the "black box" of traditional machine learning models, and improved the convenience of rule calling.

[0112] Feature data is recursively matched according to rule priority, and the physical cause explanation is recorded and confidence is calculated simultaneously. This ensures that the preliminary identification results include not only the defect type, but also the cause and confidence, which increases the amount of information in the results and reduces the trouble of having to trace the cause again. At the same time, matching according to priority also reduces the probability of mismatch.

[0113] We collected over 10,000 historical detection data points from the past three years, extracted feature patterns from false alarm cases to build a database, covering common false alarm types such as molten pool spatter and surface scratches. This enhanced the targeting of subsequent false alarm screening, reduced the problem of missed screening due to incomplete false alarm patterns, and provided a foundation for long-term optimization of false alarm handling.

[0114] The preliminary results were compared with the false alarm library using cosine similarity (threshold 0.85), and the false alarm rate was controlled to be below 3% by 10% sampling verification. This directly reduced the false alarm rate of the final identification results, improved the reliability of defect identification, and reduced misjudgments in process adjustments caused by false alarms.

[0115] In one embodiment of the present invention, S43 includes:

[0116] For each 256-dimensional feature data in the fused feature representation dataset, supplement the basic information of the corresponding sample (such as X-ray image coordinates, detection time, and process parameter labels) to establish a feature data-sample information association mapping table; remove invalid feature data with missing association information to generate a structured list of feature data to be matched;

[0117] Based on the severity of defects (cracks > lack of fusion > porosity > inclusions), the 32 core rules in the defect type determination rule base are prioritized (crack rules are at level 1, the highest, and inclusion rules are at level 4, the lowest). The rule matching process is constructed in the order of level 1, level 2, level 3, and level 4, and the matching triggering conditions of each rule are clarified (such as feature data satisfying all the determination thresholds in the rule).

[0118] Extract a single data point from the list of feature data to be matched, first substitute it into the level 1 crack rule for matching. If the triggering condition is met, the defect type is directly marked as crack. If not, recursively substitute it into the level 2 non-fusion rule, and so on, until the corresponding level rule is matched. For feature data that does not match any rule, mark it as no defect and generate a preliminary type determination result list.

[0119] For each successfully matched feature data, extract the physical cause description associated with its triggering rule (e.g., when triggering a rule with a temperature gradient > 500℃ / mm, simultaneously record the cause: the molten pool cooling rate is too fast, leading to thermal stress concentration); bind the cause explanation with the preliminary type determination result to form a three-part data set of feature data - defect type - cause explanation;

[0120] The confidence score is calculated as follows: Confidence score = (match score × 0.6) + (mean of feature importance × 0.4). The calculated confidence score (rounded to two decimal places) is then added to the triplet dataset.

[0121] For all feature data, defect type, cause explanation, and confidence score quadruple data sets, sort them by sample detection time, extract the defect location from the basic information of the samples (converted from ray image coordinates to three-dimensional spatial coordinates), integrate them to form structured data containing defect location, type, physical cause, confidence score, and sample label, and finally generate a preliminary identification result dataset.

[0122] The working principle and effects of the above technical solution are as follows:

[0123] Supplementing each feature data with basic sample information, establishing an association mapping table, and removing invalid data improves the completeness and effectiveness of the data to be matched and reduces matching errors caused by missing information.

[0124] Prioritizing rules based on the severity of defects clarifies the matching process and triggering conditions, allowing high-risk defects (such as cracks) to be identified first, enhancing the targeting of the matching and reducing the risk of missing important defects.

[0125] By recursively matching feature data according to priority and labeling the type, each piece of data can be matched with the most appropriate defect type or judged as defect-free, which improves the accuracy of the initial type judgment and reduces misjudgment caused by disordered matching.

[0126] The physical cause explanation of the binding trigger rules forms a three-part data set, which makes the defect identification results not only have type but also specific cause, thus enhancing the explanatory power of the results and reducing the trouble of having to analyze the cause separately later.

[0127] The confidence score is calculated based on the rule matching degree and feature importance and added to the data set. The reliability of the result is quantified numerically, which improves the efficiency of judging the credibility of the recognition result and reduces subjective guessing about the reliability of the result.

[0128] By integrating data to generate a structured dataset containing information such as location, type, and cause, the initial identification results are made clear and easy to process, improving the practicality of the data and reducing obstacles to subsequent work caused by fragmented information.

[0129] In one embodiment of the present invention, step S5 includes:

[0130] S51. Based on the optimized identification result dataset, design the core functions of the self-evolution module; the core functions include: first, a novel defect sample screening mechanism, which automatically identifies novel defect samples not included in the existing defect type classification system (such as intergranular cracks under special processes); second, a sample annotation mechanism, which adopts an "automatic annotation + manual review" mode to annotate the location, type, cause and other information of novel defect samples to ensure that the annotation accuracy is higher than 95%; and third, a dataset update mechanism, which automatically incorporates the annotated novel defect samples into the training dataset to realize the dynamic expansion of the training set.

[0131] S52. Employ the federated average online incremental learning algorithm, using a coupled inference network as the base model. Input the expanded training dataset into the model in batches (100 samples per batch). Set a learning rate decay strategy (initial learning rate 0.001, decaying by 10% every 10 batches). Update the parameters of the model's convolutional and fully connected layers using gradient descent to avoid catastrophic forgetting of old data (use the Elastic Weight Consolidation (EWC) algorithm with a weight consolidation coefficient of 0.01). After each model update, evaluate the model performance using a validation dataset (the validation subset of S32) to ensure a defect identification accuracy improvement of ≥0.5%, generating an updated intelligent detection model.

[0132] S53. Set the model iteration cycle (iterate once every 500 new detection data collected), input the subsequent real-time detection data of arc additive manufacturing (preprocessed and feature extracted according to the process of S2-S4) into the updated intelligent detection model; through the real-time inference and result output of the model, feed back to the self-evolution module. If a new defect sample is found, repeat the update process of S51-S52; after 3-5 iteration cycles, a self-evolutionary arc additive manufacturing defect intelligent detection system with a closed loop of data acquisition-model update-performance improvement is formed.

[0133] S54. Integrate the defect identification results output by the self-evolving system. The information integration includes: defect spatial location (accurate to 0.1 mm, represented by three-dimensional coordinates), defect type (classified according to industry standards, such as GB / T39260-2020), physical cause (detailed analysis combined with melt pool behavior data), and confidence level (accurate to 0.1%). Using a structured document format (e.g., PDF), logically organize the integrated information according to the following order: inspection object information - defect statistical overview - detailed analysis of single defects - inspection conclusion suggestions. Generate a comprehensive inspection report data containing data visualization charts (e.g., defect location distribution map, type percentage pie chart).

[0134] The working principle and effects of the above technical solution are as follows:

[0135] The self-evolution module can automatically screen out new defects not included in the database, such as intergranular cracks under special processes. It also ensures that the labeling accuracy rate exceeds 95% through automatic labeling and manual review. At the same time, it dynamically expands the training set, which not only enhances the system's ability to identify new defects, but also reduces the workload of manually collecting and labeling new samples, and improves the comprehensiveness of the training data.

[0136] The model is updated using federated average online incremental learning, combined with a learning rate decay strategy and EWC algorithm to prevent catastrophic forgetting. It also ensures that the accuracy improves by ≥0.5% after each update, which improves the efficiency of model iteration, reduces the loss of the model's ability to identify old defects, and ensures that the model performance is steadily improved after the update.

[0137] By setting an iteration cycle of 500 new data points, a closed loop of data acquisition, model update, and performance improvement is formed, which enhances the long-term adaptability of the system, reduces the cost of rebuilding the model every time there is new detection data, and allows the system to continuously adapt to the process changes of arc additive manufacturing.

[0138] Defect information is integrated according to industry standards, with location accuracy down to 0.1mm and confidence level down to 0.1%. Visual charts such as location distribution maps and type percentage pie charts are also added, which improves the accuracy and readability of the comprehensive inspection report, reduces the time staff spend interpreting the report, and provides a more intuitive and reliable basis for subsequent process optimization.

[0139] One embodiment of the present invention provides an automatic defect identification system based on mathematical X-ray inspection in arc additive manufacturing, comprising:

[0140] One or more processors;

[0141] Memory, used to store one or more programs.

[0142] Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0143] The working principle and effects of the above technical solution are as follows:

[0144] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for automatic identification of defects based on mathematical radiography of electric arc additive manufacturing, characterized in that, The method comprises: S1, based on the material microstructure characteristics in the electric arc additive manufacturing process, a dynamic ray absorption model is established; the model is used to simulate and calculate the ray absorption characteristics under different process parameters, and theoretical background map data is generated; the detection area grid is divided according to the theoretical background map data, and the ray detection key parameter configuration scheme is determined; S2, the original ray image data of the molten pool area is collected through the real-time ray detection system, the molten pool behavior inversion calculation is carried out combined with the dynamic ray absorption model, the molten pool temperature field and flow field distribution data are generated; the defect prior probability calculation is carried out on the molten pool behavior data based on the Bayesian inference method, and the initial occurrence probability distribution data of different types of defects are obtained; the original ray image data is subjected to scattering noise suppression treatment, and preprocessed ray image data is generated; S3, a coupled reasoning network architecture including a physical feature extraction branch and a visual feature extraction branch is constructed; the preprocessed ray image data is input into the visual feature branch to extract defect morphology features, and the molten pool behavior data and the defect prior probability data are input into the physical feature branch to extract defect cause features; the spatial alignment and semantic fusion of multi-modal features are realized through a feature fusion module, and fusion feature representation data are generated; S4, a defect classification decision tree is constructed based on the fusion feature representation data, an interpretable machine learning method is used to generate a defect type judgment rule library; the fusion features are recursively matched and analyzed through the rule library, and preliminary recognition result data are generated; a false alarm mode library is constructed using historical detection data, the preliminary recognition results are compared and screened for false alarm modes, and optimized recognition result data are generated; S5, a self-evolution module including a new defect sample learning mechanism is established, new defect samples confirmed in the optimized recognition results are automatically included in the training data set; the coupled reasoning network parameters are updated through an online incremental learning algorithm, and an updated intelligent detection model is generated; the subsequent detection data are iteratively processed based on the updated model, forming an electric arc additive manufacturing defect intelligent detection system with self-evolution ability, and outputting comprehensive detection report data.

2. The method of claim 1, wherein, The S1 comprises: S11, for the electric arc additive manufacturing material to be detected, the microstructure parameters of the material are collected by a collection device to generate a material microstructure original data set; the original data set is subjected to outlier rejection and data standardization treatment to obtain a standardized microstructure parameter set; S12, based on the standardized microstructure parameter set, combined with the Beer-Lambert law of ray and matter interaction, a microstructure factor is introduced to modify the traditional ray absorption model, and a dynamic ray absorption model is constructed; S13, the model parameters are fitted and optimized by the least square method, and a verified dynamic ray absorption model is generated; the key process parameter range of electric arc additive manufacturing is determined, and a process parameter combination matrix is constructed; S14, the parameter combination matrix is input into the verified dynamic ray absorption model, the absorption intensity distribution of the material to the ray under different parameters is simulated and calculated, and a ray absorption simulation data set is generated; the ray detection theoretical background map corresponding to the process parameters is drawn based on the data set, and a theoretical background map data set is formed; S15, perform gray value gradient analysis on the images in the theoretical background map data set, identify areas with significant differences in ray absorption, and determine the key detection area; use an adaptive grid division algorithm to divide the detection area into grids, generating a weighted grid map of the detection area; combine the ray absorption characteristics of the grid map with the detection accuracy requirements to optimize the key parameters of the ray source and generate a ray detection key parameter configuration scheme.

3. The method of claim 1, wherein, The S2 comprises: S21, based on the generated ray detection key parameter configuration scheme, calibrate the parameters of the real-time ray detection system, and continuously image the molten pool area at a sampling frequency of 20-50 frames per second during the electric arc additive manufacturing process, to obtain a sequence of original ray images of the molten pool area, and generate an original ray image data set; S22, extract the gray scale information of the molten pool area of each image in the original ray image data set, and input it into the dynamic ray absorption model verified in S12; calculate the energy absorption distribution inside the molten pool by a model inversion algorithm, and further derive the spatiotemporal distribution data of the molten pool temperature field and flow field to generate a molten pool behavior spatiotemporal data set; S23, obtain common defect types in electric arc additive manufacturing, and establish a defect type classification system; input the generated molten pool behavior spatiotemporal data set into the Bayes inference formula as an evidence variable, and calculate the initial occurrence probability of each defect under different molten pool behavior conditions; S24, perform spatial interpolation processing on the probability results to generate an initial occurrence probability distribution heat map of different types of defects, and obtain a defect prior probability distribution data set; S25, analyze the noise types in the original ray image data set, and use a combination of adaptive median filtering and wavelet threshold denoising to suppress noise in the images; perform gray scale stretching and contrast enhancement on the denoised images, eliminate image artifacts through edge-preserving smoothing algorithm, and finally generate a preprocessed ray image data set.

4. The method of claim 1, wherein, The S3 comprises: S31, based on the multi-modal data requirements of defect identification, construct a coupled reasoning network architecture with a double-branch + fusion module; S32, normalize the generated preprocessed ray image data set, and divide it into a training subset and a verification subset according to a ratio of 7.5:2.5; input the training subset into the visual feature extraction branch, and gradually extract the local morphological features and global gray scale features of the defects through the convolution layer, pooling layer and attention layer of the improved ResNet-50 network; fine-tune the branch model using the verification subset, and finally generate a defect visual feature data set; S33, standardize the molten pool behavior spatiotemporal data set and the defect prior probability distribution data set to construct a physical feature input matrix; input the input matrix into the physical feature extraction branch, and learn the physical laws of molten pool temperature field anomalies and pore formation through the hidden layer of the fully connected neural network; S34, optimize the network parameters through the loss function to generate a defect physical cause feature data set; use a spatial attention alignment algorithm to match the spatial positions of the defect visual feature data set and the defect physical cause feature data set; S35, input the aligned bimodal features into a feature fusion module, adopt an element weighted summation+feature splicing fusion strategy, perform feature fusion at a semantic level, and perform dimension compression on the fused features to generate a fusion feature representation dataset.

5. The method of claim 4, wherein, The S35 comprises: S351, perform spatial consistency verification on the aligned bimodal features output by S34, sort the bimodal features that pass the verification according to a time sequence, and construct a feature fusion input queue; S352, based on the weight coefficients determined through cross-validation, perform weighted summation calculation on corresponding dimension elements of the bimodal features in the input queue to generate a preliminary weighted fusion feature vector; S353, retain independent information of the original bimodal features, directly splice the visual feature vector and the physical cause feature vector according to the dimension order to form a spliced feature vector, and perform standardization processing on the spliced feature through a BatchNorm layer; S354, input the weighted fusion feature and the spliced feature into a gated fusion unit, learn a feature importance mask through a sigmoid activation function, and dynamically screen and recombine the two fusion results; S355, perform dimension compression on the optimized fusion feature by adopting a principal component analysis algorithm, retain principal components with a cumulative variance contribution rate of 95%, and reduce the feature dimension to 256; and perform normalization processing on the compressed feature to finally generate a fusion feature representation dataset.

6. The method of claim 5, wherein, The S355 comprises: Perform data cleaning on the output optimized fusion feature to eliminate null values and abnormal values; convert the processed feature data into a matrix format required by the PCA algorithm to generate a feature matrix to be reduced in dimension; Perform PCA decomposition on the feature matrix to calculate the variance contribution rate of each principal component; add the variance contribution rates from large to small, select a principal component set with a cumulative variance contribution rate of 95%, determine the number of principal components to be retained, record the feature vectors of the corresponding principal components, and generate a PCA transformation matrix; Perform matrix multiplication on the feature matrix to be reduced in dimension and the PCA transformation matrix to map the 512-dimensional feature to the selected 256-dimensional principal component space, complete the dimension compression, and perform integrity verification on the compressed feature vector to generate a reduced feature dataset; Calculate the maximum and minimum values of each dimension in the reduced feature dataset, map all feature values to the range of 0-1 by using a min-max normalization formula, verify the normalization effect through a standard deviation test, and generate a normalized feature dataset; Index the normalized feature dataset according to the sample ID to supplement the feature source label; convert the processed data into a standardized data format to finally generate a fusion feature representation dataset.

7. The method of claim 1, wherein, The S4 comprises: S41, adopt a random forest feature importance evaluation algorithm to sort the 256-dimensional features in the generated fusion feature representation dataset according to importance, and select the top 50 key features; based on the key features, adopt a C4.5 algorithm to construct a defect classification decision tree; optimize the decision tree through 10-fold cross-validation to generate an initial defect classification decision tree model; S42, based on the initial defect classification decision tree model, using the explainable machine learning method combining LIME and SHAP, extracting the judgment logic of each decision node; standardizing and arranging the extracted judgment logic to form a structured defect type judgment rule library; S43, each feature data in the fusion feature representation dataset is represented, and recursive matching is performed according to the rule order of the defect type judgment rule library; in the matching process, the physical cause explanation corresponding to each rule is recorded synchronously; the matching result is calculated for confidence, and a preliminary identification result dataset is generated; S44, collect historical data of electric arc additive manufacturing radiographic testing in the past N years, screen false positive cases, extract feature patterns of false positive cases, and build a false positive pattern library; S45, compare the preliminary identification result dataset with the false positive pattern library, and eliminate false positive results with a matching degree higher than a threshold; sample verification is performed on the remaining results, and an optimized identification result dataset is finally generated.

8. The method of claim 5, wherein, The S43 comprises: For each 256-dimensional feature data in the fusion feature representation dataset, the basic information of the corresponding sample is supplemented, and a feature data-sample information association mapping table is established; invalid feature data with missing association information is removed, and a structured matching feature data list is generated; Based on the defect hazard degree, 32 core rules in the defect type judgment rule library are prioritized; the rule matching process is constructed in the order of level 1, level 2, level 3 and level 4, and the matching trigger condition of each rule is clearly defined; Extract a single data from the matching feature data list, first substitute it into the level 1 crack rule for matching, if the trigger condition is met, directly mark the defect type as crack; if not, recursively substitute it into the level 2 unfusion rule, and so on until the corresponding level rule is matched; for feature data without any rule matching, mark it as no defect, and generate a preliminary type judgment result list; For each matching successful feature data, extract the physical cause description associated with the trigger rule; bind the cause explanation with the preliminary type judgment result to form a feature data-defect type-cause explanation triad data group; The rule matching degree accounts for 60% of the weight, and the feature importance mean accounts for 40% of the weight; the calculated confidence is supplemented to the triad data group; Sort all feature data-defect type-cause explanation-confidence quadruple data groups by sample detection time, extract the defect position in the sample basic information, integrate to form structured data, and finally generate a preliminary identification result dataset.

9. The method of claim 1, wherein, The S5 comprises: S51, based on the optimized identification result dataset, design the core function of the self-evolution module; S52, using the federated average online incremental learning algorithm, taking the coupled reasoning network as the basic model, inputting the expanded training dataset into the model in batches; setting the learning rate decay strategy, updating the convolution layer and fully connected layer parameters of the model by gradient descent method; every time the model is updated, the validation dataset is used to evaluate the model performance, and an updated intelligent detection model is generated; S53, set the model iteration period, input the real-time detection data of subsequent electric arc additive manufacturing into the updated intelligent detection model; through the real-time inference and result output of the model, feedback to the self-evolution module, if a new defect sample is found, repeat the update process of S51-S52; after 3-5 iteration periods, the self-evolution electric arc additive manufacturing defect intelligent detection system is formed; S54, information integration is performed on the defect identification result output by the self-evolution system; the integrated information is logically organized in a structured document format; and a comprehensive detection report data is generated.

10. An automatic defect identification system based on electric arc additive manufacturing mathematical ray detection, comprising: one or more processors; a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 9.

Images