Method for detecting defects during laser additive manufacturing, data processing device for implementing the method, computer program and storage medium

Abstract

A method (B2) for detecting defects during laser additive manufacturing includes the following steps: B21) capturing a first image (I INT_ i) The first image is an image captured in the infrared spectrum of the upper surface of a powder layer exposed to laser scanning; B23) The first image is processed using a first convolutional neural network (AE1) of the autoencoder type to generate a defect mask (Mi) indicating the location of a defect on the upper surface of the powder layer. A method for manufacturing a part, wherein the presence of a defect is detected by means of the above method. A data processing apparatus, a computer program, and a storage medium for implementing the method.

Description

Technical Field

[0001] This invention relates to the field of manufacturing quality control methods in specific cases of laser additive manufacturing. Background Technology

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

[0003] This method is implemented using a powder bed (in this case, metal powder) selective melting manufacturing machine 100. Machine 100 includes a manufacturing plate 110, a powder storage plate 130, and a laser source 140, on which the part 120 to be manufactured is manufactured layer by layer. Plates 110 and 130 include actuators 112 and 132, which can be controlled in a manner that controls the height position of each platform.

[0004] The machine 100 also includes a controllable reflector 150 and a control unit 160, which are configured to drive the orientation of the reflector and the height of the platforms 110 and 130.

[0005] The manufacturing of parts involves the continuous production of parts layer by layer.

[0006] The alloy used to manufacture the parts is initially in powder form. The following operations are performed in each manufacturing cycle:

[0007] A layer of powder is deposited by a scraper 170;

[0008] The laser beam 142 emitted by the laser source 140 is guided by the mirror 150 to selectively scan the deposited powder layer, thereby melting the powder particles in areas that must be part of the part (these areas are defined based on the 3D digital definition of the part to be manufactured).

[0009] At point 144 where the laser beam impacts the powder layer, the temperature can reach 2000℃, melting the upper powder layer, but also melting one or more lower layers, thus creating a localized liquid bath. The solidification of the continuous layers gradually forms the part.

[0010] As manufacturing proceeds, the powder storage tray 130 rises to allow the scraper 170 to deposit a new powder layer; conversely, the manufacturing plate 110 descends by an increment equal to the thickness of the molten layer.

[0011] During this manufacturing process, manufacturing defects may occur for a variety of reasons, such as laser beam focusing defects, excessive melting at the laser beam impact point, and impurities in the powder.

[0012] Therefore, in order to ensure the quality of the manufactured parts, it is necessary to detect these defects as soon as possible so that appropriate corrective measures can be taken.

[0013] Traditionally, defect detection is accomplished by taking photographs of the surface of a powder layer scanned by a laser beam; these photographs are then interpreted by operators specializing in ongoing manufacturing to determine whether the manufacturing is proceeding normally or, on the other hand, whether defects have occurred.

[0014] This control method is naturally expensive and prone to errors. Therefore, a more reliable and cheaper method is needed to detect defects during manufacturing via selective laser melting. Summary of the Invention

[0015] According to a first aspect of this disclosure, a method for detecting defects during laser additive manufacturing is provided, thereby fulfilling this requirement. The method is implemented by a computer.

[0016] The method includes the following steps:

[0017] B21) Capture a first image, which is an image of the upper surface of the powder layer exposed to laser scanning, captured in the infrared spectrum;

[0018] B23) The first image is processed using a first convolutional neural network of the autoencoder type to generate a defect mask that indicates the location at the upper surface of the powder layer.

[0019] In some implementations, step B23 of processing the first image is performed by a relatively simple neural network, such as a first neural network of the autoencoder type as defined above. However, in some implementations, step B23 of processing the first image is performed by a group of neural networks, including one or more additional neural networks, particularly of the autoencoder type, in addition to the first neural network as defined above.

[0020] Processing step B23 may be completed during the manufacturing of the part (or multiple parts) or subsequently.

[0021] In this document, a defect mask refers to a single image or a set of images (tensors); the image of the mask, or each image, represents the location of one or more types of defects exposed on the upper surface of the powder coating. These one or more defects may be represented, for example, by a specific color, multiple specific colors, or a specific range of colors in the image of the mask, or each image.

[0022] Each image of the defect mask can be, for example, a grayscale image (indicating the probability or salience of the presence of the defect at the location of each pixel) or, where applicable, a binary image.

[0023] The types of defects that may be represented by a defect mask include, for example, incomplete melting, charred areas, or part contamination due to impurities in the powder layer.

[0024] In some embodiments, several types of defects are represented in a single image: for example, in an image of a defect mask, a color or range of colors may correspond to a certain type of defect.

[0025] In some embodiments, different images of the mask represent the locations of different types of defects.

[0026] In some embodiments, during the process of applying the first image to the first image and generating a defect mask using the first convolutional neural network, the first image is only transmitted to the input of the first convolutional neural network.

[0027] It is evident that, based on at least one image of a surface exposed to laser scanning captured in the infrared spectrum, this image is transmitted only to the input of a neural network. An autoencoder-type convolutional neural network is capable of processing the at least one image so that defects occurring during laser additive manufacturing can be detected very effectively.

[0028] In other words, such a neural network is able to efficiently process the information contained in the first image, and therefore, it is meaningless to subsequently refer to the first image again to obtain the defect mask.

[0029] In some implementations, the only information used to generate the defect mask is the information provided as input to the first neural network. Therefore, in these implementations, it is impossible to consider supplementary information beyond the information provided as input to the first neural network to compute the defect mask. Thus, the processing only considers the information provided as input to the first neural network, specifically including the first image.

[0030] Here, the surface exposed to laser scanning is naturally the surface of a powder layer deposited by a doctor blade, and it is scanned by a laser beam so that the powder is melted in certain predefined areas so as to be incorporated into a part of the manufactured part.

[0031] The term "laser scanning" refers to the operation of using a laser beam to scan or travel across the entire surface or part of the surface, with the point of impact moving across the surface.

[0032] The efficiency of neural networks in detecting defects can be improved by implementing all or some of the following improvements:

[0033] In one implementation, the first image is an integral image of the upper surface. The term "integral image" as used herein refers to an image in which, for each pixel, the intensity represents the cumulative light energy received by that pixel during a capture period, particularly a capture period lasting greater than 0.1 seconds. The light intensity is thus integrated over time, and the resulting value is assigned to the pixel under consideration, thereby forming the integral image.

[0034] Different architectures can be envisioned for one or more neural networks involved in the processing steps.

[0035] In one implementation, in processing step B23, the defect mask is directly generated by a convolutional neural network.

[0036] The statement "directly generated by a neural network (convolution in this case)" implies that the defect mask is the output data of a convolutional neural network (or one of them). Therefore, in this case, it is not possible to obtain the defect mask by applying processing to the output data of that neural network that includes additional data or information (such as other images) besides those data or information provided to the neural network as input.

[0037] In this situation, it is also impossible to obtain a defect mask by applying the processing of the data that was once again provided to the neural network as input to the neural network to the output data of the neural network.

[0038] Nevertheless, the direct generation of output masks by neural networks does not preclude the use of one or more operations such as equalization, formatting, or thresholding performed by the neural networks to produce their output data, which includes defect masks.

[0039] In some implementations, the defect mask is directly generated by the first neural network. In other words, the first neural network is configured to directly generate the defect mask as output.

[0040] In some implementations, one or more additional images, in addition to the previously indicated first image, can be used to obtain a defect mask.

[0041] For example, in one implementation, during acquisition step B21, a second image of the upper surface exposed to laser radiation is also captured; the second image is an image of the surface captured in the infrared spectrum, wherein for each pixel, the intensity of the pixel represents the maximum light energy received by that pixel during the acquisition period; and during processing step B23, the processing performed takes the second image as input in addition to the first image.

[0042] In one implementation, during acquisition step B21, a third image of the upper surface exposed to the laser scan is also captured; the third image is an image of the surface captured in the visible spectrum; and during processing step B23, the processing performed takes the third image as input in addition to the first image.

[0043] In one or the other of the two aforementioned implementation modes, the first convolutional neural network may be configured to receive a first image and at least one first additional image from the second image and the third image as input.

[0044] In this case, preferably, the defect mask is directly generated by the first convolutional neural network.

[0045] However, more complex architectures can also be used.

[0046] Therefore, in the first variant of one of the two implementation modes described above, during processing step B23, a neural network group is used to perform the processing, which includes, in particular, an output neural network of the autoencoder type as the output and a first convolutional neural network as the input. The output neural network is configured to receive the output of the first convolutional neural network as the input and generate a defect mask based on the input (preferably directly).

[0047] Advantageously, in this first set of variants, the first convolutional neural network is sufficient to handle several (at least two) types of images provided as input.

[0048] Conversely, in a second variant of one of the two implementation modes described above, during processing step B23, a neural network is used to perform the processing, including:

[0049] - As input, the first convolutional neural network and at least a first additional neural network of an autoencoder type are configured to receive a first additional image from either the second image or the third image as input; and

[0050] - As output, the output neural network, particularly the autoencoder type output neural network, is configured to receive the outputs of the first neural network and the at least one first additional neural network as input, and to produce the defect mask as output.

[0051] In some variants of the second set of variants defined above, the neural network group further includes a second additional neural network of the autoencoder type, which is configured to receive a second additional image, different from the first additional image, from the second image and the third image as input.

[0052] In one implementation, at least one neural network in a group of self-encoding type neural networks, such as, in particular, the first neural network, includes connections linking a neural layer indexed y to a neural layer indexed ny, wherein the total depth of the at least one of the neural networks under consideration is equal to n.

[0053] In one implementation mode, during the manufacturing of parts by powder bed laser melting, a photographic or video camera is used to perform the capture step B21.

[0054] By extension, this disclosure also relates to a method for manufacturing parts by powder bed laser melting, wherein at least one part is manufactured by powder bed laser melting, and defects are detected during or after the manufacture of said at least one part by employing the method described above. The powder used in this method can be a powder of any composition, particularly a metal powder.

[0055] In a particular embodiment, the different steps of the method for detecting defects during laser additive manufacturing are determined by computer program instructions.

[0056] Therefore, this disclosure also relates to a computer program including instructions that, when executed by at least one processor, instruct the at least one processor to perform the steps of the aforementioned method. The program can use any programming language and can be in the form of source code, object code, or intermediate code between source code and object code, such as in a partially compiled form, or in any other desired form.

[0057] This disclosure also relates to a computer-readable non-volatile storage medium on which the aforementioned computer program is stored. The information medium can be any entity or device capable of storing a program. For example, the medium can include storage devices such as ROMs, such as CD-ROMs or microelectronic circuit ROMs, or magnetic recording devices such as floppy disks or hard disks. Alternatively, the information medium can be an integrated circuit with a program integrated into it, the circuit being adapted to perform the methods discussed or to be used in performing the methods discussed.

[0058] By extension, this disclosure also relates to a data processing apparatus, including at least one processor and a memory therein for recording instructions that, when executed by the at least one processor, direct the at least one processor to perform the steps of one of the methods described above.

[0059] Data processing equipment (or a portion thereof) can be integrated into a system that manufactures parts by laser melting of metal powder beds.

[0060] Therefore, this disclosure also includes a system for manufacturing parts by powder bed laser melting, comprising a machine for manufacturing parts by powder bed laser melting and a data processing device as described above. Attached Figure Description

[0061] [ Figure 1 ] Figure 1 (Presented) is a schematic representation of a conventional machine used to manufacture parts by laser melting of powder bed;

[0062] [ Figure 2 ] Figure 2 A system for manufacturing parts by powder bed laser melting according to the present disclosure is shown;

[0063] [ Figure 3 ] Figure 3 A first neural network used for detecting defects in a method for detecting defects according to this disclosure is shown;

[0064] [ Figure 4 ] Figure 4 A neural network group used for detecting defects in a method for detecting defects according to this disclosure is shown;

[0065] [ Figure 5 ] Figure 5 It is a representation of the integral image calculated during the implementation of the method according to this disclosure;

[0066] [ Figure 6 ] Figure 6 It is a representation of the largest image calculated during the implementation of the method according to this disclosure;

[0067] [ Figure 7 ] Figure 7 This is a schematic representation of a defect mask obtained by implementing the method according to this disclosure; and

[0068] [ Figure 8 ] Figure 8 This is a flowchart schematically illustrating the steps of a method for manufacturing parts by selective powder bed melting according to the present disclosure. Detailed Implementation

[0069] For example, now we will combine Figures 2 to 8 The present disclosure describes a method and system for manufacturing by selective powder bed melting.

[0070] These manufacturing methods can be used as follows: Figure 2 The system 1000 shown is implemented for manufacturing parts by laser, including a machine 100 for manufacturing parts by laser melting of powder bed and a data processing device 200.

[0071] Machine 100 is almost related to Figure 1 The manufacturing machine 100 described is the same; for this reason, the same or substantially the same components in the two machines are used in... Figure 1 and Figure 2 They have the same reference numerals. (And) Figure 1 Compared to machine 100, the special feature of machine 100 in system 1000 is that it also includes capture device 180.

[0072] The capture device 180 includes two video cameras capable of capturing images in the infrared and visible spectra, respectively. The images captured by these cameras are transmitted to the data processing device 200.

[0073] The data processing device 200 has a computer hardware architecture, such as... Figure 2 The diagram is schematically illustrated. Generally, any such data processing device can be used, comprising at least one memory capable of recording data and a program, which will be further described below, and one or more processors capable of executing the program. The data processing device may be located near machine 100, or conversely, away from machine 100, and may be accessible, for example, via a network such as the Internet.

[0074] In this embodiment, the data processing 200 specifically includes a processor 201, a read-only memory 202, a non-volatile flash memory 203, and a communication device 204 for communication with other components of the system 1000 (specifically including the control unit 160).

[0075] The non-volatile memory 203 of the data processing unit 200 constitutes a recording medium according to the present disclosure, which is readable by the processor 201 and on which a computer program according to the present disclosure is recorded, the computer program according to the present disclosure including steps for performing a method according to the present disclosure for manufacturing parts by powder bed laser melting (particularly including steps of its defect detection subroutine).

[0076] The program can take different forms. In the first implementation mode ( Figure 3 This constitutes the first neural network NN1 within the scope of this disclosure. In a second, more complex implementation mode ( Figure 4 This constitutes the convolutional neural network group NN2.

[0077] exist Figure 3 The diagram schematically illustrates a network NN1 constituting a first convolutional neural network according to the present disclosure.

[0078] In this embodiment, the neural network NN1 is used to detect three types of defects based on the input image.

[0079] The network NN1 is an autoencoder, which consists of an encoder E, a decoder D, and three normalization layers F1, F2, and F3.

[0080] In this embodiment, network NN1 is a U-net type autoencoder. A U-net type network is an autoencoder network that specifically includes connections C that directly connect blocks of encoder E to blocks of decoder D. Typically, a block at position i is connected to a block at position ni, where n is the total number of blocks.

[0081] The architecture of U-net type networks is specifically explained in the publication of Olaf Ronneberger, Philipp Fischer, and Thomas Brox: “U-net: Convolutional networks for biomedical image segmentation” (International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015).

[0082] The encoder E consists of several consecutive convolutional blocks E10, E20, E30, E40, and E50 (typically 3 to 5 convolutional blocks). Each of these convolutional blocks contains several consecutive layers of an artificial neural network. Within each block, the upstream layers (typically two or three) are convolutional layers, and the last downstream layer is an undersampling (or "pooling") layer that reduces the resolution of the processed data.

[0083] Here, the convolutional layer refers to the layer that performs the convolution operation, followed by the batch normalization process, and then the ReLU (Revised Linear Unit) correction step.

[0084] Similar to the encoder E, the decoder D also comprises several consecutive convolutional blocks D10, D20, D30, D40, and D50. Each of these convolutional blocks contains several consecutive artificial neural layers. In each of these blocks, the downstream layers (typically two or three) are convolutional layers, and the last downstream layer is an oversampling (or "upsampling") layer that increases the resolution of the processed data.

[0085] Naturally, a first neural network with an architecture different from that of network NN1 can be used, while remaining within the scope of this disclosure.

[0086] As input, encoder E receives input data D at each iteration of index i. IN i This data is processed by encoder E, which generates intermediate data D. INTERMi As output; this data D INTERM i The data is transmitted as input to decoder D. Decoder D processes the data and outputs data D. Out i Provided to each of the normalization layers F1, F2, and F3.

[0087] Each of these normalization layers affects the data D received from the decoder D. OUT i The Softmax or normalized exponential function is applied, producing the final images (M1i, M2i, and M3i, respectively) as output. 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 that a defect exists at the considered pixel (i.e., the location corresponding to that pixel on the surface of the considered layer) in relation to a defect associated with the defect being considered.

[0088] Therefore, in this embodiment, the defect mask Mi is directly generated by the first neural network NN1.

[0089] Optionally, a thresholding (also known as binarization) function can then 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 obtained mask. For each pixel, the thresholding function is to compare the value of the pixel with a predetermined threshold, and thus assign a binary value of 0 or 1 to the pixel based on whether the pixel has a value less than or greater than the threshold.

[0090] Input data D IN i It can be simply an integral image of a powder layer exposed to laser scanning. INT .

[0091] However, data D IN i It can also be a tensor obtained by cascading two or more images: especially an integral image I. INT And, where applicable, the largest image I MAX and / or visible image I VIS .

[0092] Consider using two or more images as input, which can improve the performance of neural network NN1, namely, improve the accuracy of the defect masks predicted by it. Advantageously, in this embodiment, the first neural network NN1 has been shown to be sufficient to handle different input images (integral image I). INT and the largest image I MAX and / or visible image I VIS ).

[0093] In a preferred embodiment, such as Figure 3 As can be seen, based on these input data, the first neural network NN1 directly generates the output mask Mi as the output.

[0094] exist Figure 4 The diagram schematically represents the neural network group NN2 (also simply referred to as "NN2") constituting the second embodiment.

[0095] NN2 is a neural network group, which, according to an embodiment, includes one to three convolutional neural networks of the autoencoder type as input: AE INT AE MAX AE VIS And the output is also an autoencoder-type neural network AE. OUT Network AI INT This constitutes the first neural network within the scope of this disclosure; optional network AE MAX AE VIS First and second additional networks are formed. An output autoencoder AE is included downstream of one or more input autoencoders. OUT This architecture allows for improved prediction accuracy of neural networks.

[0096] When the NN2 network can process not only one image, but two or more images as input, its performance is improved compared to the architecture of the NN2 network that can only consider a single image as input.

[0097] To accommodate two or more images as input, in one embodiment, the neural network NN2 includes only the first neural network as input, preferably the neural network AE. INT (and therefore excludes network AE) MAX and AE VIS Output neural network AE OUT The first neural network (e.g., AE) receives the signal. INT The output of ) is used as input, and the defect mask Mi is directly generated as output.

[0098] In other embodiments, such as Figure 4 As can be seen, the neural network NN2 includes at least two neural networks as inputs (in the case shown, it includes three networks AE).INT AE MAX AE VIS ) and the neural network AE as output OUT The input to the neural network includes different images received as input.

[0099] The outputs of different input neural networks are transmitted to the output neural network AE. OUT This is used to generate the output mask Mi.

[0100] Preferably, the defect mask Mi is generated by the output neural network AE. OUT It is generated directly.

[0101] in this case( Figure 5 In each iteration (at index i), NN2 receives three images: the integral image I. INTi Maximum Image I MAXi and images captured in the visible spectrum I VISi .

[0102] Neural Network AE INT AE MAX AE VIS and AE OUT Each of them has the same Figure 3 The neural network NN1 represented in the diagram has the same or roughly the same architecture. However, in network AE... INT AE MAX AE VIS In the process, the existence of one (or more) normalization layers in the downstream network is optional.

[0103] In order to merge the three input networks AE INT AE MAX and AE VIS The generated data, NN2 group, consists of neural network layers G, used to cascade images I generated by decoders from these different networks as outputs. INT_OUTi I MAX_OUTi I VIS_OUTi .

[0104] Finally, in the neural network group NN2, the output neural network AE is... OUT It also includes a normalization layer F as output. This normalization layer can be compared with... Figure 3 The normalization layer F is arranged in the same way, for example, several Softmax layers are arranged in parallel to produce different images as output, each image representing the probability of a defect at each point (or at each pixel); each image represents the defect probability of one of the defects being searched.

[0105] Therefore, in this embodiment, these different images constitute the output mask Mi, which is generated by the output neural network AE. OUT It is generated directly.

[0106] Parts manufacturing

[0107] The manufacturing of a part is carried out iteratively by performing the process at each layer of the part. Figure 8 It is accomplished through the different steps indicated in the document.

[0108] During the iteration at index i, the manufacturing of the n°i-th layer of the part is accomplished by performing the following operations:

[0109] In the first step A, a new powder layer is deposited using a doctor blade 170. For this purpose, the powder storage plate is raised, while the fabrication plate is lowered from a corresponding height.

[0110] Next (step B1), a laser scan is performed to scan all points of the layer under consideration, which must be part of the part 120 to be manufactured. The laser scan produces localized melting of the powder, which welds the powder particles to the lower part of the already formed part in the area exposed to the laser scan.

[0111] In parallel, the data processing unit executes a program to perform method B2 for detecting defects.

[0112] Method B2 for detecting defects that occur during manufacturing operations in system 1000 includes the following steps:

[0113] (B21) Different image sequences or series are captured in parallel on the surface scanned by the laser beam. More precisely, for the duration of laser scanning B1, each camera of the capturing device 180 captures an image sequence. In the described implementation mode, a first camera operating in the visible spectrum captures an image sequence in the visible spectrum, and a second camera operating in the infrared spectrum captures an image sequence in the infrared spectrum. Each image represents the entire surface scanned by the laser beam 142 (alternatively, capture may be limited to the area around the point of impact 144 of the laser beam).

[0114] The infrared image is captured from an optical tomographic sensor. In this case, it is an infrared S-CMOS sensor with a resolution of 2000 pixels × 2000 pixels. The image captured by the sensor in near-infrared light represents the temperature field at the surface of the layer.

[0115] The captured images are captured sequentially, completed throughout the entire laser scanning process of the layer.

[0116] B22) The data processing unit 200 applies preliminary processing to the captured image sequence: Based on the captured images, the data processing unit 200 calculates three images:

[0117] First Image I INT_ i is an integrated image of the upper surface, integrated during the integral image capture period during the fabrication of layer n°i. In this embodiment, the capture period of the integral image is equal to the duration of the laser scan of layer n°i. Alternatively, it can be a set duration that includes the moment the laser beam strikes the point under consideration.

[0118] Second image I MAX_ i represents the image with 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 implementation, the capture period of the maximum image is equal to the duration of the laser scan. Alternatively, it can be a set duration that includes the moment the laser beam strikes the point under consideration.

[0119] ·Third Image I VIS_ i is an image captured in the visible spectrum of the upper surface of the layer that has undergone laser beam scanning during the fabrication of layer n°i. (This image may optionally be an image I captured in the visible spectrum during the duration of the laser scan.) VIS_ The integral image I is calculated by integrating (or averaging) i. VIS_INT_ i).

[0120] The process captures three images within the data processing unit 200. INT i、I MAX_ i and i VIS_ i.

[0121] Image I INT i and i MAX_ i respectively as Figure 5 and Figure 6 As shown in the diagram.

[0122] Therefore, in the first image I, which is the integral image INT_ i (for example) Figure 5 In the diagram shown, the value of each pixel is given by the following formula:

[0123]

[0124] Where 'e' is the infrared emission intensity value measured by a pixel. This image can be interpreted as a photograph, with an exposure time equal to the laser scanning time of the layer.

[0125] In the second image, which is the largest image, I MAX_ i (for example) Figure 6In the diagram shown, the value of each pixel is given by the following formula:

[0126]

[0127] Where t layer It is the laser scanning cycle of this layer.

[0128] This maximum image can be interpreted as the envelope of infrared emission intensity during laser scanning of the layer. It can be noted that the overlapping line L between adjacent parallel bands exposed to the laser scan (or “laser band”) appears in the integral image I. INT_ i( Figure 5 The lines shown in white in the largest image I) MAX_ Not very visible in i ( Figure 6 Due to this property, using the maximum image is particularly effective in limiting the number of false positives. Specifically, if only integral images are used, neural networks tend to overemphasize these overlapping lines (BLs). Therefore, the information contributed by the maximum image appears to help the neural network avoid interpreting overlapping lines between adjacent laser bands as defect regions, and thus allows for improved quality of the defect mask generated by the neural network group NN2.

[0129] B23) Uses a neural network group NN2 to process three images I INT_ i、I MAX_ i and i VIS_ i.

[0130] These images are provided as input to NN2: in the algorithm's iteration n°i, the network AE INT AE MAX AE VIS Receive integral images I respectively INT_ i, Maximum Image I MAX_ i and visible image I VIS_ i is used as input. Based on these, the network AE... INT AE MAX AE VIS Generate output image I respectively INT_OUT i I MAX_OUT i and I VIS_OUT i .

[0131] These output images are then concatenated by a cascaded layer G, forming a 3rd-order tensor. This tensor is then fed as input to the output neural network AE. OUT .

[0132] As output, the output neural network AE OUTA defect mask Mi is generated. In this embodiment, the neural network group NN2 is configured to generate a defect mask as output, which is an image Mi. This image Mi is compared with image I. INT i、I MAX i and i VIS i has the same dimension and indicates the location of the defect in the image.

[0133] Alternatively, the data processing performed in step B23 can be accomplished using a neural network NN1 (configured in a way that suits the number of images being considered as input) instead of neural network NN2.

[0134] Figure 7 This represents a binary mask obtained by applying a threshold function to a mask Mi. Therefore, each pixel has only a value of 0 or 1. Thus, pixels located at defect locations are represented in white, while pixels in areas without defects are represented in black. Note that... Figure 7 The obvious defects also appeared in Figure 5 The right side of the integral image in the image.

[0135] Using three images together INT i、I MAX_ i and i VIS_ i is the combination that allows the use of a neural network group NN2 to obtain optimal performance. However, according to the first group of neurons of this disclosure (configured and therefore trained), such as network NN1, it is possible to rely solely on the image captured in infrared, particularly on the integral image I. INT i is used to generate a defect mask.

[0136] (B24) Based on the defect mask Mi, determine the actions for the remainder of manufacturing. Actions can also be determined based on the defect mask Mi of the ongoing step and one or more defect masks obtained in previous steps. Depending on one or more defects, it can be decided to continue manufacturing without any modifications (and thus begin manufacturing the next layer n°i+1 of the part); to continue manufacturing by modifying one or more operating parameters of the manufacturing machine 100; or to stop manufacturing the part. In the second case, the modified parameters could be the displacement velocity of the laser beam impact point, the power of the laser beam, the thickness of the deposited powder layer, etc.

[0137] Based on the decisions made, commands are determined and transmitted to control unit 160 for use in the rest of the manufacturing process.

[0138] (Multiple) neural network drivers

[0139] The training of the neural network(s) is performed using a database of training data in a manner known per se. This database includes input data and output data; these output data are the output masks (“truth values”) that one would expect the neural network or, where applicable, a group of neural networks to produce, when the input data is provided as input to the neural network.

[0140] Depending on the number of images that the neural network or neural network group can receive as input, the input data can be images, pairs of images, or n-tuples of images. For example, for training NN2, the input data is the image triplet I computed by performing step B22 of the method. INT i、I MAX_ i and i VIS_ i.

[0141] The following explanation is given in the case that the neural network or neural network assembly is configured to receive triples of images as input; however, it should be understood that this disclosure is applicable regardless of the number of images received by the neural network or neural network assembly as input.

[0142] The desired output mask (“truth”) can be a mask illustrating the different types of defects that the neural network (NN) must identify, such as hotspots (points where the temperature reaches excessively high values), lack of melt, and contamination of the melt bath. These defect masks can be prepared by experts involved in the manufacturing of the part and represented by these experts based on the input image triples I. INT i I MAXi and I VIS i The identified manufacturing defects consist of images (binary images, if applicable).

[0143] In one embodiment, a neural network (or a group of neural networks) is driven such that each image of the defect mask is used to evaluate the presence of a specific defect.

[0144] In this scenario, the database is trained and then a set of input image triples (I0.05) is included for each defect. INT i、I MAX_ i and i VIS_ i) Each triple has an associated output image (optionally binary). For the corresponding input image triple, the output image represents the best possible estimate of the probability that a considered defect exists for each pixel of the image (and therefore for each corresponding point of the layer being laser-scanned).

[0145] In this case, after training, the image obtained as the output of the neural network (as the output of the normalization layer) is a probabilistic image, which represents the probability of the presence of the considered defect for each pixel.

[0146] In another embodiment, for at least one image of the defect mask, the neural network is trained to simultaneously evaluate the presence of several defects. The neural network is thus trained to predict the presence or absence of several defect types, such as, for example, incomplete melting, hot spots, or part contamination due to impurities in the powder layer, via the resulting output image.

[0147] In this case, for images that consider defect masks, the training database consists of a set of input image triples (I... INT i、I MAX_ i and i VIS_ i), and for each triple, including the associated output image. This output image is an image in which each pixel is assigned a specific value (or color) associated with a certain type of defect identified at the pixel's location. For example, the output image could consist of pixels with one of the following values: 0 (no defect), 1 (insufficient melt), 2 (hot spot), or 3 (part contamination).

[0148] To enrich the training database, additional images can be added using so-called "enhancement" methods. These images can be generated by applying rotation, image shifting, left / right flipping, top / bottom flipping, etc., to images that originally exist in the sample database.

[0149] The cost function used can be, in particular, a binary cross-entropy function.

[0150] Training can be performed end-to-end (end-to-end learning) in all proposed architectures: whether using a simple architecture that includes only the first neural network NN1, or a complex architecture that uses a group of neural networks NN2.

[0151] Furthermore, the dimensions of the image can be optimized. The image fed as input to the neural network can be computed through undersampling. For example, an image generated by an optical tomography sensor may have a high resolution (2000×2000 pixels or higher), but for the purpose of implementing this method, only a lower resolution image, such as 1000×1000 pixels, is used.

[0152] When the image fed as input to the neural network(s) has high resolution and undersampling is avoided, another feasible technique involves dividing the input image (initially captured at "high resolution") into lower-resolution images, each with a resolution compatible with the processing power of the neural network(s) used. The different sub-images are then processed by the neural network(s) to obtain corresponding individual output masks. These output masks, having the same dimensions or resolution as the sub-images, are then recombined, allowing the acquisition of a high-resolution output mask with the same resolution as the initially captured image. For example, an input image with a resolution of 2000×2000 pixels can be divided into 16 sub-images with dimensions of 500×500 pixels. To obtain an output mask for the initial image with a resolution of 2000×2000 pixels, the output masks generated based on the different sub-images are assembled.

[0153] During network training, a search for the optimal architecture is provided to automatically determine different architectural parameters of the neural network, particularly the following parameters: the dimension of the input data of the neural network (i.e., the dimension of the input image I). INT I MAX and I VIS Size); Each network AE INT AE MAX AE VIS and AE OUT The depth (number of convolutional blocks); the dimension of intermediate data or "feature maps" 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 in particular using grid analysis (grid search).

[0154] Although the invention has been described with reference to specific exemplary embodiments, it will be apparent that various modifications and alterations can be made to these examples without departing from the general scope of the invention as defined by the claims. Furthermore, the various features of the different embodiments described can be combined in additional embodiments. For example, the method can be modified by only viewing image I. INT i or only image I INT i and i MAX i instead of three images I INT i、I MAX i and i VIS i is provided as input to the neural network group for implementation. Therefore, the specification and figures should be considered illustrative, not restrictive.

Claims

1. A method (B2) for detecting defects during laser additive manufacturing, comprising the following steps: B21) captures the first image (I INT_ i), the first image is an image captured in the infrared spectrum of the upper surface of a powder layer exposed to laser scanning; B23) The first image is processed using a first convolutional neural network (NN1, NN2) of the autoencoder type to generate a defect mask (Mi) indicating the location of a defect on the upper surface of the powder layer; During the capture step B21, a second image of the upper surface exposed to laser radiation is also captured. The second image is an image of the surface captured in infrared spectrum (I MAX_i (), where for each pixel, the intensity of the pixel represents the maximum light energy received by the pixel during the capture period; as well as During processing step B23, in addition to the first image (I) INTi In addition to the second image (I), the processing also considers the second image (I) MAX_i ) as input, The first convolutional neural network (AE) INT ) is configured to receive a first image and a second image (I MAX ) as input, During processing step (B23), processing is performed using a neural network group (NN2), which includes the first convolutional neural network (AE) as input. INT ) and the output neural network (AE) as the output OUT The output neural network (AE) OUT It is configured to receive the output of a first convolutional neural network as input and generate a defect mask (Mi) based on that input.

2. The method for detecting defects during laser additive manufacturing according to claim 1, wherein the first image is an integral image of the upper surface (I0). INT (i), where for each pixel, the intensity of the pixel represents the cumulative light energy received by the pixel during the capture period.

3. The method for detecting defects during laser additive manufacturing according to claim 1, wherein during the capture step (B21), a third image of the upper surface exposed to the laser scan is also captured; The third image is an image of the surface captured in the visible spectrum (I VIS_ i); and During processing step (B23), except for the first image (I) INT In addition to i), the processing performed also considers the third image (I). VIS_ i) as input.

4. The method for detecting defects during laser additive manufacturing according to claim 3, wherein the first convolutional neural network (AE) INT ) is configured to receive a first image and a second image (I MAX ) and the third image (I VIS () as input.

5. The method for detecting defects during laser additive manufacturing according to claim 1, wherein the defect mask (Mi) is directly generated by a first convolutional neural network (NN1).

6. The method for detecting defects during laser additive manufacturing according to claim 3, wherein during processing step (B23), processing is performed using a neural network group (NN2), the neural network group (NN2) comprising: - The first convolutional neural network as input and at least one first additional neural network of autoencoder type (AE) MAX AE VIS ), the first convolutional neural network and at least one first additional neural network of the autoencoder type (AE) MAX AE VIS The image is configured to receive a first additional image as input, wherein the first additional image is either the second image or the third image; and - Output neural network (AE) as the output of an autoencoder type OUT It is configured to receive the outputs of a first convolutional neural network and the at least one first additional neural network as inputs, and to produce the defect mask (Mi) as output.

7. The method for detecting defects during laser additive manufacturing according to claim 6, wherein the neural network group (NN2) further comprises a second additional neural network (AE) of the autoencoder type. MAX AE VIS ), the second additional neural network (AE) MAX AE VIS The image is configured to receive a second additional image, which is different from the first additional image, from the second image and the third image.

8. The method for detecting defects during laser additive manufacturing according to claim 1, wherein the first convolutional neural network (NN1) includes connections linking a neural layer indexed y to a neural layer indexed ny, and the total depth of at least one of the neural networks under consideration is equal to n.

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

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

11. The method for detecting defects during laser additive manufacturing according to claim 1, wherein during the application of processing to the first image and generation of a defect mask (Mi) using the first convolutional neural network, the first image is transmitted only to the input of the first convolutional neural network (NN1, NN2).

12. A method (B2) for detecting defects during laser additive manufacturing, comprising the following steps: B21) captures the first image (I INT_ i), the first image is an image captured in the infrared spectrum of the upper surface of a powder layer exposed to laser scanning; B23) The first image is processed using a first convolutional neural network (NN1, NN2) of the autoencoder type to generate a defect mask (Mi) indicating the location of a defect on the upper surface of the powder layer; During the capture step B21, a second image of the upper surface exposed to laser radiation is also captured. The second image is an image of the surface captured in infrared spectrum (I MAX_i (), where for each pixel, the intensity of the pixel represents the maximum light energy received by the pixel during the capture period; as well as During processing step B23, in addition to the first image (I) INTi In addition to the second image (I), the processing also considers the second image (I) MAX_i ) as input, During the capture step (B21), a third image of the upper surface exposed to the laser scan is also captured; The third image is an image of the surface captured in the visible spectrum (I VIS_ i); and During processing step (B23), except for the first image (I) INT In addition to i), the processing performed also considers the third image (I). VIS_ i) As input, During processing step (B23), processing is performed using a neural network group (NN2), which includes: - The first convolutional neural network as input and at least one first additional neural network of autoencoder type (AE) MAX AE VIS ), the first convolutional neural network and at least one first additional neural network of the autoencoder type (AE) MAX AE VIS The image is configured to receive a first additional image as input, wherein the first additional image is either the second image or the third image; and - Output neural network (AE) as the output of an autoencoder type OUT It is configured to receive the outputs of a first convolutional neural network and the at least one first additional neural network as inputs, and to produce the defect mask (Mi) as output.

13. The method for detecting defects during laser additive manufacturing according to claim 12, wherein the first image is an integral image (I0) of the upper surface. INT (i), where for each pixel, the intensity of the pixel represents the cumulative light energy received by the pixel during the capture period.

14. The method for detecting defects during laser additive manufacturing according to claim 12, wherein the first convolutional neural network (AE) INT ) is configured to receive a first image and a second image (I MAX () as input.

15. The method for detecting defects during laser additive manufacturing according to claim 14, wherein the defect mask (Mi) is directly generated by a first convolutional neural network (NN1).

16. The method for detecting defects during laser additive manufacturing according to claim 14, wherein... During processing step (B23), processing is performed using a neural network group (NN2), which includes the first convolutional neural network (AE) as input. INT ) and the output neural network (AE) as the output OUT The output neural network (AE) OUT It is configured to receive the output of a first convolutional neural network as input and generate a defect mask (Mi) based on that input.

17. The method for detecting defects during laser additive manufacturing according to claim 12, wherein the neural network group (NN2) further comprises a second additional neural network (AE) of the autoencoder type. MAX AE VIS ), the second additional neural network (AE) MAX AE VIS The image is configured to receive a second additional image, which is different from the first additional image, from the second image and the third image.

18. The method for detecting defects during laser additive manufacturing according to claim 12, wherein the first convolutional neural network (NN1) includes connections linking a neural layer indexed y to a neural layer indexed ny, and the total depth of at least one of the neural networks under consideration is equal to n.

19. The method for detecting defects during laser additive manufacturing according to claim 12, wherein a capture step (B21) is performed using a photographic camera or a video camera during the manufacturing of a part by laser melting of a powder bed.

20. The method for detecting defects during laser additive manufacturing according to claim 12, wherein in the processing step (B23), the defect mask (Mi) is directly generated by a convolutional neural network.

21. The method for detecting defects during laser additive manufacturing according to claim 12, wherein during the application of processing to the first image and generation of a defect mask (Mi) using the first convolutional neural network, the first image is transmitted only to the input of the first convolutional neural network (NN1, NN2).

22. A method for manufacturing a part by powder bed laser melting, wherein at least one part is manufactured by powder bed laser melting, and defects are detected during or after the manufacture of said at least one part by employing the method of any one of claims 1 to 21.

23. A data processing apparatus (200) comprising at least one processor and a memory therein recording instructions that, when executed by the at least one processor, direct the at least one processor to perform the steps of the method as claimed in any one of claims 1 to 21.

24. A system (1000) for manufacturing parts by powder bed laser melting, comprising a machine (100) for manufacturing parts by powder bed laser melting and a data processing device (200) as claimed in claim 23.

25. A computer program product comprising instructions that, when executed by at least one processor, cause the at least one processor to perform the steps of the method as claimed in any one of claims 1 to 21.

26. A computer-readable non-volatile storage medium having thereon stored a computer program product including instructions as claimed in claim 25.

Images