Method for detecting defects during laser additive manufacturing, data processing device for carrying out this method, computer program and storage medium

Abstract

In the method B2 for detecting defects during laser additive manufacturing, B21) a first image I INT B23) a first image is captured in the infrared spectrum of the top surface of the powder layer exposed to laser scanning, and B24) the first image is processed using a first convolutional neural network AE1 of the self-encoding type to generate a defect mask Mi indicating the locations of defects in the top surface of the powder layer. A method for manufacturing a part, wherein the presence of defects is detected during manufacturing using the above method. A data processing device, a computer program, and a storage medium for implementing this method.

Description

[Technical Field]

[0001] This invention relates to a field of manufacturing quality control methods in special cases of laser additive manufacturing. [Background technology]

[0002] Figure 1 schematically illustrates a method for selective melting of a powder bed.

[0003] This method is carried out using a selective melting apparatus 100 for a powder bed (in this case, metal powder). The apparatus 100 includes a manufacturing plate 110, parts 120 to be manufactured layer by layer on the manufacturing plate 110, a powder reserve plate 130, and a laser source 140. Plates 110 and 130 have actuators 112 and 132 that can be controlled in a manner that controls the respective height positions of the platform.

[0004] The device 100 further includes a controllable mirror 150 and a control device 160 configured to drive the orientation of the mirror and the height of the platforms 110 and 130.

[0005] The manufacturing of parts involves sequentially producing parts layer by layer.

[0006] The alloys that make up the components are initially in powder form. The following steps are performed in each manufacturing cycle:

[0007] The scraper 170 deposits a powder layer, and the laser beam 142 emitted by the laser source 140 is directed by the mirror 150 in such a way that it selectively scans the deposited powder layer, melting the powder particles in areas that will become part of the part (these areas are defined based on a 3D digital definition of the part to be manufactured).

[0008] At the point 144 where the laser beam strikes the powder layer, the temperature can reach 2000°C, melting the upper powder layer and one or more of the lower layers, thus creating a localized liquid chamber. The solidification of the successive layers gradually forms the component.

[0009] As manufacturing progresses, the tray of preliminary powder 130 rises, allowing the scraper 170 to deposit a new powder layer, while the manufacturing plate 110 descends by an increment equivalent to the thickness of the molten layer.

[0010] During this manufacturing process, defects can occur for a variety of reasons, including laser beam focusing defects, excessive melting at the laser beam collision point, and the presence of impurities in the powder.

[0011] Therefore, it is necessary to detect these defects as quickly as possible and ensure the quality of the manufactured parts so that appropriate corrective actions can be taken.

[0012] Traditionally, defect detection has been performed by taking photographs of the surface of the powder layer scanned with laser light. These photographs are then interpreted by operators who specialize in ongoing manufacturing and are responsible for detecting whether the manufacturing is proceeding normally, as well as whether defects have occurred. [Overview of the project] [Problems that the invention aims to solve]

[0013] This control method is, naturally, expensive and prone to errors. Therefore, there is a need for a more reliable and less expensive method of detecting defects during manufacturing using selective laser melting. [Means for solving the problem]

[0014] According to a first aspect of this disclosure, a method is provided for detecting defects during laser additive manufacturing, which enables this need to be met.

[0015] This method is implemented by a computer. This delicious, B21) A first image is captured, the first image being an image captured in the infrared spectrum of the upper surface of a powder layer exposed to laser scanning, step, B23) The first image is processed using a first autoencoded convolutional neural network in such a way as to generate a defect mask indicating the location of the upper surface of the powder layer.

[0016] In certain embodiments, step B23, which processes the first image, is performed by a relatively simple neural network, for example, by a first neural network of the autoencoder type as defined above. However, in certain embodiments, step B23, which processes the first image, is performed by a set of neural networks comprising the first neural network defined above, plus one or more additional neural networks, particularly of the autoencoder type.

[0017] Processing step B23 can be performed during or after the manufacturing of a part (or a set of parts).

[0018] A defect mask, as used herein, refers to a single image or a set of images (tensor), where each of the images in the mask represents the location of one or more types of defects revealed on the upper surface of the powder coating. These defects or defects can be represented, for example, within the mask image or within each of the images by a specific color, a set of specific colors, or a specific range of colors.

[0019] Each image in the defect mask may be, for example, a grayscale image (showing the probability or significance of the presence of a defect at the location of each pixel), or a binary image, where applicable.

[0020] Types of defects that tend to be represented by defect masks include, for example, insufficient melting, combustion areas, or contamination of parts due to the presence of impurities in the powder layer.

[0021] In certain embodiments, several types of defects are represented in a single image, for example, in a defect mask image, where the color or range of colors corresponds to a specific type of defect.

[0022] In certain embodiments, different images of the mask represent the locations of different types of defects.

[0023] In a particular embodiment, during the process of applying the first image to the first convolutional neural network and generating the defect mask, the first image is transmitted only to the input of the first convolutional neural network.

[0024] It has been found that an autoencoded convolutional neural network, based on at least one image captured in the infrared spectrum of a laser-scanned surface, which is transmitted only to this input of the neural network, can process the at least one image in such a manner that it enables the detection of defects occurring during laser additive manufacturing.

[0025] In other words, such a neural network can effectively process the information contained in the first image, and as a result, it is pointless to refer back to the first image to obtain a defect mask.

[0026] In certain embodiments, only the items of information used to generate the defect mask are the items of information supplied as input to the first neural network, and these items of information are supplied only as input to the first neural network. Therefore, in these embodiments, it is not possible to consider any auxiliary items of information in addition to those supplied as input to the first neural network for calculating the defect mask. Thus, this process considers only the information supplied as input to the first neural network, particularly the information including the first image.

[0027] The surface exposed to laser scanning is, naturally, the surface of the powder layer deposited by the scraper, which is scanned by a laser beam so that the powder melts in a specific predetermined area so that it is incorporated into a part of an already manufactured component.

[0028] The term "laser scan" refers to an action that uses a laser beam to scan or move across all or part of a surface, with the point of collision moving across the surface.

[0029] The efficiency of defect detection using neural networks can be increased by implementing all or some of the following improvements.

[0030] In one embodiment, the first image is an integrated image of the top surface. The term “integrated image” as used herein refers to an image in which, for each pixel, the intensity of that pixel represents the cumulative luminous energy received by that pixel during the capture period, particularly during a capture period longer than 0.1 seconds. In this way, the luminous energy is integrated over time, the resulting value is assigned to the pixel in question, and an integrated image is formed.

[0031] Different architectures may be conceivable for the neural network or multiple neural networks involved in the processing steps.

[0032] In one embodiment, in processing step B23, the defect mask is generated directly by a convolutional neural network.

[0033] The representation directly generated by a neural network (in this case, a convolutional type) means that the defect mask is the output data of the convolutional neural network. Therefore, in this case, the defect mask cannot be obtained by applying processing to the output data of this neural network that involves additional data or information other than what is supplied as input to this neural network (such as data accompanied by other images).

[0034] Furthermore, in this case, it is impossible to obtain the defect mask by applying further processing to the output data of the neural network, which involves the data supplied as input to the neural network at this stage.

[0035] Nevertheless, the direct production of output masks by neural networks does not preclude one or more operations—equalization, formatting, or thresholding—that are performed by the neural network to generate the output data, including the defective mask.

[0036] In certain embodiments, the defect mask is generated directly by the first neural network. In other words, the first neural network is configured to directly generate the defect mask as an output.

[0037] In certain embodiments, one or more additional images other than the first image shown above can be used to obtain a defect mask.

[0038] For example, in the embodiment, during the capture step B21, a second image of the upper surface exposed to the laser beam is also captured, this second image being an image of the upper surface captured in the infrared spectrum, where for each pixel, the pixel intensity represents the maximum light energy received by the pixel during the capture period, and during the processing step B23, the processing performed takes the second image as input in addition to the first image.

[0039] In one embodiment, during the capture step B21, a third image of the upper surface exposed to the laser scan is also captured, the third image being an image of the upper surface captured in the visible spectrum, and during the processing step B23, the processing performed takes the third image into consideration as input in addition to the first image.

[0040] In one or another embodiment of the two prior embodiments, the first convolutional neural network may be configured to receive a first image and at least one first additional image from the second and third images as input.

[0041] In this case, preferably, the defect mask is generated directly by the first convolutional neural network.

[0042] However, more complex architectures can also be used.

[0043] Accordingly, in a modification of the first set of two prior embodiments, the processing performed during processing step B23 is carried out using a set of neural networks, the first convolutional neural network as input and the output neural network as output, which is in particular an autoencoded type, and is configured to take the output of the first convolutional neural network as input and generate a defect mask (preferably directly) based on this input.

[0044] Advantageously, in this first variant, the first convolutional neural network is sufficient to process several (at least two) types of images supplied as input.

[0045] Conversely, in a modified example of the second set of one of the two prior embodiments, during processing step B23, The inputs include the first convolutional neural network and at least one autoencoded additional neural network configured to receive a first additional image from the second or third image as input. The processing is performed using a set of neural networks, which include, as an output, an autoencoded output neural network, configured to take the output of a first neural network and the output of at least one additional first neural network as input and generate the defect mask as an output.

[0046] In a particular variation of the second set of variations defined above, the neural network of this set further comprises a second autoencoder-type additional neural network configured to take as input a second additional image which is from the second and third images other than the first additional image.

[0047] In one embodiment, at least one of the above set of autoencoded neural networks and, for example, the first neural network in particular, comprises a connection linking a neural layer at index y to a neural layer at index ny, and the total depth of at least one of the neural networks in question is equal to n.

[0048] In one embodiment, during the manufacturing of a part by powder bed laser melting, the capture step B21 is performed using a photographic camera or a video camera.

[0049] Furthermore, the present disclosure also relates to a method for manufacturing a component by powder bed laser melting, wherein at least one component is manufactured by powder bed laser melting, and defects are detected during or after the manufacture of the at least one component by employing the aforementioned method. The powder used in this method may be a powder of any composition, in particular a metal powder.

[0050] In certain embodiments, different steps in a method for detecting defects during laser additive manufacturing are determined by computer program instructions.

[0051] Accordingly, the disclosure also relates to a computer program comprising instructions that, when the program is executed by at least one processor, direct at least one processor to perform the steps of the method described above. The program may use any programming language and may be in the form of source code, object code, intermediate code between source code and object code such as a partially compiled form, or any other preferred form.

[0052] Furthermore, this disclosure relates to a computer-readable non-volatile storage medium in which the above-mentioned computer program is stored. This information medium may be any entity or device capable of storing the program. For example, the medium may include storage means such as a CD-ROM or a ROM of a miniature electronic circuit, or other magnetic recording means such as a diskette (floppy disk) or a hard disk. Alternatively, the information medium may be an integrated circuit into which the program is incorporated, which is suitable for or used to perform the method.

[0053] Furthermore, the present disclosure relates to a data processing device comprising at least one processor and a memory in which several instructions are recorded, the memory in which, when these instructions are executed by at least one processor, the instructions cause at least one processor to perform one of the steps of the aforementioned method.

[0054] This data processing device (or part thereof) can be integrated into a system for manufacturing parts by metal powder bed laser melting.

[0055] Accordingly, the present disclosure also includes an apparatus for manufacturing a component by powder bed laser melting and a system for manufacturing a component by powder bed laser melting, which comprises a data processing apparatus as described above. [Brief explanation of the drawing]

[0056] [Figure 1] Figure 1 (already shown) is a schematic diagram of a conventional apparatus for manufacturing parts by powder bed laser melting. [Figure 2] Figure 2 shows a system for manufacturing parts by powder bed laser melting according to the present disclosure. [Figure 3] Figure 3 shows a first neural network used to detect defects in the method for detecting defects according to this disclosure. [Figure 4] Figure 4 shows a first set of neural networks used to detect defects in the method for detecting defects according to the present disclosure. [Figure 5] Figure 5 is a diagram of the integrated image calculated during the implementation of the method according to this disclosure. [Figure 6] Figure 6 is a diagram of the maximum image calculated during the implementation of the method according to this disclosure. [Figure 7] Figure 7 is a schematic diagram of a defect mask obtained by carrying out the method according to this disclosure. [Figure 8]Figure 8 is a flowchart illustrating the steps of the method for manufacturing a part by selective powder bed melting according to this disclosure. [Modes for carrying out the invention]

[0057] As an example, the method and system for production by selective powder bed melting according to this disclosure are described here in reference to Figures 2 to 8.

[0058] These manufacturing methods can be carried out using a laser-based component manufacturing system 1000, as shown in Figure 2, which includes an apparatus 100 for manufacturing components by powder bed laser melting and a data processing device 200.

[0059] This apparatus 100 is substantially identical to the manufacturing apparatus 100 described in relation to Figure 1, and for this reason, components that are identical or substantially identical in the two apparatuses have the same reference numerals as in Figures 1 and 2. A special feature of apparatus 100 in system 1000 is that, compared to apparatus 100 in Figure 1, it also has a capture device 180.

[0060] The capture device 180 has two video cameras, one capable of capturing images in the infrared spectrum and the other in the visible spectrum. The images captured by these cameras are transmitted to the data processing device 200.

[0061] The data processing device 200 has a computer hardware architecture, as schematically shown in Figure 2. In general, any data processing device can be used, comprising at least one memory capable of recording data, a program as described below, and one or more processors capable of executing this program. The data processing device may be located near the device 100, or conversely, remotely, and may be accessible via a network, such as the Internet.

[0062] In this embodiment, the data processing device 200 includes, in particular, a processor 201, a read-only memory 202, a non-volatile flash memory 203, and communication means 204 with other components of the system 1000, including, in particular, a control device 160.

[0063] The non-volatile flash memory 203 of the data processing device 200 is readable by the processor 201 and constitutes a recording medium according to the disclosure on which a computer program according to the disclosure is recorded, which has instructions for performing steps of a method for manufacturing a part by powder bed laser melting (in particular, including the step of a defect detection subprogram).

[0064] This program can take on different forms. In the first embodiment (Figure 3), the program constitutes a first neural network NN1 within the scope of the meaning of this disclosure. In a more complex second embodiment (Figure 4), the program constitutes a pair of convolutional neural networks NN2.

[0065] The network NN1, which constitutes the first convolutional neural network according to this disclosure, is schematically represented in Figure 3.

[0066] In this embodiment, a neural network NN1 is used to detect three types of defects based on the input image.

[0067] Network NN1 is a self-encoder having an encoder E, a decoder D, and three normalization layers F1, F2, and F3 in sequence.

[0068] In this embodiment, network NN1 is a U-net type self-encoder. A U-net type network is a self-encoding type network that specifically has connections C that directly connect the encoder E blocks and the decoder D blocks. Generally, the block at position i is connected to the block at position ni, where n is the total number of blocks.

[0069] The architecture of U-net type networks is described in particular in the publication "U-net: Convolutional Networks for Biomedical image segmentation" by Olaf Ronneberger, Philipp Fischer, and Thomas Brox (International Conference on Medical image computing and computer-assisted intervention, Springer, Cham, 2015).

[0070] Encoder E has several consecutive convolutional blocks E10, E20, E30, E40, and E50 (generally 3 to 5 convolutional blocks). Each of these convolutional blocks has several consecutive layers of an artificial neural network. In each of these blocks, the upstream layers (usually two or three in number) are convolutional layers, and the last downstream layer is an undersampling (or "pooling") layer, thus reducing the resolution of the processed data.

[0071] Here, the convolutional layer is shown as the layer that performs the convolution operation, followed by the batch normalization process, and then the ReLU (Modified Linear Unit) correction step.

[0072] Similar to encoder E, decoder D also has several consecutive convolutional blocks D10, D20, D30, D40, and D50. Each of these convolutional blocks has several consecutive artificial neural layers. In each of these blocks, the downstream layers (usually two or three in number) are convolutional layers, and the last downstream layer is an oversampling (or "upsampling") layer, thus increasing the resolution of the processed data.

[0073] Of course, for the first neural network with an architecture other than that of the network NN1, it can be used within the scope of the present disclosure.

[0074] As input, for each iteration of the index i, the encoder E receives the input data D IN i This data is processed by the encoder E that generates an output as intermediate data D INTERM i and this data D INTERM i is transmitted as input to the decoder D. The decoder D processes this data and supplies the output data D OUT i to each of the normalization layers F1, F2, and F3.

[0075] Each of these normalization layers applies a functional SoftMAX or a normalization exponential function to the data D [[ID=--]] OUT [[ID=--]] i received from the decoder D and generates it as the final image (M1i, M2i, M3i respectively). The images M1i, M2i, and M3i together form the defect mask Mi. In these images, each pixel has a value representing an estimate (or prediction) of the probability of a defect at the pixel corresponding to the position on the surface of the target layer for the defect associated with the target normalization layer.

[0076] In this embodiment, therefore, the defect mask Mi is directly generated by the first neural network NN1.

[0077] Optionally, a thresholding (also called binarization) function can be applied to all or part of the mask Mi obtained at the output of the normalization layer to facilitate the processing or interpretation of the resulting mask. The thresholding function consists of comparing the value of each pixel with a predetermined threshold for each pixel, and thus assigning a binary value 0 or 1 to the pixel according to whether the pixel is smaller or larger than the threshold respectively.

[0078] Input Data D IN i This is simply an integrated image I of a powder layer exposed to laser scanning. INT It can be done.

[0079] However, Data D IN i This can also be a tensor obtained by concatenating two or more images, in particular, the integrated image I INT , and, where applicable, maximum image I MAX and / or visible image I VIS That is the case.

[0080] By considering two or more images as input, the performance of the neural network NN1, i.e., the accuracy of the defect mask it predicts, can be improved. In this embodiment, the first neural network NN1 uses different input images (integrated image I) INT , and maximum image I MAX and / or visible image I VIS It turns out to be advantageous to be sufficient to handle ).

[0081] In a preferred embodiment, as shown in Figure 3, based on these input data, the first neural network NN1 directly generates an output mask Mi as an output.

[0082] A set of neural networks NN2 (or simply "NN2") that constitute the second embodiment are schematically shown in Figure 4.

[0083] NN2, according to its embodiment, takes one to three autoencoded convolutional neural networks (AE) as input. INT , AE MAX , AE VIS ) and, as output, an autoencoded neural network AE OUT It is a set of neural networks that are equipped with the following: Network AI INTIn the context of this disclosure, the first neural network constitutes an arbitrarily selected network AE. MAX , AE VIS These form the first and second additional networks. Output self-encoders AE located downstream of one or more input self-encoders. OUT This architecture, which incorporates [specific features], can improve the prediction accuracy of neural networks.

[0084] When a network NN2 can process two or more images as input, rather than just one, its performance improves compared to an architecture where the network NN2 can only consider a single image as input.

[0085] When considering two or more images as input, in one embodiment, the neural network NN2 is a first neural network, preferably a neural network AE INT (Therefore, Network AE) MAX and AE VIS The input is only (excluding ), and the output is the neural network AE. OUT This is the first neural network (for example, AE INT It takes the output of ) as input and directly generates a defect mask Mi as output.

[0086] In other embodiments, as shown in Figure 4, the neural network NN2 has at least two neural networks as input (three networks AE, if represented as such) INT , AE MAX , AE VIS (Having), Neural Network AE OUT It has the following as its output. The input to the input neural network consists of different images received as input.

[0087] The output of different input neural networks is the output neural network AE. OUT It is forwarded to generate the output mask Mi.

[0088] The defect mask Mi is used in the output neural network AE. OUT It is preferable that it be produced directly by [this method].

[0089] In this case (Figure 5), NN2 produces three images, i.e., integrated image I, for each iteration (of index i). INT i Maximum image I MAX i Visible spectrum I VIS i Receive the captured image.

[0090] Neural Network AE INT , AE MAX , AE VIS , AE OUT Each of these has the same or nearly the same architecture as the neural network NN1 shown in Figure 3. However, network AE INT , AE MAX , AE VIS Therefore, the presence of one (or more) normalization layers downstream of the network is optional.

[0091] Three input network AE INT , AE MAX , AE VIS To combine the data generated by these, this pair NN2 generates image I, which is produced as the output by the decoders of these different networks. INT_OUT i , I MAX_OUT i , I VIS_OUT i It has a single neural network G that plays the role of connecting them.

[0092] Finally, in this pair of neural networks NN2, the output neural network AE OUThe model further has a normalization layer F as an output. This normalization layer can be arranged in a similar manner to the normalization layer F described in relation to Figure 3, for example, by arranging several SoftMAX layers in parallel and generating different images as shown in the output, each representing the probability of a defect present at each point (or at each pixel), and each image representing the defect probability for one of the defects being searched for.

[0093] Thus, in this embodiment, these different images are output to the neural network AE. OUT This constructs the output mask Mi that is directly generated by this.

[0094] Manufacturing of parts The manufacturing of the component is carried out by repeatedly performing the different steps shown in Figure 8 for each layer of the component.

[0095] During the iteration of index i, the manufacturing of layer n°i of the component is carried out by performing the following steps:

[0096] In the first stage A, a new powder layer is deposited using the scraper 170. For this purpose, the powder preparation plate is raised, and conversely, the production plate is lowered from a corresponding height.

[0097] Next (step B1), the laser scan is performed in such a manner that it scans all points in the target layer that must be part of the part 120 to be manufactured. The laser scan generates local melting of the powder and bonds the crystal grains of the powder to the underside of the already formed part in the area exposed to the laser scan.

[0098] In parallel, the data processing device executes a program in a manner that performs a method for detecting defect B2.

[0099] Method B2, which helps detect defects occurring during the manufacturing process in System 1000, has the following steps:

[0100] B21) Different series or rows of images of the surface scanned by the laser beam are captured in parallel. More precisely, during the duration of the laser scan B1, each of several cameras of the capture device 180 captures a series of images. In the embodiment described, a first camera operating in the visible spectrum captures a series of images in the visible spectrum, and a second camera operating in the infrared spectrum captures a series of images in the infrared spectrum. Each of these images represents the entire surface scanned by the laser beam 142 (or the capture may be limited to the region surrounding the laser beam collision point 144).

[0101] The infrared image captured is from an optical tomography sensor. In this case, the sensor is an infrared S-CMOS sensor with a resolution of 2000 x 2000 pixels. The near-infrared image generated by the sensor represents the temperature field at the surface of the layer.

[0102] The captured images are continuously captured throughout the entire laser scan of the layer.

[0103] B22) The captured series of images are preprocessed using the data processing device 200, and based on the captured images, the data processing device 200 calculates three images. Image I INT_ i is the combined image of the upper surface, which is combined during the combined image acquisition period while layer n°i is being fabricated. In this embodiment, the acquisition period of the combined product image is equal to the duration of the laser scan on layer n°i. It could instead be a set duration comprising the moment of impact of the laser beam at the point of interest. Second image I MAX_ i is an image representing the maximum light energy received by each pixel during the capture period of the maximum image during the fabrication of layer n°i. In this embodiment, the capture period of the maximum image is equal to the duration of the laser scan. It could instead be a set duration encompassing the moment of collision of the laser beam at the point of interest. Third image I VISThis is an image of the top surface of layer n°i, scanned with laser light and captured in the visible spectrum during the manufacturing of layer n°i. (This image is an image of Image I, which was optionally captured in the visible spectrum during the duration of the laser scan.) VIS_ Integrated image I calculated by integration (integration) (or averaging) of i VIS_INT_ This can be expressed as i.) This process is performed by the data processing device 200, which processes three images I INT_ i, I MAX_ i and I VIS_ Capture i. Image I INT_ i and I MAX_ i is shown in Figures 5 and 6, respectively.

[0104] Therefore, the first image I is an integrated image (as shown in Figure 5 for example). INT In _i, the values ​​of each pixel are given as follows:

number

[0105] The second image I is the largest image (an example is shown in Figure 6). MAX In _i, the values ​​of each pixel are given as follows:

number

[0106] This maximum image can be interpreted as the envelope of the intensity of infrared emission during laser scanning of the layer. Note that L represents the overlapping line between adjacent parallel bands (or "lasing bands") exposed to laser scanning, and I represents the integrated image. INT_ The white line in i (Figure 5) is the maximum image I MAXIt is not so visible in _i (FIG. 6). Due to this characteristic, the use of the maximum image is particularly effective in limiting the number of false positives. Specifically, when only the integrated image is used, the neural network tends to assign excessive importance to these overlapping lines BL. Thus, the information contributed by the maximum image seems to help the neural network not to interpret the overlapping lines between adjacent laser oscillation strips as defect regions. Therefore, the quality of the defect mask generated by this pair of neural networks NN2 can be improved.

[0107] B23) Three images I INT_ i, I MAX_ i, I VIS_ i is processed using this pair of neural networks NN2.

[0108] These images are supplied as inputs to NN2, and in iteration n°i of the algorithm, the network AE INT , AE MAX , AE VIS receives, respectively, the integrated image I INT _i, the maximum image I MAX _i and the visible image I VIS _i as inputs. Based on these, the network AE INT , AE MAX , AE VIS generates, respectively, the output images I INT _ OUT i , I MAX _ OUT i , I VIS _ OUT i .

[0109] Next, these output images are concatenated by a concatenation layer G that forms a third-order tensor. This tensor is supplied as an input to the output neural network AE OUT .

[0110] As an output, the output neural network AE OUTThis generates a defect mask Mi. In this embodiment, the paired neural network NN2 is configured to generate a defect mask, which is image Mi, as an output. This image Mi is image I INT i, I MAX i and I VIS It has the same dimensions as i and indicates the location of the defect in the image.

[0111] Alternatively, the data processing performed in step B23 can be carried out using neural network NN1 (configured in a way that is appropriate for the number of images to be considered as input) instead of neural network NN2.

[0112] Figure 7 shows a binary mask obtained by applying a thresholding function to the mask Mi. Therefore, each pixel can only have a value of 0 or 1. Thus, pixels located at the location of defects are represented in white, while pixels in defect-free regions are represented in black. Note that the defects clearly visible in Figure 7 also appear on the right side of the integrated image in Figure 5.

[0113] Three images I INT _i, I MAX _i, I VIS The joint use of _i is a combination that can obtain the best performance using a set of neural networks NN2. However, the first set of neurons (configured and consequently trained) according to this disclosure, for example, network NN1, is based only on infrared-captured images, in particular on integrated image I INT A defect mask can be generated based on _i.

[0114] B24) Based on the defect mask Mi, measures are determined for the remaining defects. These measures may also be determined based on the defect mask Mi of the step in progress and one or more defect masks obtained in the previous step. Depending on the defect(s), it may be decided to continue manufacturing without making any corrections (and thus to start manufacturing the next layer n°i+1 of the part), to continue manufacturing by correcting one or more operating parameters of the manufacturing apparatus 100, or to stop manufacturing the part. In the second case, the corrected parameters may be the displacement velocity of the laser beam impact point, the power of the laser beam, the thickness of the deposited powder layer, etc.

[0115] Based on the decision made, an order is determined and sent to the control device 160 for the remainder of the manufacturing process.

[0116] Driving neural networks Training a neural network is carried out in a known manner using a database of training data. This database contains input and output data, the output data being the output mask (ground truths) that the neural network, or a set of neural networks if applicable, is expected to generate when the input data is fed to it as input.

[0117] Depending on the number of images that a neural network or set of neural networks can accept as input, the input data can be an image, a pair of images, or n sets of images. For example, for training NN2, the input data is three images I calculated by performing step B22 of this method. INT _i, I MAX _i and I VIS It is _i.

[0118] The following explanation is given when a neural network or set of neural networks is configured to receive three images as input. However, it will be understood that this disclosure is naturally applicable regardless of the number of images received as input by the neural network or set of neural networks.

[0119] The desired output mask ("ground truth") can be a mask that shows different types of defects that the neural network (NN) must identify, such as defects at high temperature points (points where the temperature reaches excessive values), insufficient melting, and contamination of the melting bath. These defect masks can be composed of images (binary images where applicable) prepared by experts in the manufacturing of the relevant parts, and the manufacturing defects identified by these experts are then processed from three input images. INT _i, I MAX _i and I VIS It can be expressed based on _i.

[0120] In one embodiment, a neural network (or a set of neural networks) is driven to evaluate the presence of a specific defect in each image of a defect mask.

[0121] In this case, the training database contains one set of three input images (I) for each defect. INT _i, I MAX _i and I VIS i) has, for each of the three input images, a corresponding output image (selectively binary). This output image represents the best possible estimate of the likelihood of a defect, which is considered for each of several pixels in the image (and thus for each corresponding point in each layer that has been laser scanned) for the three corresponding input images.

[0122] In this case, the image obtained after training as the output of the neural network (as the output of the normalization layer) is a probabilistic image, which represents the probability of the target defect existing for each pixel.

[0123] In another embodiment, a neural network is trained to evaluate the presence of several defects at once with respect to at least one image of a defect mask. In this way, the neural network is trained to predict the presence or absence of several types of defects, such as under-melting, high-temperature points, or contamination of the component due to the presence of impurities in the powder layer, via the generated output image.

[0124] In this case, the training database then applies a defect mask to the target image using a set of three input images (I INT _i, I MAX _i and I VIS i) and each of the three images have associated output images. These output images are then images in which each pixel is assigned a specific value (or color) associated with a specific type of defect identified at the pixel's location. For example, these output images may consist of pixels having one of the following values: 0 (no defect), 1 (melting defect), 2 (high temperature location), or 3 (part contamination).

[0125] To enrich the training database, additional images can be added using so-called "augmentation" methods. These images can be generated by applying effects such as rotation, image shifting, horizontal flipping, and vertical flipping to the initial images in the database.

[0126] The cost function used can, in particular, be the binary cross-entropy function.

[0127] Training can be performed end-to-end (end-to-end learning) in all proposed architectures, whether it is a simple architecture consisting of only a first neural network NN1, or a complex architecture consisting of a set of multiple neural networks NN2.

[0128] Furthermore, the image dimensions can be optimized. Images supplied as input to the neural network can be computed by undersampling. For example, images generated by an optical tomography sensor may have high resolution (2000 x 2000 pixels or more), but for the implementation of this method, only lower resolution images, such as 1000 x 1000 pixels, are used.

[0129] If the images supplied as input to the neural network have high resolution and undersampling is avoided, another practical technique involves splitting several input images (initially acquired at "high resolution") into lower-resolution images, each with a resolution that matches the processing power of the neural network used. These different sub-images are then processed by the neural network to obtain corresponding output masks. These output masks, having the same dimensions or resolution as the sub-images, are then recombined, making it possible to obtain a high-resolution output mask with the same resolution as the initially captured image. For example, an input image with a resolution of 2000x2000 pixels can be split into 16 sub-images of 500x500 pixels each. To obtain an output mask for the initial image with a resolution of 2000x2000 pixels, output masks are created based on the different sub-images.

[0130] During network training, different architectural parameters of the neural network are considered, particularly the following parameters, namely the dimensions of the input data to the neural network (i.e., input image I). INT , I MAX and I VIS (Size), Network AE INT , AE MAX , AE VIS and AE OUTA search for the best architecture is provided to automatically determine the depth of each convolutional block (number of convolutional blocks), the dimensions of the intermediate data or "feature map" in different convolutional blocks of different neural networks (i.e., the number of neurons per layer), or the number of convolutional layers in different convolutional blocks. This search can be performed using grid analysis (grid search) in particular.

[0131] While the present invention has been described with reference to specific exemplary embodiments, it is evident that different modifications and changes can be made to these examples without departing from the general scope of the invention as defined by the claims. Furthermore, individual features of the different embodiments described may be combined in further embodiments. For example, this method may involve a set of neural networks processing image I INT i By supplying only one image as input, or by supplying three images I INT _i, I MAX _i and I VIS Image I instead of _i INT _i and I MAX This can be achieved by supplying only _i. Therefore, the specification and drawings should be considered in an illustrative sense, not in an restrictive sense. Furthermore, this disclosure includes the following inventions. The first aspect is, In a method for detecting defects during laser additive manufacturing (B2), The above method (B2) is, B21) First image (I INT Step i) captures the first image, which is an image captured in the infrared spectrum of the upper surface of the powder layer exposed to laser scanning, B23) A method (B2) for detecting defects during laser additive manufacturing, comprising the steps of: processing the first image using a first autoencoded type convolutional neural network (NN1, NN2) to generate a defect mask (Mi) indicating the location of defects on the upper surface of the powder layer. The second aspect is, The first image is the integrated image (I) of the upper surface. INT i) a method for detecting defects in a first embodiment, wherein for each pixel, the intensity of the pixel indicates the cumulative luminous energy received by the pixel during the capture period, particularly during a capture period longer than 0.1 seconds. The third aspect is, During the capture step B21, a second image of the upper surface exposed to the laser beam is also captured, and the second image is the image of the upper surface captured in the infrared spectrum (I MAX i) where, for each pixel, the intensity of the pixel represents the maximum light energy the pixel receives during the capture period, and the processing performed during processing step B23 is the first image (I INT In addition to the above second image (I MAX A method for detecting defects in the first or second embodiment, which takes _i) as input. The fourth aspect is, During the capture step (B21), a third image of the upper surface exposed to the laser scan is also captured, and the third image is the image of the upper surface captured in the visible spectrum (I VIS i) and during the processing step (B23), the first image (I INT In addition to _i), the third image (I VIS This is a method for detecting a defect in any one of the first to third embodiments, where _i) is considered as input. The fifth aspect is, The first convolutional neural network (AE) INT ) comprises the first image and at least one first additional image (I) from among the second image and the third image. MAX 、I VIS A method for detecting a defect in a third or fourth embodiment, configured to accept ) as input. The sixth aspect is, The defect mask (Mi) is a method for detecting defects in a fifth embodiment, which is directly generated by the first convolutional neural network (NN1). The seventh aspect is, During the processing step (B23), the processing takes the first convolutional neural network (AE) as input. INT ) and an output neural network (AE), which is configured to receive the output of the first convolutional neural network as input and generate the defect mask (Mi) based on this input, and is a particularly autoencoded type output neural network as output. OUT This is a method for detecting defects in a fifth aspect, performed using a set of neural networks (NN2) comprising the following: The eighth aspect is, During the processing step (B23), the processing is performed The inputs include the first convolutional neural network and at least one autoencoded first additional neural network configured to receive the second image and a first additional image from the third image as inputs. An autoencoded output neural network (AE) configured to take the outputs of the first convolutional neural network and the at least one additional neural network as inputs and to generate the defect mask (Mi) as an output. OUT A method for detecting defects in a third or fourth embodiment, which is performed using a set of neural networks (NN2) comprising the following: The ninth aspect is, The aforementioned pair of neural networks (NN2) is configured to accept a second additional neural network (AE) of autoencoder type as input, which is selected from the second image and the third image and is not the first additional image. MAX , AE VIS A method for detecting defects in an eighth aspect, further comprising the above. The tenth aspect is, The first convolutional neural network (NN1) is a method for detecting a defect in any one of the first to ninth embodiments, wherein the first convolutional neural network (NN1) comprises a connection linking a neural layer at index y to a neural layer at index ny, and the sum of the depths of at least one of the neural networks is considered to be equal to n. The eleventh aspect is, The capture step (B21) is a method for detecting a defect in any one of the first to tenth embodiments, which is performed using a photographic camera or a video camera during the manufacturing of a part by powder bed laser melting. The twelfth aspect is, In the processing step (B23), the defect mask (Mi) is directly generated by a convolutional neural network, and this is a method for detecting defects in any one of the first to eleventh embodiments. The 13th aspect is, A method for detecting defects in any one of the first to twelfth embodiments, wherein the first image is transmitted only to the inputs of the first convolutional neural network (NN1, NN2) during the process of applying the first image to the first image using the first convolutional neural network and generating the defect mask (Mi). The 14th aspect is, In a method for manufacturing parts by powder bed laser melting, At least one component is manufactured by powder bed laser melting, A method for detecting a defect in at least one of the aforementioned parts during or after its manufacture by employing one of the methods described in the first to thirteenth embodiments. The 15th aspect is, In the data processing device (200), The data processing device (200) is At least one processor, A data processing device (200) comprising: a memory containing instructions that, when executed by the at least one processor, direct the at least one processor to perform a step of the method in any one of the first to thirteenth embodiments. The 16th aspect is, In a system (1000) for manufacturing parts by powder bed laser melting, The aforementioned system (1000) is Apparatus (100) for manufacturing parts by powder bed laser melting, A system (1000) for manufacturing parts by powder bed laser melting, comprising a data processing device (200) according to the 15th embodiment. The 17th aspect is, A computer program, wherein, when the computer program is executed by at least one processor, the computer program comprises instructions that cause the at least one processor to perform a step of the method in any one of the first to thirteenth embodiments. The 18th aspect is, A computer-readable non-volatile storage medium on which a computer program according to the 17th aspect is stored.

Claims

1. In a method for detecting defects during laser additive manufacturing (B2), The above method (B2) is, B21) First image (I INT i) is captured, and the first image is an image captured in the infrared spectrum of the upper surface of the powder layer exposed to laser scanning, step, B23) The first image is processed using a first autoencoded convolutional neural network (NN1, NN2) to generate a defect mask (Mi) indicating the location of defects on the upper surface of the powder layer, comprising the steps of: During the capture step B21, a second image of the upper surface exposed to the laser beam is also captured, and the second image is the image of the upper surface captured in the infrared spectrum (I MAX i) where, for each pixel, the intensity of the pixel represents the maximum light energy the pixel receives during the capture period, and the processing performed during processing step B23 is the first image (I INT In addition to i), the second image (I MAX A method for detecting defects during laser additive manufacturing, which takes i) as input (B2).

2. The first image is the integrated image (I) of the upper surface. INT i) a method for detecting a defect according to claim 1, wherein for each pixel, the intensity of the pixel indicates the cumulative luminous energy received by the pixel during the capture period, particularly during a capture period longer than 0.1 seconds.

3. During the capture step (B21), a third image of the upper surface exposed to the laser scanning is also captured, and the third image is the image of the upper surface captured in the visible spectrum (I VIS i) and during the processing step (B23), the first image (I INT In addition to i), the third image (I VIS A method for detecting the defect described in claim 1, wherein i) is considered as input.

4. The first convolutional neural network (AE INT ), which is configured to receive, as an input, the first image and at least one first additional image (I MAX , I VIS ) selected from the second image and the third image. The method for detecting a defect according to claim 1.

5. The method for detecting a defect according to claim 4, wherein the defect mask (Mi) is directly generated by the first convolutional neural network (NN1).

6. During the processing step (B23), the processing takes the first convolutional neural network (AE) as input. INT ) and an output neural network (AE), which is configured to receive the output of the first convolutional neural network as input and generate the defect mask (Mi) based on this input, and is a particularly autoencoded output neural network as output. OUT A method for detecting the defect according to claim 4, comprising a set of neural networks (NN2) comprising the following:

7. During the processing step (B23), the processing is performed The inputs include the first convolutional neural network and at least one autoencoded first additional neural network configured to receive the second image and a first additional image from the third image as inputs. An autoencoded output neural network (AE) configured to take the outputs of the first convolutional neural network and the at least one additional neural network as input and to generate the defect mask (Mi) as output. OUT A method for detecting the defect described in claim 1, comprising a set of neural networks (NN2) comprising the following:

8. The aforementioned pair of neural networks (NN2) is configured to receive as input a second additional image selected from the second and third images and other than the first additional image, and is an autoencoded second additional neural network (AE). MAX AE VIS A method for detecting a defect according to claim 7, further comprising:

9. A method for detecting a defect according to claim 1, wherein the first convolutional neural network (NN1) comprises a connection linking a neural layer at index y to a neural layer at index n-y, and the sum of the depths of at least one of the neural networks is considered to be equal to n.

10. The method for detecting a defect according to claim 1, wherein the capture step (B21) is performed using a photographic camera or a video camera during the manufacturing of a part by powder bed laser melting.

11. The method for detecting a defect according to claim 1, wherein in the processing step (B23), the defect mask (Mi) is directly generated by a convolutional neural network.

12. A method for detecting defects according to claim 1, wherein, during the process of applying the first image using the first convolutional neural network and generating the defect mask (Mi), the first image is transmitted only to the inputs of the first convolutional neural network (NN1, NN2).

13. In a method for manufacturing parts by powder bed laser melting, At least one component is manufactured by powder bed laser melting, A method for detecting a defect during or after the manufacture of at least one of the aforementioned components by employing the method according to any one of claims 1 to 12.

14. In the data processing device (200), The data processing device (200) is At least one processor, A data processing device (200) comprising: a memory which, when executed by the at least one processor, records instructions that cause the at least one processor to perform a step of the method according to any one of claims 1 to 12.

15. In a system (1000) for manufacturing parts by powder bed laser melting, The aforementioned system (1000) Apparatus (100) for manufacturing parts by powder bed laser melting, A system (1000) for manufacturing parts by powder bed laser melting, comprising a data processing device (200) according to claim 14.

16. A computer program comprising, when the computer program is executed by at least one processor, instructions that cause the at least one processor to perform a step of the method according to any one of claims 1 to 12.

17. A computer-readable non-volatile storage medium storing the computer program described in claim 16.

Images