Internal Defect Monitoring Method for Laser Additive Manufacturing Process Based on Three-Level Spatiotemporal Graph Convolution

Abstract

The present invention belongs to the technical field of laser additive defect detection, and specifically discloses a method for monitoring internal defects in the laser additive manufacturing process based on three-level spatio-temporal graph convolution, including the following steps: Step 1: Continuously collect molten pool images in the laser-material interaction area under a certain time series; Step 2: Use a three-level spatio-temporal graph convolution network monitoring model to process the molten pool images to monitor internal defects in the laser additive manufacturing process; The present invention comprehensively captures multi-level spatio-temporal features in the dynamic evolution process of the molten pool. Based on the physical mechanism of point-by-point scanning, line-by-line overlapping, and layer-by-layer stacking in the laser additive manufacturing process, a three-level spatio-temporal framework is proposed; A hierarchical graph structure is constructed and graph convolution is applied to effectively capture multi-level spatio-temporal features in the dynamic evolution of the molten pool, accurately extract key information, and realize efficient online monitoring of internal defects in the laser additive manufacturing process.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser additive manufacturing defect detection, and particularly relates to a method for monitoring internal defects in the laser additive manufacturing process based on three-level spatio-temporal graph convolution. Background Art

[0002] Laser additive manufacturing manufactures complex-shaped parts by adding materials layer by layer. During the laser additive manufacturing process, dynamic changes in the molten pool, thermal stress, and instability of other process parameters may lead to the generation of local defects, such as cracks, pores, material shortages, etc. These defects have a significant impact on the quality and performance of the final product. Therefore, real-time monitoring and prediction of the occurrence of internal defects are important research directions for improving the quality of the laser additive manufacturing process.

[0003] Currently, the quality monitoring of laser additive manufacturing mainly relies on the method combining sensors and data-driven. Advanced imaging technologies have greatly promoted the condition monitoring and quality control of the laser additive manufacturing process. The development of various image sensing systems (such as infrared cameras, high-resolution cameras, optical sensors) enables the collection of rich information related to the quality characteristics of the laser additive manufacturing build, which helps to quickly identify the defect types, thereby improving the forming quality in the laser additive manufacturing process. However, the existing methods usually only focus on the transient spatial characteristics of the molten pool and fail to fully consider the defect-related temporal characteristics hidden in the dynamic evolution process of the molten pool, especially in the extraction of temporal characteristics of molten pool state fluctuations and local thermal state evolution. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for monitoring internal defects in the laser additive manufacturing process based on three-level spatio-temporal graph convolution; this method comprehensively captures multi-level spatio-temporal characteristics in the dynamic evolution process of the molten pool. Based on the physical mechanism of point-by-point scanning, line-by-line overlapping, and layer-by-layer stacking in the laser additive manufacturing process, a three-level spatio-temporal framework is proposed; a hierarchical graph structure is constructed and graph convolution is applied to effectively capture multi-level spatio-temporal characteristics in the dynamic evolution of the molten pool, accurately extract key information, and realize the efficient online monitoring of internal defects in the laser additive manufacturing process.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A method for monitoring internal defects in the laser additive manufacturing process based on three-level spatio-temporal graph convolution includes the following steps:

[0007] Step 1: Continuously collect molten pool images in the laser-material interaction area under a certain time series;

[0008] Step 2: Use a three-level spatio-temporal graph convolution network monitoring model to process the molten pool images to monitor internal defects in the laser additive manufacturing process;

[0009] Among them, the three - level spatio - temporal graph convolution network monitoring model includes an in - lane module, an inter - lane module, and an inter - layer module connected in sequence;

[0010] The in - lane module first chain - constructs the molten pool image path map of each path according to the temporal relationship of the printing path, then extracts the spatial features in the molten pool image path map, then extracts the high - dimensional features in the spatial features, and then extracts the global node features through global average pooling;

[0011] The inter - lane module first constructs the inter - lane path map with the global node features according to the temporal relationship of the printing path, then extracts the global features in the inter - lane path map, and finally extracts the global features of a fixed dimension through global average pooling;

[0012] The inter - layer module is used to construct the inter - layer path map with the global features of a fixed dimension according to the temporal relationship of the printing path; then extract the defect features in the inter - layer path map, and finally output the defect state in the target local area of the multi - pass and multi - layer specimen through aggregation.

[0013] Further, the in - lane module includes an image processing module, a two - dimensional convolution module, a graph convolution module, and a first global average pooling module connected in sequence;

[0014] The image processing module is used to represent the molten pool image sequence of each path as a chain structure based on the temporal dependence relationship of the printing path to form the molten pool image path map;

[0015] The two - dimensional convolution module is used to extract the spatial features in the molten pool image path map;

[0016] The graph convolution module is used to extract the high - dimensional features in the spatial features;

[0017] The first global average pooling module is used to extract the global node features from the high - dimensional features.

[0018] Further, the inter - lane module includes a feature processing module, a graph convolution aggregation module, and a second global average pooling module connected in sequence; the feature processing module constructs the inter - lane path map with the global node features extracted by the in - lane module in the printing order;

[0019] The graph convolution aggregation module is used to establish the temporal dependence relationship between the modules and lanes, and accumulate global information through the aggregation and update of the global node features to obtain global features;

[0020] The second global average pooling module is used to extract the global features of a fixed dimension from the global features.

[0021] Furthermore, the inter-layer module includes an inter-layer graph processing module, a graph convolution update module, a third global average pooling module, and a fully connected output module that are connected in sequence; the inter-layer graph processing module is used to construct an inter-layer path graph according to the timing relationship of the printing path for the fixed-dimension global features;

[0022] The graph convolution update module is used to establish the temporal dependence relationship between layers, and accumulate and update defect information by aggregating the fixed-dimension global features, so as to extract node defect features;

[0023] The third global average pooling module is used to extract defect features from the node defect features;

[0024] The fully connected output module is used to output the defect state in the target local area of the multi-channel multi-layer specimen after aggregating the defect features.

[0025] The present invention has the following beneficial effects:

[0026] (1) The molten pool images in the laser additive manufacturing process strictly follow the printing order, and the path graph can accurately reflect the relationship between the printing path and the time series, forming a chain topological structure for the molten pool images in the printing time series; at each time point, edges are established between a node and its front and rear neighbor nodes. This topological structure can not only capture the dynamic interaction features in the time series, but also ensure the physical consistency of the data at a low computational cost, providing an efficient and reliable basis for subsequent graph neural network modeling.

[0027] (2) By gradually modeling the spatio-temporal interaction relationships within a pass, between passes, and between layers, constructing a hierarchical graph structure and applying graph convolution, the multi-level spatio-temporal features in the dynamic evolution of the molten pool are effectively captured, key information is accurately extracted, and efficient online monitoring of internal defects in the laser additive manufacturing process is realized.

[0028] (3) The three-level spatio-temporal graph convolution network monitoring model can accurately and reliably predict the internal defect state in the laser additive manufacturing process. Description of the Drawings

[0029] Figure 1 It is a schematic structural diagram of the three-level spatio-temporal graph convolution network monitoring model of the present invention.

[0030] Figure 2 It is a confusion matrix diagram of an embodiment of the present invention.

[0031] Figure 3 It is a clustering result diagram of the model of an embodiment of the present invention; (a) three-dimensional diagram, (b) two-dimensional diagram. Detailed Embodiment

[0032] A method for monitoring internal defects in the laser additive manufacturing process based on three-level spatio-temporal graph convolution provided in this embodiment includes the following steps:

[0033] Step 1: Continuously collect the molten pool images of the laser-material interaction area in a certain time series;

[0034] Step 2: Use a three-level spatio-temporal graph convolutional network monitoring model to process the molten pool images to monitor the internal defects in the laser additive manufacturing process;

[0035] Among them, as Figure 1 shown, the three-level spatio-temporal graph convolutional network monitoring model includes an intra-track module, an inter-track module, and an inter-layer module connected in sequence;

[0036] The intra-track module first chains the molten pool images of each path into a molten pool image path graph according to the timing relationship of the printing path, then extracts the spatial features in the molten pool image path graph, then extracts the high-dimensional features in the spatial features, and then extracts the global node features through global average pooling.

[0037] The intra-track module in this embodiment includes an image processing module, a two-dimensional convolutional module, a graph convolutional module, and a first global average pooling module connected in sequence; the image processing module is used to represent the molten pool image sequence of each path as a chain structure based on the timing dependence of the printing path to form a molten pool image path graph, ensuring that the printing order and its associated features can be accurately reflected.

[0038] The two-dimensional convolutional module is used to extract the spatial features in the molten pool image path graph; the two-dimensional convolutional module is implemented by 5 layers of two-dimensional convolutional layers, and the convolutional kernel size in each layer of two-dimensional convolution is 3×3, and the stride is 1, respectively extracting low, medium, and high-level spatial features.

[0039] The graph convolutional module is used to extract the high-dimensional features in the spatial features; the graph convolutional module is implemented by 5 layers of graph convolutional layers, and uses the topological structure and temporal connectivity of the nodes in the path graph to extract the high-dimensional features from the spatial features.

[0040] The first global average pooling module is used to extract the global node features from the high-dimensional features.

[0041] The inter-track module first constructs the global node features into an inter-track path graph according to the timing relationship of the printing path, then extracts the global features in the inter-track path graph, and finally extracts the global features of a fixed dimension through global average pooling.

[0042] The inter-track module in this embodiment includes a feature processing module, a graph convolution aggregation module, and a second global average pooling module connected in sequence; the feature processing module constructs an inter-track path graph according to the printing order of the global node features extracted by the intra-track module; each intra-track path graph is represented as a one-dimensional global vector; the association relationship between multiple printing paths in the same layer is modeled to capture their spatial interdependence and dynamic interaction characteristics.

[0043] The graph convolution aggregation module is used to establish the temporal dependence relationship between tracks, and accumulate global information to obtain global features through the aggregation and update of global node features; the graph convolution aggregation module is implemented by using 5 graph convolution layers, and 5 graph convolution layers are used to model the complex dependence relationship between different printing paths; it can not only capture the local temporal dependence between tracks, but also gradually accumulate global information through the aggregation and update of node features.

[0044] The second global average pooling module is used to extract fixed-dimensional global features from the global features; the global features are aggregated into a fixed-dimensional global feature; this is not only beneficial to capturing the global dependence relationship between tracks, but also can effectively reduce the dimension, alleviate overfitting and enhance the robustness of the model; the inter-track module can effectively model the dependence relationship between different printing paths and provide a more accurate feature representation for subsequent tasks.

[0045] The inter-layer module is used to construct an inter-layer path graph according to the temporal relationship of the printing paths of the fixed-dimensional global features; then extract the defect features in the inter-layer path graph, and finally output the defect state in the target local area of the multi-track multi-layer specimen through aggregation.

[0046] The inter-layer module includes an inter-layer graph processing module, a graph convolution update module, a third global average pooling module, and a fully connected output module connected in sequence; the inter-layer graph processing module is used to construct an inter-layer path graph according to the temporal relationship of the printing paths of the fixed-dimensional global features, and the fixed-dimensional global features output by the inter-track module are regarded as nodes in the inter-layer path graph, and these nodes capture the printing path information, temporal features and their cross-track interaction dependencies of each layer, reflecting the temporal and spatial dependence relationships between different printing layers.

[0047] The graph convolution update module is used to establish the temporal dependence relationship between layers, and accumulate defect information to obtain node defect features through the aggregation and update of fixed-dimensional global features; the graph convolution update module is implemented by using 5 graph convolution layers, and 5 graph convolution layers are used to model the complex dependence relationship between different layers; it can not only capture the local temporal dependence between layers, but also gradually accumulate defect information through the aggregation and update of features.

[0048] The third global average pooling module is used to extract defect features from the node defect features; the fully connected output module is used to aggregate the defect features and output the defect states within the target local area of the multi-channel multi-layer specimen; the inter-layer module can more accurately capture the complex temporal relationship and spatial dependence between different printing layers, providing a more accurate feature representation for the final defect prediction task.

[0049] Next, experimental data is used to verify the effectiveness of the method of this embodiment.

[0050] Collection of sample data: By designing process parameter combinations with different laser powers and scanning speeds, a 3-pass 30-layer process experiment is carried out, and a CCD camera is used to collect the molten pool images during the laser additive deposition process. The molten pool images are cropped and normalized to reduce the invalid redundant information therein, facilitating subsequent data processing.

[0051] Enrichment of sample data: Based on a sliding time window (window size is 0.4s, step size is 0.04s), continuous sampling of the molten pool image data is carried out on the time axis, and the defect features within the spatial region corresponding to this time window are combined for annotation; not only the local dynamic changes of the molten pool morphology are retained, but also the spatio-temporal correspondence between the molten pool morphology and the defect states is effectively established.

[0052] Annotation of defect information: Industrial CT is used to scan and detect the internal defects of the deposited samples offline. After scanning, the coordinates and volumes of each pore defect in the samples are extracted, and the total defect volume within the local area is calculated.

[0053] According to fixed thresholds (0.37 and 1.4), the defects are divided into three size categories: negligible (0), medium (1), and large (2). Finally, a high-quality data set containing 4360 samples is constructed, and the sample distribution is 2878 negligible (0), 926 medium (1), and 556 large (2).

[0054] The data set is divided into a training set and a test set in a ratio of 8:2.

[0055] The three-level spatio-temporal graph convolutional network monitoring model is trained using the training set, and the weight parameters and bias parameters of the model are iteratively optimized using the stochastic gradient descent algorithm and the error backpropagation algorithm to minimize the cross-entropy loss function, obtaining the optimized network parameters.

[0056] After training and optimization, the network hyperparameters of the three-level spatio-temporal graph convolutional network monitoring model are finally determined as follows: the learning rate adjustment method is the cosine annealing algorithm with warmup, the initial learning rate is 0.001, the minimum learning rate is 0, the maximum cosine period is 40, the period decay constant is 2, the iteration starting point number is -1, the minimum batch size is 128, and the iteration period is 200 generations.

[0057] To avoid the contingency of training, the three-level spatio-temporal graph convolutional network monitoring model was trained 5 times and the average value was taken, obtaining a predicted average accuracy of 95.18%. The confusion matrix of the prediction results is as follows Figure 2 . The accuracy of the model for the negligible state is 98.60%, and the recall rate is 97.91%; the accuracy for the medium state is 86.17%, and the recall rate is 92.05%; the accuracy for the large state is 92.98%, and the recall rate is 86.89%; the micro-F1 index of the model is 0.9518, and the macro-F1 index of the model is 0.9236.

[0058] The t-SNE technique was used to perform a visual analysis of the deep features, as shown in Figure 3 . By performing dimensionality reduction on the high-dimensional features, the t-SNE graph can intuitively display the distribution of different category samples in the feature space. It not only reflects the discrimination ability of the model for various categories but also reveals the relationships between samples of various categories, including the clustering effect and potential feature overlap. This analysis provides a more intuitive and in-depth perspective for understanding the classification effect of the model.

[0059] The three-level spatio-temporal graph convolutional network monitoring model described in this embodiment can accurately and reliably predict the internal defect state during the laser additive manufacturing process.

[0060] The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solutions and inventive concepts provided by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for monitoring internal defects in the laser additive manufacturing process based on three-level spatio-temporal graph convolution, characterized in that, The method includes the following steps: Step 1: Continuously collect the molten pool images of the laser-material interaction area in a certain time series; Step 2: Use a three-level spatio-temporal graph convolutional network monitoring model to process the molten pool images to monitor the internal defects in the laser additive manufacturing process; Among them, the three-level spatio-temporal graph convolutional network monitoring model includes an in-track module, an inter-track module, and an inter-layer module connected in sequence; The in-track module first chains the molten pool images of each path into a molten pool image path graph according to the time sequence relationship of the printing path, then extracts the spatial features in the molten pool image path graph, then extracts the high-dimensional features in the spatial features, and then extracts the global node features from the high-dimensional features through global average pooling; The inter-track module first constructs an inter-track path graph with the global node features according to the time sequence relationship of the printing path, then extracts the global features in the inter-track path graph, and finally extracts the global features of a fixed dimension through global average pooling; The inter-layer module is used to construct an inter-layer path graph with the global features of a fixed dimension according to the time sequence relationship of the printing path; then extract the defect features in the inter-layer path graph, and finally output the defect status in the target local area of the multi-track and multi-layer specimen through aggregation.

2. The internal defect monitoring method for the laser additive manufacturing process based on three-level spatio-temporal graph convolution according to claim 1, wherein, The in-track module includes an image processing module, a two-dimensional convolutional module, a graph convolutional module, and a first global average pooling module connected in sequence; The image processing module is used to represent the molten pool image sequence of each path as a chain structure based on the time sequence dependence relationship of the printing path to form a molten pool image path graph; The two-dimensional convolutional module is used to extract the spatial features in the molten pool image path graph; The graph convolutional module is used to extract the high-dimensional features in the spatial features; The first global average pooling module is used to extract the global node features from the high-dimensional features.

3. The internal defect monitoring method for the laser additive manufacturing process based on three-level spatio-temporal graph convolution according to claim 2, wherein The inter-track module includes a feature processing module, a graph convolutional aggregation module, and a second global average pooling module connected in sequence; the feature processing module constructs an inter-track path graph with the global node features extracted by the in-track module according to the printing order; The graph convolutional aggregation module is used to establish the time sequence dependence relationship between the molds and between the tracks, and accumulate the global information by aggregating and updating the global node features to obtain the global features; The second global average pooling module is used to extract the global features of a fixed dimension from the global features.

4. The internal defect monitoring method for the laser additive manufacturing process based on three-level spatio-temporal graph convolution according to claim 3, characterized in that The inter-layer module includes an inter-layer graph processing module, a graph convolutional update module, a third global average pooling module, and a fully connected output module connected in sequence; the inter-layer graph processing module is used to construct an inter-layer path graph with the global features of a fixed dimension according to the time sequence relationship of the printing path; The graph convolutional update module is used to establish the time sequence dependence relationship between layers, and accumulate the defect information and extract the node defect features by aggregating and updating the global features of a fixed dimension; The third global average pooling module is used to extract the defect features from the node defect features; The fully connected output module is used to aggregate the defect features and output the defect status in the target local area of the multi-track and multi-layer specimen.

Images