Metal AM component defect contour detection method

Abstract

The invention belongs to the technical field of metal AM component defect detection, and particularly relates to a metal AM component defect contour detection method which comprises the following steps: dividing a defect-containing surface of a to-be-detected metal AM component into a plurality of pixel grids with the same size, and carrying out laser scanning by adopting a matrix grid scanning mode to obtain an LIBS spectrum of each pixel grid, dimension reduction processing is carried out; inputting the LIBS spectrums subjected to dimension reduction processing of the pixel grids into a grid classification model corresponding to the defect type of the metal AM component to be detected so as to identify areas corresponding to the pixel grids; and drawing the defect contour of the to-be-detected metal AM component based on the pixel grid identified as the defect contour area. According to the method, the defect contour in the metal AM component can be identified with high precision, and accurate form information of the defect surface is provided, so that an important basis is provided for quality evaluation and defect analysis.

Description

Technical Field

[0001] This application belongs to the technical field of metal AM component defect detection, and particularly relates to a method for detecting the defect contour of metal AM components. Background Art

[0002] Metal additive manufacturing (AM) technology can achieve complex shapes and precise structures that are difficult to process by traditional manufacturing methods through layer-by-layer deposition of metal materials, and is gradually changing the pattern of the manufacturing industry. Compared with traditional subtractive manufacturing technology, metal AM technology has significant advantages in dealing with internal defects. Its layer-by-layer stacking method provides an opportunity to detect and process defects that appear on the surface of each stacked layer, thus effectively avoiding the generation of internal defects. In metal AM technology, the surface contour information of different types of defects is crucial in the subsequent repair process. The contour of the defect not only determines the material filling method during repair, but also affects the control of the cooling rate to ensure the consistency of the repaired surface with the original structure. Therefore, accurately detecting the contour information of surface defects in each stacked layer can provide an important basis for subsequent defect repair, thereby ensuring the quality level of metal AM components, especially in high-demand application fields such as aerospace, automotive industry, and medical treatment, etc.

[0003] Traditional metal AM component surface defect detection technologies include non-destructive testing methods such as ultrasonic, eddy current, and thermal imaging. The above technologies can capture the surface defect information of metal AM components without damaging them. However, in practical applications, these methods still have certain limitations. Specifically, ultrasonic and eddy current technologies are often affected by the geometric shape and material heterogeneity of metal AM components, resulting in a reduction in the accuracy of detection results; the thermal response of thermal imaging technology is affected by multiple factors such as material thermal conductivity, surface smoothness, and external environmental temperature, making its detection results often inaccurate and even possibly unable to fully cover the defect area.

[0004] Therefore, in order to find an effective method for accurately detecting the surface defect information of metal AM components, many scholars have used neural networks (NN) and convolutional neural networks (CNN) to analyze and construct additive manufacturing defects in hierarchical images. The following achievements have been obtained: 1) CNN performs better than traditional NN under different lighting conditions, especially in terms of generalization ability, and can maintain a high accuracy rate on unseen datasets. In addition, lack of fusion defects may be related to process ejecta; 2) A deep learning defect detection scheme based on the YOLOv4 model and attention mechanism is applied to wire arc additive manufacturing (WAAM). This method improves three existing object detection models, including channel attention mechanism, multi-spatial pyramid pooling, and exponential moving average. In addition, the evaluation of the WAAM defect dataset shows that the model reaches an average precision (mAP) of 94.5% and can process at least 42 frames per second; 3) An infrared thermography non-destructive testing method based on an improved detection probability (POD) function studies the correlation between defect size, depth, and detection probability, and evaluates the detectability of spherical defects in 3D printing using an enhanced probability function. However, the above defect detection methods mainly focus on the detection of defect types and specific sizes, and the research on the detection of defect contours of metal AM components is still relatively scarce. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method for detecting the defect contour of a metal AM component, which can accurately identify the defect contour in the metal AM component and provide accurate morphological information of the defect surface, thereby providing an important basis for quality assessment and defect analysis.

[0006] This application provides a method for detecting the defect contour of a metal AM component, including: Dividing the defective surface of the metal AM component to be detected into several pixel grids of the same size, and performing laser scanning in a matrix grid scanning manner to obtain the LIBS spectra of each pixel grid, and performing dimensionality reduction processing; Inputting the LIBS spectra after dimensionality reduction processing of each pixel grid into the grid classification model corresponding to the defect type of the metal AM component to be detected to identify the regions corresponding to each pixel grid; wherein, the regions include: defect-free region, defect contour region, and defect internal region; Based on the pixel grids identified as the defect contour region, draw the defect contour of the metal AM component to be detected.

[0007] Furthermore, the side length of the pixel grid, the diameter of the laser spot, and the step size of the laser scanning are the same.

[0008] Furthermore, the grid classification models corresponding to each defect type are obtained through the following methods: Using a metal plate of the same material as the metal AM component to be inspected, defect samples of different sizes under each defect type are prepared respectively; wherein the defect types include: pit defects, crack defects, and convex defects; the sizes are defect depth and defect width; For each defective sample, several pixel grids of the same size are divided and each region is determined. LIBS spectra of multiple pixel grids are randomly collected in each region and dimensionality reduction is performed. For each defect type, the LIBS spectra collected from different areas of each defect sample under this defect type and processed with reduced dimension are used to train the CNN model to obtain the grid classification model corresponding to this defect type.

[0009] Furthermore, after identifying the areas corresponding to the pixel grids, the method further includes: For each pixel grid, the KNN concept is used to monitor the number of neighboring pixel grids of the pixel grid that are identified as regions; Whether to adjust the area of ​​the pixel grid is determined according to a preset threshold.

[0010] Furthermore, the step of drawing the defect contour of the metal AM component to be inspected based on the pixel grid identified as the defect contour area includes: Determine the positional relationship between pixel grids identified as defect contour areas; wherein the positional relationship includes directly adjacent, diagonally adjacent, and non-adjacent; The pixel grids of directly adjacent defect contour areas are connected in sequence, and the pixel grids of two diagonally adjacent defect contour areas are connected to obtain the defect contour of the metal AM component to be inspected.

[0011] Furthermore, after obtaining the defect profile of the metal AM component to be inspected, the method further includes: Determining whether the defect contour constitutes a closed area; If the defect contour does not constitute a closed area, the middle positions of the adjacent pixel grids identified as the non-defective area and the defect internal area in the open area of ​​the defect contour are used as connection points to complete the defect contour.

[0012] Furthermore, the LIBS spectrum is subjected to dimensionality reduction processing by the following method: For each LIBS spectrum, after preprocessing by baseline removal, noise reduction and smoothing, and maximum and minimum standardization, the principal component analysis method was used to select the first 18 principal components with a total explained variance greater than 99% to characterize the characteristics of the LIBS spectrum, so as to obtain the LIBS spectrum after dimensionality reduction.

[0013] The metal AM component defect contour detection method provided by this application can accurately identify the defect contours in metal AM components and provide accurate morphological information of the defect surfaces, thereby providing an important basis for quality assessment and defect analysis. Brief Description of the Drawings

[0014] Figure 1 Shows the structural diagram of the LIBS spectrum acquisition device provided by an embodiment of this application; Figure 2 Shows the scanned sample diagrams of pit defects and crack defects provided by an embodiment of this application; Figure 3 Shows the average spectrograms of the defect-free area, defect contour area, and defect internal area of the scanned sample provided by an embodiment of this application; Figure 4 Shows the flowchart of the metal AM component defect contour detection method provided by an embodiment of this application; Figure 5 Shows the schematic diagram of the matrix grid scanning method provided by an embodiment of this application; Figure 6 Shows the schematic diagrams of pit defect samples and crack defect samples provided by an embodiment of this application; Figure 7 Shows the defect contour drawing effect diagram of a pit defect sample provided by an embodiment of this application; Figure 8 Shows the defect contour drawing effect diagram of another pit defect sample provided by an embodiment of this application. Detailed Embodiments

[0015] To make the objectives, technical solutions, and advantages of this technical solution clearer and more understandable, the following further details this technical solution in combination with specific embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of this technical solution.

[0016] First, the application scenario of this application is introduced. This application can be applied to metal AM components with clear defect types.

[0017] Secondly, in combination with Figure 1The structural diagram of the LIBS spectral acquisition device is shown to introduce the LIBS spectral acquisition method of this application. Specifically, a Q-switched Nd:YAG laser (Vlite-200 type, wavelength 1064 nm, energy 60 mJ, pulse width 8 ns, repetition rate 1 Hz, beam mode TEM00) is used as the ablation source. The laser pulse is reflected by a mirror and a lens and then focused 2 mm below the sample surface to induce a plasma. An optical fiber probe is used to collect the optical signal emitted by the plasma and transmit the signal to a spectrometer (SR-500i-a, Andor). Subsequently, an ICCD detector (1024×1024 pixels, Andor) converts the optical signal into digital information and stores it in a computer. The acquisition range of the spectral wavelength is 300 nm to 600 nm, and the resolution is 0.07 nm. At the same time, a digital pulse delay generator (BNC-575) is used for synchronization, and the gate width and delay time of the ICCD are set to 1 μs and 2 μs respectively. The sample is fixed on a three-dimensional moving platform to adjust its position. The experimental environment is at atmospheric pressure, temperature 23 °C, and relative humidity 26%. Example 1:

[0018] Through metal AM technology, pit defects and crack defects are respectively simulated as scanning samples using the same raw materials as the metal AM components to explore and analyze the LIBS spectral characteristics of different regions for different defect types.

[0019] Here, please refer to Figure 2 the scanning sample diagram of the pit defect shown in (a) and the scanning sample diagram of the crack defect shown in 2(b). In the figure, the size of the pit defect is a diameter of 4 mm and a depth of 2.5 mm, and the size of the crack defect is a length of 6 mm, a width of 4 mm, and a depth of 1.5 mm. Using the LIBS spectral acquisition device, the LIBS spectra of the defect-free region, defect contour region, and defect internal region of the scanning samples of the pit defect and the crack defect are respectively collected by the matrix grid scanning method.

[0020] This application selects the emission spectral lines of Fe, Cr, and Mn for analysis and finds that in different regions, the emission line intensities of Fe, Cr, and Mn show complex change patterns. Specifically, please refer to Figure 3 the average spectrograms of the defect-free region, defect contour region, and defect internal region of the scanning sample shown, where Figure 3 (a) is the average spectrogram of the scanning sample of the pit defect, Figure 3 and (b) is the average spectrogram of the scanning sample of the crack defect.

[0021] Consistent with expectations, the eight spectral intensity ratios of pit defects and crack defects have similar spectral characteristics, but the spectral line change patterns are relatively complex. Specifically, the ratio of FeI 357.39 / FeI 424.74nm in the internal area of the defect is higher than that in the defect-free area and the defect contour area. Among other spectral ratios, the spectral ratios in the internal area of the defect are relatively low. At the same time, for the same area, there are also differences in the relative intensities of the eight spectral ratios of pit defects and crack defects. For example, for the defect contour area, the ratio of FeII 519.91 / FeI 424.74nm of crack defects in this area is significantly higher than that of pit defects. This phenomenon can be attributed to the differences in defocus amount and spatial effect between the two types of defects. To avoid the similar spectral characteristics of pit defects and crack defects from affecting the detection accuracy of the defect area contour, the present application respectively establishes grid classification models for pit defects and crack defects, so as to prepare for defect contour detection. Embodiment 2:

[0022] Please refer to Figure 4 the flowchart of the method for detecting the defect contour of a metal AM component as shown in Figure 4 As shown, the method includes: S101: Divide the surface of the metal AM component to be detected with defects into several pixel grids of the same size, and perform laser scanning in a matrix grid scanning manner to obtain the LIBS spectra of each pixel grid, and perform dimensionality reduction processing.

[0023] Among them, the side length of the pixel grid, the diameter of the laser spot, and the step size of the laser scanning are all the same.

[0024] In this step, when the side length of the pixel grid, the diameter of the laser spot, and the step size of the laser scanning are all the same, it means that one scanning point corresponds to one pixel grid, so that the target area can be scanned by laser in a full area and without coverage, and at the same time, the efficiency of laser scanning can be improved. Here, please refer to Figure 5 the schematic diagram of the matrix grid scanning method as shown in

[0025] In addition, the following method is used to perform dimensionality reduction processing on the LIBS spectra: For each LIBS spectrum, after performing preprocessing such as baseline removal, noise reduction and smoothing, and maximum-minimum normalization in sequence, the first 18 principal components with a total explained variance greater than 99% are selected by the principal component analysis method to characterize the characteristics of the LIBS spectrum, so as to obtain the LIBS spectrum after dimensionality reduction processing.

[0026] Here, the maximum-minimum normalization process is performed on the LIBS spectra to align the spectral intensities within the same order of magnitude. Then, the spectral lines of the characteristic elements of the LIBS spectra are identified through the National Institute of Standards and Technology (NIST) database, and the cumulative contribution of the principal components is calculated. In order to retain as much of the original information of the LIBS spectra as possible, the first 18 principal components with an explained variance greater than 99% are selected to characterize the characteristics of the LIBS spectra, so as to achieve the dimensionality reduction processing of the LIBS spectra.

[0027] S102. Input the LIBS spectra after dimensionality reduction processing of each pixel grid into the grid classification model corresponding to the defect type of the metal AM component to be detected, so as to identify the regions corresponding to each pixel grid.

[0028] Among them, the regions include: defect-free regions, defect contour regions, and defect internal regions.

[0029] In specific implementation, the grid classification models corresponding to each defect type can be obtained through the following methods: Step 201. Use metal plates of the same material as the metal AM component to be detected to prepare defect samples of different sizes under each defect type.

[0030] Among them, the defect types include: pit defects, crack defects, and protrusion defects; the sizes are the defect depth and the defect width.

[0031] In this step, since the defect contours of protrusion defects and pit defects are similar, therefore, in the embodiments of the present application, only pit defects and crack defects are taken as examples to elaborate in detail the construction process of the grid classification models of the two types of defects. As an example, defect samples of pit defects as shown in Figure 6 (a) and crack defect samples as shown in 6(b) are prepared using metal plates of ferrochrome alloy materials. In the figure, there are 50 pit defect samples and 50 crack defect samples each. Among them, the sizes of the pit defects are as follows: the diameter range is 1 - 3.5 mm, and the depth range is 1 - 5 mm; the sizes of the crack defects are as follows: the length is fixed at 6 mm, the width range is 1 - 3.5 mm, and the depth range is 1 - 5 mm. Of course, protrusion defect samples or other types of defect samples can also be prepared using the same material and the same process, and the present application does not make any limitations here.

[0032] Step 202. For each defect sample, divide a number of pixel grids of the same size, and determine each region. Randomly collect the LIBS spectra of multiple pixel grids in each region, and perform dimensionality reduction processing.

[0033] As an example, each pothole defect sample is divided into a 30×30 pixel grid, and each crack defect sample is divided into a 34×30 pixel grid. The side length of the pixel grid, the diameter of the laser spot, and the step size of the laser scan are all set to 300 μm. At the same time, the defect-free area, the defect contour area, and the defect internal area of the pothole defect sample and the crack defect sample are respectively determined, and each area is mapped onto the pixel grid. For each area, multiple pixel grid LIBS spectra within the area are randomly collected, and the LIBS spectra are subjected to dimensionality reduction processing.

[0034] Step 203: For each defect type, use the LIBS spectra collected and dimensionally reduced in different areas of each defect sample under that defect type to train a CNN model to obtain a grid classification model corresponding to that defect type.

[0035] As an example, use the LIBS spectra collected and dimensionally reduced in different areas of each pothole defect sample to train a CNN model to obtain a grid classification model for pothole defects, and use the LIBS spectra collected and dimensionally reduced in different areas of each crack defect sample to train a CNN model to obtain a grid classification model for crack defects.

[0036] In addition, after identifying the areas corresponding to each pixel grid, the method further includes: Step 301: For each pixel grid, use the KNN idea to monitor the number of times the multiple neighboring pixel grids of the pixel grid are recognized as each area.

[0037] In this step, the core idea of the KNN classification algorithm is that if most of the K most similar (i.e., the closest in the feature space) samples of a sample in the feature space belong to a certain category, then the sample also belongs to this category. That is to say, the area of the pixel grid is related to the areas of multiple neighboring pixel grids. Therefore, obtain the areas of multiple neighboring pixel grids and respectively count the number of each area.

[0038] Step 302: Judge whether to adjust the area of the pixel grid according to a preset threshold.

[0039] In this step, for a multi-classification problem, the category can be judged correct by setting a threshold, and this threshold represents the proportion of the number of any category in the total number. That is to say, when exceeding the preset threshold, it is judged that the area of the pixel grid needs to be adjusted, that is, adjusted to the majority area.

[0040] S103: Based on the pixel grids identified as the defect contour area, draw the defect contour of the metal AM component to be detected.

[0041] In specific implementation, the defect contour of the metal AM component to be detected can be drawn in the following manner: Step 1031: Determine the positional relationship between the pixel grids in the area recognized as the defect contour area.

[0042] Among them, the positional relationship includes directly adjacent, diagonally adjacent, and non - adjacent.

[0043] Step 1032: Connect the pixel grids of the directly adjacent defect contour areas in sequence, and connect the pixel grids of two diagonally adjacent defect contour areas to obtain the defect contour of the metal AM component to be detected.

[0044] As an example, please refer to Figure 7 the defect contour drawing effect diagram of a pothole defect sample as shown. In the figure, the pixel grids recognized as the defect - free area, defect contour area, and defect internal area are marked orange, red, and green respectively for easy observation and distinction. Of course, other markings beneficial for observation and distinction can also be made for different areas, and the present application does not make any limitation here.

[0045] In addition, after obtaining the defect contour of the metal AM component to be detected, the method further includes: Step 1033: Determine whether the defect contour forms a closed area.

[0046] Step 1034: If the defect contour does not form a closed area, use the middle position of the adjacent pixel grids of the defect - free area and the defect internal area in the open area of the defect contour as the connection point to complete the defect contour.

[0047] As an example, please refer to Figure 8 the defect contour drawing effect diagram of another pothole defect sample as shown. The same color is used to mark each area in the figure. In addition, the position indicated by the arrow is the middle position of the pixel grids of the defect - free area and the defect internal area. Figure 7

[0048] The above content is only the preferred embodiment of the present invention. For those of ordinary skill in the art, based on the idea of the present technical content, many changes can be made in the specific implementation manner and application scope. As long as these changes do not deviate from the concept of the present invention, they all fall within the protection scope of the present invention.​

Claims

1. A method for detecting defect contours of metal AM components, characterized in that: The method comprises: The defective surface of the metal AM component to be inspected is divided into several pixel grids of the same size, and laser scanning is performed using a matrix grid scanning method to obtain the LIBS spectrum of each pixel grid, and then a dimensionality reduction process is performed; The LIBS spectrum of each pixel grid after dimension reduction processing is input into the grid classification model corresponding to the defect type of the metal AM component to be detected, so as to identify the area corresponding to each pixel grid; wherein the area includes: a defect-free area, a defect contour area and a defect internal area; Based on the pixel grid identified as the defect contour area, the defect contour of the metal AM component to be inspected is drawn.

2. The metal AM component defect contour detection method according to claim 1, characterized in that: The side length of the pixel grid, the diameter of the laser spot, and the step length of the laser scanning are all the same.

3. The metal AM component defect contour detection method according to claim 1, characterized in that: The grid classification model corresponding to each defect type is obtained in the following way: Using a metal plate of the same material as the metal AM component to be inspected, defect samples of different sizes under each defect type are prepared respectively; wherein the defect types include: pit defects, crack defects, and convex defects; the sizes are defect depth and defect width; For each defective sample, several pixel grids of the same size are divided and each region is determined. LIBS spectra of multiple pixel grids are randomly collected in each region and dimensionality reduction is performed. For each defect type, the LIBS spectra collected from different areas of each defect sample under this defect type and processed with reduced dimension are used to train the CNN model to obtain the grid classification model corresponding to this defect type.

4. The metal AM component defect contour detection method according to claim 1, characterized in that: After identifying the areas corresponding to the pixel grids, the method further includes: For each pixel grid, the KNN concept is used to monitor the number of neighboring pixel grids of the pixel grid that are identified as regions; Whether to adjust the area of ​​the pixel grid is determined according to a preset threshold.

5. The metal AM component defect contour detection method according to claim 1, characterized in that: The step of drawing the defect contour of the metal AM component to be inspected based on the pixel grid identified as the defect contour area comprises: Determine the positional relationship between pixel grids identified as defect contour areas; wherein the positional relationship includes directly adjacent, diagonally adjacent, and non-adjacent; The pixel grids of directly adjacent defect contour areas are connected in sequence, and the pixel grids of two diagonally adjacent defect contour areas are connected to obtain the defect contour of the metal AM component to be inspected.

6. The metal AM component defect contour detection method according to claim 5, characterized in that: After obtaining the defect profile of the metal AM component to be inspected, the method further includes: Determining whether the defect contour constitutes a closed area; If the defect contour does not constitute a closed area, the middle positions of the adjacent pixel grids identified as the non-defective area and the defect internal area in the open area of ​​the defect contour are used as connection points to complete the defect contour.

7. The metal AM component defect contour detection method according to claim 1, characterized in that: The LIBS spectrum was subjected to dimensionality reduction by the following method: For each LIBS spectrum, after preprocessing by baseline removal, noise reduction and smoothing, and maximum and minimum standardization, the principal component analysis method was used to select the first 18 principal components with a total explained variance greater than 99% to characterize the characteristics of the LIBS spectrum, so as to obtain the LIBS spectrum after dimensionality reduction.

Images