Anomaly detection in an aeronautical part

Abstract

The invention relates to a method for detecting an anomaly in a plurality of predefined zones of a single aeronautical component, the method comprising: - supplying an image to be inspected (Img) to a detection system (300) comprising a backbone neural network (BB) configured to supply an intermediate representation (RI) of the image to be inspected (Img), and localisation neural networks (LA, LB) that are respectively associated with the zones and each designed, on the basis of the intermediate representation (RI), to search for one or more anomalies and to locate each anomaly found in the image to be inspected (Img); the detection system (300) having been previously trained, for each zone, by a step of supplying, to the backbone network (BB), training images, and, on the basis of the training images of the relevant zone, an adaptive adjustment step during which parameters of the backbone network (BB) and of the localisation network (LA, LB) that is associated with the relevant zone are adjusted without adjusting the parameters of the one or more other localisation networks (S, LB); and - obtaining the one or more anomalies and their location as output from the localisation network (LA, LB) associated with the zone of the image to be inspected (Img).

Description

Description TITLE: DETECTION OF ANOMALY IN AN AERONAUTICAL PART Technical field of the invention

[0001] The present invention relates to a method for detecting an anomaly in several predefined room areas, an associated computer program, and an anomaly detection device.

[0002] The present invention relates in particular to the field of non-destructive testing of parts, in particular aeronautical parts. Technological background

[0003] In the US patent application published under number US 20110182495 A1, conventional descriptors characterizing the texture of the part (wavelet analysis) are used, and a deviation value of the image is calculated with respect to an image with reference texture of the healthy part, to detect possible anomalies without any statistical method. This approach, however, has limitations because it requires obtaining a reference texture, which can be difficult due to the complexity and heterogeneity of the geometry of the part itself. Thus, to treat different areas of the part, it would be necessary to establish different mathematical models with different textures for the different areas.

[0004] In the US patent published under number US 8,238,635 B2, a comparison of a test image with a healthy reference image (the latter possibly obtained by averaging several healthy images) is carried out. An elastic or rigid registration to superimpose the two images and to calculate the residual image (subtraction of the two registered images) is carried out beforehand. This solution therefore has the disadvantage of requiring a registration, which is a fairly expensive operation in terms of calculation. In addition, the exploitation of the residual image proves complicated as soon as the acquisition conditions (sensor brightness, etc.) of the images change slightly.

[0005] In the US patent application published under number US 20090066939 A1, instead of using statistical learning, descriptors are manually designed to detect each type of defects sought, for example by exploiting the Hough transform for the search for right, or connected filters for shape filtering. This approach is clearly not optimal in the presence of anomalies and very heterogeneous areas of the part.

[0006] In the French patent application published under number FR 3058816, automatic statistical learning is proposed. First, the image acquired using an X-ray generator is first divided into several thumbnail images. Then, the thumbnails are automatically classified with a label representing the presence of anomalies in the thumbnail. This solution has the disadvantage of requiring a prior division of the images into small thumbnails, then a reconstruction. In addition, this solution does not allow for precise localization of anomalies.

[0007] It may thus be desirable to provide an anomaly detection method which makes it possible to overcome at least some of the aforementioned problems and constraints. Summary of the invention

[0008] There is therefore proposed a method for detecting an anomaly in several predefined zones of the same part such as a blade, the zones having different structures such as different geometries and / or different materials and / or different thicknesses and / or the presence or absence of a cavity and / or cavities of different sizes, each zone being associated with at least one predefined view of this zone, characterized in that it comprises: - obtaining an image, said to be inspected, of one of the zones according to one of the predefined views of this zone;- providing the image to be inspected to a detection system comprising: a backbone neural network designed to receive the image to be inspected and to provide an intermediate representation of the image to be inspected, and localization neural networks, respectively associated with the zones, and each designed, from the intermediate representation, to search for one or more anomalies and locate each anomaly found in the image to be inspected; the detection system having been previously trained, for each zone, by: providing, to the backbone network, training images according to each predefined view of the zone considered, and, from the training images of the zone; considered, an adaptive adjustment during which parameters of the backbone network and the localization network associated with the area considered are adjusted, but not parameters of the other localization network(s), to improve detection and localization; and - obtaining the anomaly(ies) and their location at the output of the localization network associated with the area of ​​the image to be inspected.

[0009] Thus, compared to US patent application US 20110182495 A1, the invention requires only one trained detection system (with the localization networks respectively dedicated to the areas), which avoids establishing several mathematical models. Compared to US patent US 8,238,635 B2, the invention does not require comparing the image to be tested with a reference image to calculate the residual image, since once the detection device is trained, it is sufficient to provide it with the image to be tested. Compared to US patent application US 20090066939 A1, the invention uses neural networks to avoid the manual definition of descriptors. Compared to French patent application FR 3058816, the detection device receives and processes the image to be tested in its entirety, and therefore does not require a step of cutting the image to be tested.

[0010] The invention may further comprise one or more of the following additional features, in any technically possible combination.

[0011] In practice, each localization network is associated with several reference boxes and each localization network is designed to provide, on the one hand, a position, in the image to be inspected, of one of the reference boxes of the area represented on the image to be inspected and, on the other hand, a probability of presence of an anomaly in the reference box at this position.

[0012] Thus, a more precise localization than with the state-of-the-art networks mentioned above can be obtained. Indeed, these state-of-the-art networks are classification networks using activation maps of neural areas, these neural areas being respectively associated with very general parts of the image. Thus, they only allow localization in these general parts, which is much less precise than using boxes.

[0013] Also in practice, the reference boxes are different for at least one of the location networks compared to the other location network(s), for example in number and / or size.

[0014] Also in practice, the method further comprises, for each zone: - a determination, in the training images representing the zone considered, of boxes surrounding each anomaly previously found in the training images; - a partitioning of the boxes into several groups of similar boxes; and - for each group found, a determination of a reference box from the boxes of the group.

[0015] Also conveniently, the determined reference box has dimensions derived from the dimensions of the boxes in the group.

[0016] Also in practical terms, the method further comprises, for each predefined zone, a classification network designed to classify, into predefined classes, the anomalies found by the localization network associated with the zone considered.

[0017] Also practically, the part is a blade comprising a root and a blade, and in which the predefined zones are formed from the root and the blade.

[0018] Also in practice, the blade has one or more cavities defining a certain percentage of void in relation to the total volume of the blade, this percentage of void being greater than that of the root, the percentage of void of the root being able to be zero signifying the absence of cavity.

[0019] Also practical, the blade and the foot have different thicknesses.

[0020] There is also provided a computer program downloadable from a communication network and / or recorded on a computer-readable medium, characterized in that it comprises instructions for executing the steps of an anomaly detection method according to the invention, when said program is executed on a computer.

[0021] An anomaly detection system is also proposed, characterized in that it comprises: - a backbone neural network designed to receive an image to be inspected and to provide an intermediate representation of the image to be inspected, and - localization neural networks respectively associated with predefined zones of the same part such as a blade, each zone being associated with at least one predefined view of this zone, the zones having different structures such as different geometries and / or different materials and / or different thicknesses and / or the presence or absence of a cavity and / or cavities of different sizes, and each designed, from the intermediate representation, to search for one or more anomalies and to locate each anomaly found in the image to be inspected; the detection device having been previously trained, for each area, by: - ​​a supply, to the backbone network, of training images according to each predefined view of the area considered, and - from the training images of the area considered, an adaptive adjustment during which parameters of the backbone network and of the localization network associated with the area considered are adjusted, but not parameters of the other localization network(s), to improve detection and localization. Brief description of the figures

[0022] The invention will be better understood with the aid of the following description, given solely by way of example and with reference to the appended drawings in which: - Figure 1 is a perspective view of a blade that can form a part to be inspected, - Figure 2 is a perspective view of a system for acquiring images of parts to be inspected, - Figure 3 is a functional diagram of a device for detecting an anomaly in a part to be inspected, - Figure 4 is a block diagram of a method for detecting an anomaly in a part to be inspected, - Figure 5 is similar to Figure 1, and further illustrates two areas of the blade, namely a root and a blade, - Figure 6 groups together an example of training images for training the anomaly detection device, - Figure 7 illustrates the training images of Figure 6, after annotation to define boxes surrounding anomalies respectively, - Figure 8 is a graph illustrating, for each of the two zones of the blade, the dimensions of the boxes of the training images, as well as the reference boxes determined by partitioning, - Figure 9 illustrates an example of distribution of the training images into two batches, - Figure 10 is similar to Figure 3 and illustrates the operation of the detection device during a learning phase from the training images of the first batch of Figure 9 illustrating the first zone, - Figure 11 is similar to Figure 3 and illustrates the operation of the detection device during a learning phase from the training images of the first batch of Figure 9 illustrating the second, - Figure 12 is similar to Figure 3, and illustrates the operation of the detection device during a learning phase, - Figure 13 is an image of a zone of a blade to be inspected,- figure 14 is similar to figure 3 and illustrates the operation of the detection device during a prediction phase on the image of figure 13, and - figure 15 is a functional diagram of a computer system implementing the detection device of figure 3. Detailed description of the invention,

[0023] With reference to Figure 1, a blade 100 which can form a part to be inspected in accordance with the invention, comprises a root 102 and a blade 104. The blade 100 belongs for example to a fan of a turbomachine of an aircraft.

[0024] With reference to FIG. 2, an example of a system 200 for acquiring an image of a part 202, such as the blade 100 of FIG. 1, will now be described.

[0025] The acquisition system 200 firstly comprises an X-ray generator 204 designed to generate a cone of X-rays bombarding the part 202. The acquisition system 200 may further comprise a collimator 203 fixed at the output of the generator 204 to limit the cone of X-rays. Thus, there is less dispersion of X-rays likely to cause diffusion. The acquisition system 200 further comprises a detector 206 placed behind the part 202 and designed to determine an attenuation of the X-rays, in order to provide a two-dimensional image, sometimes called a “projection”. This image is, for example, in gray levels proportional to the attenuation of the X-rays passing through the part 202.

[0026] With reference to Figure 3, an example of a system 300 for detecting an anomaly in a part, such as the blade 100 of Figure 1, will now be described.

[0027] The detection system 300 may firstly comprise an image processing module 302. The latter is designed to receive an image to be inspected Img and to possibly carry out one or more image processing operations on it. The processed image will subsequently also be called the image to be inspected and designated by the same reference Img. The image to be inspected Img represents one of Z predefined part zones. The definition of these zones will be described later in more detail.

[0028] A first possible treatment is a filtering designed to enhance the characteristic patterns of anomalies and reduce image noise.

[0029] A second possible processing is a masking around a region of interest of the image. This masking is for example carried out by an automated detection of the air and the collimator 203, for example by automatically calculating an optimal gray level threshold making it possible to separate the part from the air and the collimator 203.

[0030] A third possible operation is a normalization of the data in order to facilitate and obtain a faster convergence of the learning which will be described later. For example, the normalization is carried out to obtain a mean of 0 and a standard deviation of 1 of the gray levels of the image to be inspected.

[0031] The detection system 300 further comprises a convolutional neural network called a backbone BB. The latter is designed to receive the image to be inspected Img (after processing in the case where one or more image processing operations are planned) and to provide an intermediate representation RI of the image to be inspected Img.

[0032] The detection system 300 further comprises Hz heads (1 ≤ z ≤ Z), respectively associated with the z zones.

[0033] Each head Hz includes a localization neural network Lz, designed, from the intermediate representation RI, to search for one or more anomalies and locate each anomaly found in the image to be inspected Img, when the latter represents the associated zone z.

[0034] Each head Hz can further include a classification neural network Cz, designed, from the intermediate representation RI, to classify each anomaly found by the localization network Lz.

[0035] Each Hz head is thus designed to provide a prediction result Rz comprising the location determined by the location network Lz and, where appropriate, the classification determined by the classification network Cz.

[0036] The detection system 300 may further comprise a selection module 304, but this is not mandatory. The latter is then designed to receive an indication IND of the zone z that the image to be inspected Img represents and to deactivate the head(s) associated with zones other than that indicated by the received indication IND.

[0037] The detection system 300 further comprises a drive module 306 designed to adjust parameters of the BB backbone network and, for each Hz head, the parameters ϑ Lz of the Lz location network and, where applicable, the parameters of the Cz classification network. This adjustment is carried out during a learning phase which will be described later.

[0038] With reference to FIG. 4, a method 400 for detecting an anomaly according to the invention will now be described.

[0039] The method 400 first comprises the learning phase mentioned above, which comprises the following steps.

[0040] During a step 402, the z zones are defined. A z zone is a part of a part, in particular a homogeneous part, for example with a particular geometry and / or material.

[0041] The z zones are areas of the same part, for example of a blade like the one in Figure 1. In the latter case, a first zone z = A can be the root of the blade and a second zone z = B can be the blade of the blade, as illustrated in Figure 5. The choice of the different z zones is for example made taking into account the heterogeneity of the part to be inspected, that is to say the geometric and physical specificities linked to the different zones of the part. The z zones thus have different structures. By "different structures" we mean for example: different geometries and / or different materials and / or different thicknesses and / or the presence or absence of a cavity and / or cavities of different sizes. For example, a blade root generally has few cavity, unlike the blade. The blade thus has a certain percentage of void space compared to its total volume due to the presence of one or more cavities. The root also has a certain percentage of void space compared to its total volume (which may be zero in the absence of a cavity). The percentage of void space in the blade is then greater than the percentage of void space in the root, for example at least 10% greater.

[0042] During a step 404, for each zone z, one or more views of the zone z considered are defined. Subsequently, it will be considered that only one view is defined for each zone z.

[0043] During a step 406, training images I* are obtained. These training images I* come from several training pieces, for each view of each zone z.

[0044] For example, the acquisition system 200 is used to acquire the training images I*. Each training image I* is preferably obtained from several projections of the same part, according to the same view, for example by averaging these projections. This makes it possible to reduce the acquisition noise.

[0045] In Figure 6, an example of training images I*A1, I*A2, I*A3, I*B1, I*B2, I*B3 is shown, with areas z = A and z = B in Figure 5. The hatched parts on these training images I*A1, I*A2, I*A3, I*B1, I*B2, I*B3 represent anomalies. Of course, in practice, the number of training images I* will be much higher.

[0046] During a step 408, the training images I* can be processed by the processing module 302, in order to facilitate analysis by the detection device 300.

[0047] During a step 410, one or more anomalies are searched for in each training image I* and each anomaly found is located. This step is for example carried out by a human operator.

[0048] For example, the location of each found anomaly includes a definition of a box, called an annotated box BOI*, which is rectangular and encompasses the found anomaly. This definition includes, for example, a height and a width of the annotated box BOI*, as well as a position of the annotated box BOI* in the training image I*.

[0049] Each anomaly found can also, for example, be classified by the human operator according to a predefined classification (for example: inclusion, excess thickness, residue, excess thickness, etc.).

[0050] For example, referring to Figure 7, the human operator defines the boxes annotated BOI* A2 , BOI* A3 , BOI* B1 , BOI* B3around the anomalies of the training images I* A2 , I* A3 , I* B1 , I* B3 , respectively.

[0051] The training images I* are preferably representative of the variability of the parts to be inspected as well as of the different homogeneous areas of the part to be inspected. The number of training images with anomalies must be sufficient to be able to teach the detection device 300 characteristics representative of each anomaly.

[0052] In order to create diversity in the training images I*, it is possible, during a step 411, to add transformations to at least part of the training images I* (for example, rotations, addition of noise or blur, modification of the contrast, change of the brightness, cropping, etc.).

[0053] At the end of step 410 and, where appropriate, step 411, a training database is thus obtained comprising the training images I* and, for each of them: an indication IND of the zone z represented on the training image and a training result comprising, for each anomaly found, a location of this anomaly with possibly a classification of this anomaly in the case where the anomalies are classified. As explained previously, the location comprises for example a box annotated BOI* encompassing the anomaly found.

[0054] During an additional step 412, reference boxes AN (generally called “anchors” in English) are determined for each zone z. The reference boxes are preferably different from one zone z to another. Each reference box AN is characterized by its dimensions, for example a height and a width.

[0055] To determine the reference boxes AN, the dimensions of the annotated boxes BOI* associated) at least a part of the training images I* of each zone z are for example partitioned (from the English "clustering") to obtain one or more groups ("clusters" in English) of similar dimensions of the annotated boxes BOI*. For example, a partitioning by K-means (of (the English "K-means") is carried out. An AN reference box is then associated with each of the groups found, with dimensions derived from the dimensions of the group considered, for example by taking a centroid of the group considered.

[0056] For example, referring to Figure 8 where the box width is on the abscissa and the box height on the ordinate, the partitioning can give four groups for zone A (blade), i.e. four reference boxes AN A1 , AN A2 , AN A3 , AN A4, and three groups for zone B (blade roots), i.e. three AN reference boxes B1 , AN B2 , AN B3 .

[0057] During a step 414, the detection system 300 is trained by supervised learning from the training images I* and the associated training results. For this, each training image I* is provided to the backbone network BB and the training module 306 adjusts the parameters ϑ BB of the BB backbone network and the Hz head associated with the zone z represented on the training image I* considered, to improve detection and localization in this zone z. On the other hand, the parameters of the other head(s) are not adjusted from the training image I* considered. Thus, the parameters ϑ Lz , ϑ Cz Hz heads are adjusted with the training images of the z zones associated with them respectively. In other words, the other heads are left unchanged.

[0058] The training aims, for example, first of all to associate, with each annotated box BOI* surrounding an anomaly in the training image I*, a reference box AN among those associated with the zone z represented in the training image I* and a position of this reference box AN, so that the reference box AN at this position is close to the annotated box BOI*. The training aims, for example, furthermore to associate, with each annotated box BOI* surrounding an anomaly in the training image I*, a high probability of anomaly in the repositioned reference box AN, as well as the annotated anomaly class when classes are used.

[0059] It will be appreciated that no predefined function is provided for the BB backbone network, nor any predefined form or meaning for the RI intermediate representation. It is during training that the BB backbone network acquires a certain, not explicit, function. In other words, the output of the BB backbone network forms a hidden layer of the complete network (BB backbone network and Hz head networks) and the RI intermediate representation, also called hidden representation (of hidden representation), has internal abstract features or concepts, which the complete network learns during its training.

[0060] Unlike the method described in the French patent application published under number FR 3058816 A1, which performs a tiling of the image to work on thumbnails, each training image is provided in full to the BB backbone network. This is advantageous in particular in terms of computation time because a single pass through the detection system 300 is sufficient to obtain a prediction on the complete training image. In addition, it is not necessary to carry out a post-processing step to assemble the predictions obtained on each thumbnail.

[0061] For example, the parameters are adjusted iteratively, for example, in batches of training images, and the parameter adjustment is performed, for example, by a gradient descent algorithm. As is known, a gradient descent algorithm can find the minimum of any convex function by gradually converging towards it.

[0062] Thus, in the case of using batches, the training images I* are divided into batches Kj (1 ≤ j ≤ J), each batch Kj being able to contain training images from different z zones.

[0063] The training images I* of each batch Kj are then successively provided as input to the detection device 300 to respectively obtain prediction results Rj. These prediction results Rj are compared to the training results of the training images of batch Kj. From this comparison, a partial detection loss function ℐ ! "#( $ %% , $ "# ) and, where appropriate, a partial classification loss function , are determined for the lot Kj considered and for each zone z.

[0064] Partial detection loss functions of the same zone z are then combined, for example added, to obtain a detection loss function ℐ "#( $ %% , $ "# ) for the detection network Lz of each zone z. The detection loss function ℐ "# ($ %% , $ "# ) is for example the L1 loss function or the smooth L1 loss function (from the English “smooth L1”).

[0065] The case similarly, the partial classification loss functions ℐ ! "# ($ %% , $ &# ) of the same zone z are then combined, for example added, to obtain a classification loss function ℐ &#( $ %% , $ &# ) for the classification network Cz of each zone z. The classification loss function ℐ &# ($ %% , $ &#) is for example the cross entropy loss function or the focal loss function.

[0066] Thus, by denoting Z the total number of zones z, 2xZ loss functions are determined: Z loss functions for localization and Z loss functions for classification.

[0067] Then, for each zone z, the parameters ϑ BB of the BB backbone network, the parameters ϑ Lz of the detection network Lz and, in the case of the parameters ϑ Cz of the classification network Cz of the head Hz are updated from the loss functions ℐ "# ($ %% , $ "# ) and ℐ &# ($ %% , $ &# ) determined. For example, the gradient descent algorithm uses the following equation for each area z: [Math.1] 3.

[0068] Thus, each training image I* participates in the development of the loss function(s) ℐ "# ($ %% , $ "# ) and ℐ &# ($ %% , $ &# ) associated with the area z represented on this training image I*. Thus, by using these loss functions to update the parameters, the ϑBB parameters of the BB backbone network are adjusted from each training image I*, regardless of the area z represented, while only the ϑ parameters Lz , of the Hz head associated with the z zone represented on the training image I* are adjusted, and not the parameters of the or of the other heads.

[0069] For example, with reference to Figure 9, a first batch K1 comprises the training images I* A1 , I* A2 , I* B3 and a second batch K2 includes the training images I* B1 , I* B2 , I* A3 .

[0070] Training Images I* A1 , I* A2 , I* B3 of the first batch K1 are thus first supplied to the detection device 300. With reference to figure 10, for the training images I* A1 , I* A2 , the HA head then provides the results R A1 , R A2 . Referring to Figure 11, for the training image I* B3 , the HB head then provides the result R B3 . From the R results A1 , R A2 , R B3 , the following partial loss functions are determined: ℐ + "4( $ %% , $ "4 ) , ℐ + &4( $ %% , $ &4 ) , ℐ + "%( $ %% , $ "% ) and ℐ + &%( $ %% , $ &% ) .

[0071] In the same way, the training images I*B1, I*B2, I*A3 of the second batch K2 are provided to the detection system 300 to obtain the results RB1, RB2, RA3 and, from them, the partial loss functions ℐ5 "4 ($ %% , $ "4 ), ℐ5 &4 ($ %% , $ &4 ) and ℐ5 "% ($ %% , $ "% ) and ℐ5 &% ($ %% , $ &% ). 12, the loss functions ℐ "4 ($ %% , $ "4 ) = ℐ + "4 + , ℐ "% = ℐ + "% + , + ϑCA, ϑCB of the detection system 300, for example according to equation [Math.1].

[0073] A new training iteration can then begin, again providing the training images I* to the detection system 300.

[0074] The method 400 then comprises a prediction phase.

[0075] During a step 416, an image to be inspected Img is obtained, for example by means of the acquisition system 200. This image to be inspected Img represents one of the zones z according to a view among the predefined view(s) of this zone z. For example, as illustrated in FIG. 13, the image to be inspected Img represents the blade (zone A) of a vane, according to the single predefined view for the blade.

[0076] During a step 418, the image to be inspected Img must preferably be processed in the same way as the training images I* by the processing module 302.

[0077] During a step 420, the image to be inspected Img is provided to the detection system 300, if necessary after processing.

[0078] During a step 422, when the selection module 304 is present, the zone z represented on the image to be inspected Img is indicated to the selection module 304, which deactivates for example the head(s) associated with the zone(s) other than that indicated. For example, as illustrated in FIG. 14, the zone A is indicated to the selection module 304. In response, the latter deactivates the head HB and leaves the head HA active. In the absence of the selection module 304, all the heads inspect for example the image to be inspected Img, but only the output of the head associated with the zone of the image to be inspected Img is relevant. The outputs of the other heads will a priori be of no interest and can be rejected.

[0079] During a step 424, the backbone network BB provides an intermediate representation RI of the image to be inspected Img.

[0080] During a step 426, the head Hz associated with the zone z represented on the image to be inspected Img searches for one or more anomalies and locates each anomaly found, from the intermediate representation RI.

[0081] During a step 428, a result R of this search and this localization is recovered at the output of the head Hz associated with the zone z represented on the image to be inspected Img.

[0082] For example, the result R includes, for each reference box AN associated with the zone z represented on the image to be inspected Img, a position of this reference box AN in the image Img and a probability of presence of an anomaly in this reference box AN. An anomaly can thus be considered as found when this probability is greater than a predefined threshold.

[0083] The Hz heads can further be designed to perform a segmentation of each anomaly found, i.e. a classification of each pixel of the AN reference box containing this anomaly, for example according to two values: one value indicating a pixel with anomaly and another value indicating a pixel without anomaly. This segmentation is for example carried out by adding in each head a segmentation neural network, which will be trained in a similar way to the Lz detection and Cz classification networks.

[0084] The prediction phase can be repeated for several parts to be inspected, for example for several blades to be inspected.

[0085] With reference to FIG. 15, the detection device 300 is for example a computer system comprising a data processing unit 1502 (such as a microprocessor) and a main memory 1504 (such as a RAM memory, from the English "Random Access Memory") accessible by the processing unit 1502. The computer system further comprises for example a network interface and / or a computer-readable medium, such as for example a local medium 1506 (such as a local hard disk) or a remote medium (such as a remote hard disk and accessible via the network interface through a communication network) or even a removable medium (such as a USB key, from the English "Universal Serial Bus", or a CD, from the English "Compact Disc" or a DVD, from the English "Digital Versatile Disc") readable by means of an appropriate reader of the computer system (such as a USB port or a CD and / or DVD disk reader).A computer program P containing instructions for the processing unit 1502 is recorded on the local medium 1506 and / or downloadable via the network interface. This. computer program P is for example intended to be loaded into the main memory 1504, so that the processing unit 1502 executes its instructions. The instructions are for example organized into software modules respectively implementing the elements of the detection device 300, as described with reference to FIG. 3.

[0086] Alternatively, all or part of these modules could be implemented in the form of hardware modules, that is to say in the form of an electronic circuit, for example micro-wired, not involving a computer program.

[0087] In conclusion, it will be noted that the invention is not limited to the embodiments described above. It will indeed appear to those skilled in the art that various modifications can be made to the embodiments described above, in light of the teaching which has just been disclosed to them.

[0088] In the detailed presentation of the invention given above, the terms used should not be interpreted as limiting the invention to the embodiments set forth in this description, but should be interpreted to include all equivalents the prediction of which is within the reach of those skilled in the art by applying their general knowledge to the implementation of the teaching just disclosed to them.

Claims

Claims [1] Method (400) for detecting an anomaly in several predefined zones (A, B) of the same aeronautical part such as a blade, the zones having different structures such as different geometries and / or different materials and / or different thicknesses and / or the presence or absence of a cavity and / or cavities of different sizes, each zone (A, B) being associated with at least one predefined view of this zone (A, B), characterized in that it comprises: - obtaining an image, called an image to be inspected (Img), of one of the zones (A, B) according to one of the predefined views of this zone (A, B);- providing the image to be inspected (Img) to a detection system (300) comprising: • a backbone neural network (BB) designed to receive the image to be inspected (Img) and to provide an intermediate representation (RI) of the image to be inspected (Img), and • localization neural networks (LA, LB), respectively associated with the zones (A, B), and each designed, from the intermediate representation (RI), to search for one or more anomalies and locate each anomaly found in the image to be inspected (Img); the detection system (300) having been previously trained, for each zone (A, B), by providing, to the backbone network (BB), training images (I*; A1 , I* A2 , I* A3 , I* B1 , I* B2 , I* B3 ,) according to each predefined view of the zone (A, B) considered, and, from the training images (I* A1 , I* A2 , I* A3 , I* B1 , I* B2 , I* B3,) of the area (A, B) considered, an adaptive adjustment during which parameters of the backbone network (BB) and of the localization network (LA, LB) associated with the area (A, B) considered are adjusted, but not parameters of the other localization network(s) (LA, LB), to improve detection and localization; and - obtaining the anomaly(ies) and their location at the output of the localization network (LA, LB) associated with the area (A, B) of the image to be inspected (Img). [2] Method (400) according to claim 1, wherein each localization network (LA, LB) is associated with several reference boxes (ANA1, ANA2, ANA3, ANA4, ANB1, ANB2, ANB3) and wherein each localization network (LA, LB) is designed to provide, on the one hand, a position, in the image to be inspected (Img), of one of the reference boxes (ANA1, ANA2, ANA3, ANA4, ANB1, ANB2, ANB3) of the area (A, B) represented on the image to be inspected (Img) and, on the other hand, a probability of presence of an anomaly in the reference box (AN A1 , AN A2 , AN A3 , AN A4 , AN B1 , AN B2 , AN B3 ) at this position. [3] The method (400) of claim 2, wherein the reference boxes (AN A1 , AN A2 , AN A3 , AN A4 , AN B1 , AN B2 , AN B3 ) are different for at least one of the localization networks (LA, LB) compared to the other localization network(s), for example in number and / or in size. [4] Method (400) according to claim 2 or 3, further comprising, for each zone (A, B): - a determination, in the training images (I*) representing the zone (A, B) considered, of boxes (BOI* A2 , BOI* A3 , BOI* B1 , BOI* B3) surrounding each anomaly previously found in the training images (I*); - a partitioning of the boxes (BOI* A2 , BOI* A3 , BOI* B1 , BOI* B3) into several groups of similar boxes; and - for each group found, a determination of a reference box (ANA1, ANA2, ANA3, ANA4, ANB1, ANB2, ANB3) from the boxes of the group. [5] Method (400) according to claim 4, in which the determined reference box (ANA1, ANA2, ANA3, ANA4, ANB1, ANB2, ANB3) has dimensions derived from dimensions of the boxes of the group. [6] Method (400) according to any one of claims 1 to 5, further comprising, for each predefined zone (A, B), a classification network (CA, CB) designed to classify, in predefined classes, the anomalies found by the localization network (LA, LB) associated with the zone (A, B) considered. [7] Method (400) according to any one of claims 1 to 6, in which the part is a blade comprising a root and a blade, and in which the predefined zones (A, B) are formed from the root and the blade.[8] The method (400) of claim 7, wherein the blade has one or more cavities defining a certain percentage of void relative to the volume. total of the blade, this percentage of void being greater than that of the root, the percentage of void of the root being able to be zero signifying the absence of cavity. [9] Method (400) according to claim 7, in which the blade and the root have different thicknesses. [10] Computer program (P) downloadable from a communication network and / or recorded on a computer-readable medium, characterized in that it comprises instructions for executing the steps of a method (400) for detecting an anomaly according to any one of claims 1 to 9, when said program is executed on a computer. [11] System (300) for detecting an anomaly characterized in that it comprises: - a backbone neural network (BB) designed to receive an image to be inspected (Img) and to provide an intermediate representation (RI) of the image to be inspected (Img), and - localization neural networks (LA, LB) respectively associated with predefined zones (A,B) of the same part such as a blade, each zone (A, B) being associated with at least one predefined view of this zone (A, B), the zones having different structures such as different geometries and / or different materials and / or different thicknesses and / or the presence or absence of a cavity and / or cavities of different sizes, and each designed, from the intermediate representation (RI), to search for one or more anomalies and locate each anomaly found in the image to be inspected (Img); the detection device (300) having been previously trained, for each zone (A, B), by: - ​​a supply, to the backbone network (BB), of training images (I*A1, I*A2, I*A3, I*B1, I*B2, I*B3,) according to each predefined view of the zone (A, B) considered, and, - from the training images (I*A1, I*A2, I*A3, I*B1, I*B2, I*B3,) of the area (A, B) considered, an adaptive adjustment during which parameters of the backbone network (BB) and of the localization network (LA, LB) associated with the area (A, B) considered are adjusted, but not parameters of the other localization networks (LA, LB), to improve detection and localization.

Images