Part inspection method

Abstract

A method for inspecting a part (10) with reference to a scale comprising several material integrity grades and at least one reference image (24) associated with each grade, the method comprising: obtaining an image (20) of the part (10) in at least one imaging direction (12); obtaining a mask (22) of the part (10); applying the mask (22) to at least one reference image (24) in at least one position and at least one orientation to obtain at least one comparative image (30) associated with the grade of the corresponding reference image (24); for each comparative image (30), calculating (32) a metric representative of the difference in material integrity between the comparative image (30) and the image (20) of the part (10); based on the calculated metric, assigning (36) to the part (10) the material integrity grade for which the difference in integrity of Material is minimized. Figure for the abbreviation: Fig. 1

Description

Title of the invention: Method for inspecting a part technical field

[0001] This presentation relates to the field of industrial quality, and more specifically to a method for inspecting a part, particularly with regard to the integrity of the part's material, sometimes known as "material health". This method can be applied to parts in all sectors of activity, including aerospace. Previous technique

[0002] Evaluating the quality of a part requires an objective reference. To this end, standards, such as ASTM E2660 (“E2660-17: Standard Digital Reference Images for Investment Steel Castings for Aerospace Applications”) in the field of aerospace casting, have been established to provide a scale comprising several quality levels, typically several material integrity grades, and at least one reference image associated with each grade. Thus, by visually comparing an image of the part to be evaluated with these reference images, an operator can, by following the requirements of the appropriate standard, estimate the quality of the part.

[0003] In this document, the quality of the material is assessed based on the presence of casting defects and classified into 8 grades. "Grade 1" corresponds to a sound material, without material defects, and "Grade 8" corresponds to a material with very degraded quality.

[0004] However, the visual estimation of the part's quality by an operator relies solely on an empirical, and therefore non-deterministic, approach. This approach is thus subjective, which can lead to inaccurate estimates, and relatively slow, which limits its applicability to larger-scale applications.

[0005] There is therefore a need for a new type of part inspection process. Description of the invention

[0006] To this end, the present description relates to a method for checking a part with reference to a scale comprising several material integrity grades and at least one reference image associated with each grade, the method comprising: - obtaining an image of the part according to at least one imaging direction; - obtaining a mask of the part; - the application of the mask to at least one reference image according to at least a position and at least one orientation in order to obtain at least one image comparative associated with the grade of the corresponding reference image; - for each comparative image, the calculation of a metric representing the difference in material integrity between the comparative image and the image of the part; - based on the calculated metric, the assignment to the part of the grade for which the difference in material integrity is minimized.

[0007] As previously stated, the scale can define several grades of material integrity. Each grade is associated with at least one reference image representative of the material integrity for that grade. In this statement, and unless otherwise indicated, "a" or "a" reference image means "at least one" or "each" reference image. More generally, "an" or "an" element means "at least one" or "each" element. Conversely, the generic use of the plural may include the singular.

[0008] The reference images can be images of the same type as the image of the part obtained. Conversely, to allow for a suitable comparison, the imaging technique for the part is chosen to provide a result of the same nature as the reference images.

[0009] The imaging direction, where applicable, refers to the direction in which the image of the part is taken. For example, for a two-dimensional image, the imaging direction is a direction orthogonal to the image plane. Obtaining an image may include acquiring an image from the part or retrieving an image of the part previously acquired and recorded, typically on a recording medium.

[0010] Similarly, obtaining a mask of the part may include acquiring the mask from the part or retrieving a previously acquired and recorded mask, typically on a recording medium.

[0011] A part mask is a filter configured to, given a source image showing the part, allow a portion of the source image with the same shape as the part to pass through and mask the rest of the source image, or vice versa. Thus, the part mask can reproduce the contours of the part.

[0012] The part mask can be a mask of the previously obtained part image. More generally, the part mask can show the part in the same imaging direction as the part image.

[0013] In general, there is no guarantee that the image of the part will have dimensions and shapes comparable to those of the reference images. In order to compare the image of the part to the reference images, the part mask is applied to the reference images, like a stencil, to extract reference images from the part. These are called comparative images, which have a shape and dimensions comparable to those of the image of the part (or even exactly those of the part). Thus, the process can be applied to parts of any shape and size.

[0014] Each comparative image remains associated with the grade of the reference image (for example, a reference image from the ASTM grading scale) from which it is derived. In other words, the comparative image inherits the grade of the reference image from which this comparative image was obtained.

[0015] A metric is then calculated for each comparative image to quantify the difference in material integrity between the comparative image and the part image. The metric can be based on any image characteristic related to material integrity, for example, defect size, spacing between defects, defect density or homogeneity, etc. Further examples will be given later.

[0016] Image operations, such as calculating a metric, can be performed by associating one or more values ​​with each elementary cell of the image, for example, with each pixel. The values ​​can be representative of a color, a gray level, etc.

[0017] The metric considered is not absolute but reflects the difference between the image of the part and a comparative image in terms of material integrity. The metric therefore aims to quantify how similar the image of the part and the comparative image are. The metric generally returns a value. The metric may take into account several criteria, which may, where appropriate, be weighted relative to one another.

[0018] Based on the metric calculated for all comparative images, the grade of the part can be determined as the grade of the comparative image for which the difference in material integrity is minimized. Indeed, the metric makes it possible to determine, among the comparative images, the one whose image of the part is most similar in terms of material integrity. This comparative image is therefore representative of the material integrity of the part, which is why the part is assigned the grade associated with this comparative image.

[0019] The proposed method thus makes it possible to estimate the grade of a part, and therefore its material integrity, in a deterministic, reliable, and objective manner, by calculating a metric with reference to comparative images similar to the image of the part. Furthermore, such a method can be easily automated, resulting in significant time savings.

[0020] The steps of the control process can be implemented by computer.

[0021] In certain embodiments, the image of the part is obtained by a non-destructive imaging method. Thus, any production part can be inspected without impacting its properties, and not just a few samples which will be inevitable. tablement eliminated subsequently.

[0022] For example, the image of the part may be a radiograph, for example a two-dimensional radiograph. The radiograph may be produced using X-rays or any other radiation suitable for the structure and materials of the part.

[0023] In some embodiments, the inspection process includes obtaining an image of the part from several imaging directions, and the grade is assigned to the part based on the grades assigned for each imaging direction. Indeed, an inspected part is not necessarily isotropic in terms of material integrity. Taking into account several imaging directions minimizes the influence of selecting a single imaging direction, and thus provides a more accurate estimate of the part's material integrity. The inspection process is therefore more robust.

[0024] Optionally, the grade assigned to the part is an average of the grades assigned for each of the imaging directions or the grade for which the difference in material integrity is minimized across the different imaging directions. Depending on what is sought and the properties to be controlled, one may indeed be interested in the average grade of a part, or in the grade that is most probable regardless of the imaging direction considered.

[0025] In certain embodiments, the mask is applied to at least one reference image in a plurality of positions and / or orientations, thereby obtaining a plurality of comparative images for said at least one reference image. In other words, several comparative images can be obtained from the same reference image. The reference images may not be homogeneous, and / or the mask may be relatively small compared to a reference image. By applying the mask in a plurality of positions and / or orientations, the influence of edge effects in the mask's positioning on the reference image is minimized. This allows for better consideration of the information contained in the reference image. The control method is therefore more robust, as the metric is calculated on a larger and more representative set of comparative images.

[0026] In certain embodiments, the mask is applied using a random sampling method such as a Monte Carlo method. Random sampling methods, and more specifically the Monte Carlo method, are statistical methods, known in themselves, that allow for the random selection of several values ​​according to a given probability distribution, for example, a uniform probability distribution. In this case, the random sampling method can allow for the random selection of positions and orientations of the mask relative to the reference image under consideration. Thanks to these arrangements, a plurality of comparative images can be obtained without bias, which further contributes to the robustness of the control process.

[0027] In some embodiments, the metric compares at least one of the Shannon entropy and the size of a defective area between the comparison image and the image of the part. Shannon entropy, known as such in information theory, quantifies the amount of information contained in an input signal such as an image. The underlying idea is that two images are similar in terms of material integrity if they contain similar amounts of information in terms of the quantity of defects and perturbations in their structure.

[0028] Furthermore, the size of a defective area can refer to a real or effective dimension of the defect, for example an apparent side or an apparent radius of the defect, that is to say the square root of the area of ​​the defect or the radius of the circle with the same area as the defect.

[0029] Shannon entropy and the size of a defective area are relatively simple quantities to calculate, thus enabling efficient and robust automation of the inspection process. Furthermore, when used in combination, Shannon entropy and the size of a defective area provide complementary information for estimating the difference between the image of the part and the comparison image. The proposed metric is therefore an effective measure of the difference between two images.

[0030] In some embodiments, the metric is defined by M(Xrx, Xref) = x (H(Xre,)-H(XRX))lH(xRX) + fA x (A„f-ARX) / ARX\ , where fj A are weights associated respectively with the Shannon entropy and the size of the defective area, Ht) measures the Shannon entropy, ^ref represents the comparative image, XRX represents the image of the part, Aref represents the size of a defective area on the comparative image and represents the size of a defective area on the image of the part. The weights can be taken as equal or adjusted according to the importance given to each criterion based on the properties that one seeks to quantify.

[0031] In some embodiments, the size of a defective area is measured by the number of pixels having a value exceeding a threshold representative of sound material. Typically, for an area measurement, the number of contiguous pixels (arc-connected) having a value exceeding said threshold is counted in two dimensions.

[0032] Thus, the size of the defective area is easy to estimate automatically. The representative threshold can be a value representing a certain color or a certain shade of gray, or more generally any quantity encoded by the image, as explained previously. In order to have a value suitable for the representative images provided with the scale, the representative threshold can be determined from the reference image representing the healthiest grade, for example, as being equal to the average value of this reference image plus (or subtracted from) one standard deviation, or even of three standard deviations.

[0033] In certain embodiments, the inspection method includes adjusting at least one of the resolution, color distribution, and material thickness represented in the image of the part relative to at least one reference image. The inspection method thus transforms the image of the part to place it under conditions similar to those under which the reference images were obtained, so that the comparison (via metric calculation) is as reliable as possible.

[0034] In some embodiments, the adjustment involves assigning a value determined from a value of sound material on at least one reference image to a material pixel in the part image. By associating a value determined from a value of sound material on a reference image, for example, on the reference image corresponding to the soundest grade, each material pixel value in the part image is virtually transformed into a thickness and material identical to that of the reference. This procedure directly converts the value ranges (for example, gray levels) of the part image into those of the reference image and also geometrically transforms the part into a representation of the parts that gave rise to the reference images. This makes the comparison between the part image and the comparative images even more reliable.

[0035] In some embodiments, obtaining the image of the part includes obtaining a three-dimensional image of the part comprising a plurality of voxels, assigning a first value to the healthy material voxels of the plurality of voxels and a second value to the defect voxels of the plurality of voxels, and two-dimensionally projecting the three-dimensional image along at least one imaging direction, wherein a value of a pixel of the resulting image is calculated as a function of the values ​​of the voxels that are projected onto that pixel.

[0036] A voxel is to three dimensions what a pixel is to two dimensions, namely an elementary cell of an image. The projection can be performed along the imaging direction. The resulting image, or image resulting from the projection (i.e., the image of the part), is obtained by taking into account the voxels that overlap along the imaging direction to form a single pixel of the part image. The two-dimensional projection results in a two-dimensional image.

[0037] A voxel is determined to represent sound material or a defect according to its value and the imaging technique used. For example, those skilled in the art know that, for a given imaging technique, a void is represented by a high or low value. A threshold can be used to distinguish between sound material voxels and defect voxels; this threshold may have all or some of the characteristics mentioned above.

[0038] Thanks to these characteristics, it is possible to obtain an image, typically two-dimensional (2D), from a three-dimensional (3D) image of a part for which a 2D imaging method is not applicable or desired. In particular, this method makes it possible to obtain 2D images of digitally modeled 3D parts.

[0039] In some embodiments, the first value is a non-zero constant, for example 1, and / or the second value is zero.

[0040] In some embodiments, the inspection method includes identifying defect voxels by segmenting the three-dimensional image. Image segmentation, for example semantic segmentation, is a well-known technique that uses, for example, deep learning models. The identification of defect voxels can thus be automated.

[0041] In other embodiments, particularly when the part is a digital part in which the defects have been digitally placed, the defect voxels can be known by construction.

[0042] In certain embodiments, the value of the aforementioned pixel is equal to the sum of the voxel values ​​projected onto that pixel. This sum is an approximation of Beer-Lambert's law, which describes the absorption of an imaging ray passing through the part, for small thicknesses. The fact that the pixel value is equal to the sum of the voxel values ​​projected onto it therefore provides both a simple and physically relevant method for calculating the image of the part.

[0043] In some embodiments, the three-dimensional image of the part is a three-dimensional tomography. Tomography, for example X-ray tomography, is a technique known in itself for obtaining three-dimensional images.

[0044] In some embodiments, the inspection process includes passing or rejecting the part based on its assigned grade. Typically, if the grade exceeds a certain threshold, the part can be scrapped, and kept otherwise. In this way, production quality can be effectively maintained.

[0045] In a particular embodiment, the various stages of the control process are determined by computer program instructions. In other words, the control process can be implemented by computer.

[0046] Consequently, the present exposition also relates to a program set comprising instructions for the execution of the steps of the control process described above when said program set is executed by at least one computer or microprocessor.

[0047] This program set can use any programming language, and be in the form of source code, object code, or code intermediate between source code and object code, such as in a partially compiled form, or in any what other form would be desirable?

[0048] The present exposition also relates to an information carrier readable by a computer or by a microprocessor, and comprising instructions for a set of programs as mentioned above.

[0049] The information medium can be any entity or device capable of storing the entire program. For example, the medium can include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a floppy disk or a hard disk.

[0050] On the other hand, the information medium can be a transmissible medium such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means. The program described herein can, in particular, be downloaded from an Internet-type network.

[0051] The present disclosure also relates to a method for imaging a part, comprising obtaining a three-dimensional image of the part comprising a plurality of voxels, assigning a first value to the voxels of sound material in the plurality of voxels and a second value to the voxels of defects in the plurality of voxels, and projecting the three-dimensional image two-dimensionally in at least one imaging direction, wherein a pixel value of the resulting image is calculated based on the voxel values ​​projected onto that pixel. The imaging method may have all or some of the features mentioned above in relation to the inspection method. Brief description of the drawings

[0052] Other features and advantages of the subject matter of this presentation will become apparent from the following description of embodiments, given by way of non-limiting examples, with reference to the attached figures.

[0053] Fig. 1 is a general diagram of a control method according to one embodiment.

[0054] Figure [Fig. 2] represents a step in projecting a three-dimensional image into a two-dimensional image. Detailed description

[0055] A method for inspecting a part according to an embodiment is described with reference to Figures 1 and 2. The inspection method aims to quantify the material integrity of the part with reference to a scale comprising several material integrity grades. The material integrity scale used is, for example, the ASTM grade scale (see ASTM E2660 standard).

[0056] Figure 1 shows a part 10 whose integrity is to be checked, in the form of a mechanical test specimen. However, the inspection procedure applies to any type of part, regardless of its shape, dimensions, material, function, etc. In particular, but not exclusively, part 10 may be a cast part made of a metal alloy. Part 10 may also be a part from the aeronautical field, for example, an aircraft engine component. Examples include aircraft engine bearings or bladed wheel components such as low-pressure turbine distributors.

[0057] The inspection method includes obtaining an image 20 of the part 10 from at least one imaging direction 12. Specifically, the aim is to obtain an image of the part 10 that can be compared to reference images associated with the material integrity grades of the aforementioned scale. If these reference images are two-dimensional, the image 20 of the part 10 may also be two-dimensional. However, more generally, the image 20 of the part 10 may have the same dimensionality as the reference images.

[0058] The imaging direction 12 can be chosen randomly or according to a predetermined program.

[0059] According to one example, the image 20 of part 10 is obtained by a non-destructive imaging method such as radiography, in particular 2D radiography. However, any other method may be used, in particular a method similar to the method used to obtain the reference images.

[0060] For example, as an alternative to 2D radiography, it is possible to first obtain a three-dimensional image of part 10, for example by 3D tomography, and then to project this three-dimensional image along the imaging direction 12 in order to obtain the image 20 of part 10. 3D tomography can offer better image quality than 2D radiography.

[0061] As previously stated, the control process includes obtaining a mask 22 of the part 10. The mask 22 can be obtained as a mask of the image 20 of the part 10. The mask 22 can reproduce, in the dimensionality of the image 20 or, better, in the dimensionality of the reference images, the contours of the part 10. Obtaining such a mask 22 can be achieved by image processing techniques known, as such, to a person skilled in the art.

[0062] As previously stated, in the aforementioned scale, at least one reference image is associated with each material integrity grade 1, 2, ..., N. The reference image represents the typical material integrity observed for a part of the corresponding grade. As illustrated in [Fig. 1], the image of part 20 is compared to each grade 1, 2, ..., N. To do this, the mask 22 is applied to the reference images associated with at least two of the grades, preferably to each of the grades. For the sake of brevity, the following description and [Fig. 1] detail the application of the mask 22 to a reference image associated with one of the grades, but the same operations can be performed for all reference images associated with a given grade, and this for all desired grades in the aforementioned scale.

[0063] Thus, with reference to [Fig. 1], the mask 22 is applied to a reference image 24 in at least one position and at least one orientation, or even in a plurality of positions and / or a plurality of orientations. In this description, the position refers to the relative coordinates of a given (arbitrarily chosen) point of the mask 22 with respect to the reference image 24, while the orientation quantifies the angle of rotation of the mask 22 around said given point with respect to a starting (arbitrarily chosen) orientation of the mask 22. In this case, [Fig. 1] represents the same reference image 24 to which the mask 22 is applied in two different (position; orientation) pairs. From one application of the mask 22 to another, at least one of the position and orientation of the mask 22 with respect to the reference image 24 may vary.The number of mask 22 applications per reference image 24 can vary; however, it is desirable that the mask 22 applications be well distributed across the entire reference image 24, typically covering at least half of the reference image 24 cumulatively. In this way, the loss of information in the reference image 24 is limited and the impact of the choice of the position or orientation of mask 22 is reduced.

[0064] According to one example, the mask 22 can be applied to the image several times using a random sampling method such as a Monte Carlo method. The random sampling method can be used to determine successive positions and / or successive orientations. Geometric constraints can also be imposed to prevent the mask 22, at the determined position and orientation, from falling outside the reference image 24.

[0065] Each application of the mask 22 to the reference image 24 produces a comparative image 30 associated with the corresponding position and orientation. The comparative image 30 is an image of the same shape as the part 10 or its image 20, thanks to the mask 22, whose material integrity level reflects the grade of the reference image 24 from which it originates. In other words, the comparative image 30 can be associated with the grade of the corresponding reference image 24.

[0066] Moreover, as illustrated in [Fig.1], when the mask 22 is applied in a plurality of positions and / or a plurality of orientations, a plurality of comparative images 30 can be obtained from the same reference image 24.

[0067] To make the comparison between the comparative images 30 and the image 20 of part 10 more reliable, a fitting 14 can be performed to obtain the image 20 of part 10. The fitting 14 includes a digital transformation applied (for example, to a provisional image of part 10) to obtain the image 20. For example, Adjustment 14 concerns the resolution or color distribution of the resulting image 20. Alternatively, or in addition, image 20 can be modified to neutralize the difference in thickness between part 10 and the samples used to obtain the reference images 24.

[0068] By way of example, the reference images 24 can be obtained from flat samples having a given thickness, known on the scale considered (for example, given by a standard such as the ASTM standard). Conversely, the part 10 may not be flat but, for example, cylindrical (in particular axisymmetric) or otherwise. It is possible to apply a geometry correction function to the image (for example, a sinusoid to transform a cylinder with a circular cross-section into a plate). Alternatively, particularly when the geometry correction is more complex to formulate, it is possible to adjust the image 20 of the part based on the value of the sound material in the reference image.

[0069] More specifically, the adjustment 14 may include assigning to a pixel of material in the image 20 of the part 10 a value determined from a value of the healthy material on one of at least one reference image 24.

[0070] According to an example, the following procedure can be followed: a 3D mask of the part 10 is created and the number of voxels Ni is counted along the imaging direction 12, at each location i transverse to this imaging direction 12. A constant a; is then determined such that the product N,xa; is equal to the average value of a pixel of sound material, said average value being able to be determined with reference to the reference image 24 of the grade having the best material integrity.

[0071] The value a is then assigned to the material voxels of part 10 at location i. To account for the variance in the reference image 24, a noise matrix with the same variance as the reference image 24 (preferably a representative reference image of sound material) can be added to the image 20 of part 10. The noise can follow a normal distribution. The defect voxels are assigned a different value, for example, zero.

[0072] It should be noted that the fit 14 varies for each position i, which makes it possible to virtually transform any geometry into a plate. It is within the scope of those skilled in the art to adapt this method if the samples used as the basis for the reference images 24 have a geometry other than a plate.

[0073] In [Fig.1], image 20 represents the image of part 10 after adjustment 14, the adjustment 14 being able to be carried out on a provisional image of part 10.

[0074] The control method also includes, for at least one comparative image and preferably for each comparative image 30, the calculation 32 of a metric M(Xrx, Xref) representative of the difference in material integrity between the comparative image 30, denoted Xref, and the image 20 of the part 10, denoted XRX.

[0075] The metric can compare at least one of the Shannon entropy and the size of a defective area between the comparison image and the image of the part. However, other measures can be incorporated, as an alternative or in addition, to the metric.

[0076] In an example that combines these two measures, the metric can be of the form: m(xrx, Xref) = \fH x (H(X„f)-H(XRX))lH(xRX) + fA x ( A„.f- ARX) / ARX\ •> where ff and A are weights respectively associated with the Shannon entropy and the size of the defective area, H(.) measures the Shannon entropy of an image (here, a comparative image or part image), Aref represents the size of a defective area on the comparative image, and ARX represents the size of a defective area on the part image. Preferably, f / 4 are percentages such that fH + jA = 100%. For example, we choose values ​​of 50% for these two weights.

[0077] Shannon entropy can be calculated in a way known in itself.

[0078] The size of a defective area can be equal to the number of pixels, preferably consecutive, whose value exceeds a threshold representative of sound material. For example, starting with a first pixel whose value exceeds said threshold, all pixels contiguous to this first pixel and whose value also exceeds said threshold are counted, in all dimensions of the image (here, for example, in two dimensions). The number of pixels thus obtained represents an area of ​​the defect. The size of a defective area can correspond directly to this number of pixels, or to an effective dimension such as the apparent side of the defective area (square root of the defect area) or the apparent radius of the defective area (radius of the circle with the same area as the determined number of pixels).

[0079] The metric calculation 32 returns, for each comparative image 30, a value reflecting the level of difference between the image 20 of part 10 and the comparative image 30 under consideration, it being recalled that each comparative image 30 is also associated with a material integrity grade. Thus, the different values ​​of the metric reflect the difference between the image 20 of part 10 and the different grades on the scale.

[0080] At this stage, however, the metric values ​​obtained are dependent on the initially selected imaging direction 12. To reduce this dependence, it is possible to perform iterations 34 of the previously described steps by modifying the imaging direction 12, and thus applying the previously described steps for each chosen imaging direction 12. Thus, the inspection process can include obtaining an image 20 of the part 10 from several imaging directions 12. Several images 20 are then obtained, to which all or part of the previously described steps are applied.

[0081] The imaging directions 12 can be selected randomly or in any other way, for example in an evenly distributed manner. Each iteration 34 can be followed by a rotation 16 of the part 10.

[0082] Once the metric has been calculated, if necessary for all desired imaging directions, the inspection process assigns to the part 10 the grade for which the difference in material integrity is minimized. Typically, the smallest metric value obtained is selected, and the part 10 is assigned the grade of the comparative image 30 for which this smallest metric value was obtained. When there are several imaging directions, the grade is preferentially assigned to the part based on the grades assigned for each imaging direction. When several comparative images 30 are considered for the same grade, the metric values ​​obtained for that same grade can be combined, for example, by calculating an average for the grade in question.Other combinations are possible, however, including calculating a minimum or maximum among the metric values ​​obtained for the same grade.

[0083] Optionally, the inspection process may further include a step 38 for validating or rejecting part 10 based on the grade assigned to it. Validation, especially when the imaging method is non-destructive, confirms the quality of the industrial process in which part 10 is used. For example, a threshold grade may be predetermined based on the application of part 10, and step 38 consists of verifying whether the calculated grade for part 10 exceeds said threshold grade. Depending on the case, the part may be rejected for insufficient material integrity or, conversely, validated for acceptable material integrity.

[0084] The preceding description applies to actual parts and can be transposed to digital parts (such as digitally generated synthetic microstructures). For digital parts in particular, traditional imaging methods may not be applicable, and other methods may be necessary to obtain an image of the part along a given imaging direction.

[0085] Thus, according to one example, obtaining an image of the part may include obtaining a three-dimensional image of the part, assigning a first value to the voxels of sound material and a second value to the voxels of defects, and projecting the three-dimensional image along at least one imaging direction. The principle underlying this method is illustrated in [Fig. 2].

[0086] The three-dimensional image 40 of the part can be a three-dimensional tomography, either actually obtained by 3D tomography or simulated based on a digital model of the part. For the sake of simplification, but without loss of generality, the three-dimensional image 40 is here schematically represented as a cylinder.

[0087] In this three-dimensional image 40, it is possible to know which voxels represent sound material and which voxels represent a defect: either this information is given following the construction of the three-dimensional image (typically from a digital model), or the identification of the defect voxels can be obtained otherwise, for example by a semantic segmentation method of defects in the three-dimensional image 40.

[0088] A first value, for example a non-zero constant, is then assigned to the voxels of sound material, and a second value, for example a zero value, to the voxels of defects. This can reflect the fact that defects are most often characterized by an absence of material. More practically, this can reflect the fact that in grayscale, material appears gray while defects, often characterized by a void, appear black.

[0089] The three-dimensional image 40 is then projected along the imaging direction 12 as illustrated in [Fig. 2], as if 2D imaging of the part were being performed. For example, the value of a pixel in the resulting image is calculated based on the voxel values ​​projected onto that pixel.

[0090] In this context, Beer-Lambert's law states that the intensity of a monochromatic beam passing through a material decreases with increasing depth of the inspected material. This reduction in intensity, quantified by a decreasing exponential, can be approximated for assumed small thicknesses by a summation. Consequently, the intensity through a material can be rewritten as the sum of matter voxels along the imaging direction 12. More precisely, the value of a pixel in the image resulting from the projection (i.e., the image 20 of the part 10) can be equal to the sum of the voxel values ​​projected onto that pixel. Void voxels, since they are assigned a value of zero, do not contribute to the attenuation of the simulated imaging beam.

[0091] For the sake of completeness, it should be noted that the assumption of linear variation of the attenuation excludes the effects of beam hardening and scattering. However, the consequences of these effects are assumed to be negligible to the first order.

[0092] Thus, the proposed inspection method makes it possible not only to assign a grade to a physical (real) part for industrial validation, but also to assign a grade to a digital part, for example, to annotate images of parts (or even parts directly) to create a training dataset for a machine learning model, with the aim of better predicting part lifespan, for example, fatigue life. This avoids lengthy and costly physical testing campaigns.

[0093] Although the present description refers to specific embodiment examples, modifications may be made to these examples without departing from the scope general of the invention. Furthermore, individual features of the various embodiments illustrated or mentioned can be combined in additional embodiments. Therefore, the description and drawings should be considered in an illustrative rather than restrictive sense.

Claims

Demands

1. A method for inspecting a part (10) with reference to a scale comprising several material integrity grades and at least one reference image (24) associated with each grade, the method comprising the following computer-implemented steps: - obtaining an image (20) of the part (10) from at least one imaging direction (12); - obtaining a mask (22) of the part (10); - applying the mask (22) to at least one reference image (24) from at least one position and at least one orientation in order to obtain at least one comparative image (30) associated with the grade of the corresponding reference image (24);- for each comparative image (30), the calculation (32) of a metric representing the difference in material integrity between the comparative image (30) and the image (20) of the part (10) - on the basis of the calculated metric, the assignment (36) to the part (10) of the material integrity grade for which the difference in material integrity is minimized.;

2. A control method according to claim 1, wherein the image (20) of the part (10) is obtained by a non-destructive imaging method, optionally wherein the image (20) of the part (10) is a radiograph.

3. A control method according to claim 1 or 2, comprising obtaining an image (20) of the part (10) from several imaging directions (12), and wherein the grade is assigned to the part (10) according to the grades assigned for each of the imaging directions (12).

4. Inspection method according to claim 3, wherein the grade assigned to the part (10) is an average of the grades assigned for each of the imaging directions (12) or is the grade for which the difference in material integrity is minimized over the different imaging directions (12).

5. A control method according to any one of claims 1 to 4, in which the mask (22) is applied to at least one reference image (24) in a plurality of positions and / or a plurality of orientations, whereby a plurality of comparative images (30) is obtained for said at least one reference image (24).

6. A control method according to claim 5, wherein the mask (22) is applied according to a random sampling method such as a Monte Carlo method.

7. A control method according to any one of claims 1 to 6, wherein the metric compares at least one of the Shannon entropy and the size of a defective area between the comparison image (30) and the image (20) of the part (10); optionally wherein the metric is defined by \fH x(H(Xn.f)x (A„rARX) / ARX\-where ff A are weights associated with the Shannon entropy and the size of the defective area, H(.) measures the Shannon entropy, ^ref represents the comparison image (30), XRX represents the image (20) of the part (10), Aref represents the size of a defective area on the comparison image (30) and ARX represents the size of a defective area on the image (20) of the part (10).

8. A control method according to any one of claims 1 to 7, wherein obtaining the image (20) of the part (10) comprises obtaining a three-dimensional image (40) of the part (10) comprising a plurality of voxels, assigning a first value to the voxels of sound material of the plurality of voxels and a second value to the voxels of defect of the plurality of voxels, and projecting the three-dimensional image (40) in at least one imaging direction (12), wherein a value of a pixel of the resulting image (20) is calculated as a function of the values ​​of the voxels that are projected onto that pixel.

9. A testing method according to any one of claims 1 to 8, comprising a step of validating or rejecting (38) the part (10) according to the grade assigned to the part (10).

10. A program set comprising instructions for carrying out the steps of the control process according to any one of claims 1 to 9 when said program set is executed by at least one computer or microprocessor.

Images