FR3075376A1 - NON-DESTRUCTIVE CONTROL METHOD FOR AERONAUTICAL WORKPIECE - Google Patents

Abstract

 The invention relates to a method of non-destructive testing of a metal part (10) for aeronautics using a radiographic image acquisition system (20), comprising a step of statistically learning indications (E10). including the following sub-steps:  - creating a database of simulated images of the metal part (10), as well as at least one element for identifying an indication given in said database (40); and  - From said database, implementing a statistical analysis by supervised adversary learning to generate a statistical model of indication indications in a metal part (10);  a step of calculating the probabilities of presence of indications (E2) in the part to be checked (10) to calculate the probabilities of presence of indications in said 2D image from the statistical model (MS); and  determine whether the metal part (10) is characteristic of a sound part.

Description

NON-DESTRUCTIVE INSPECTION METHOD FOR AN AERONAUTICAL PART

GENERAL TECHNICAL AREA

The invention relates to the field of non-destructive testing (CND) on industrial parts made of metallic material, in particular in the aeronautical field, using images acquired by means of an X-ray generator.

More specifically, the invention relates to the automation of NDT methods.

STATE OF THE ART

The CND is essential to control the material health of materials. For example, high pressure turbine blades are critical parts that need to be fully checked.

Each part from the production chain is individually checked to ensure that it complies with specifications. This material health check can be carried out by means of an X-ray generator which bombards the part to be inspected. The attenuation of the X-rays is observed thanks to a detector placed behind the part forming a 2D image (cf. Figure 1). The grayscale of this projected image is proportional to the attenuation of the X-rays passing through the room.

The high pressure turbine blades have a complex, hollow internal structure, allowing air to pass through to cool them. Additional thicknesses and sub-thicknesses may be present inside the dawn. In addition, other defects may appear during the part's manufacturing process: more or less dense foreign bodies, shrinkage, inclusions, residues, oxides ...

Traditionally, after acquiring a projection of the part (2D image), an operator comes to manually check this image in search of indications (critical faults, structural anomalies).

The image is coded in 16 bits (i.e. 65536 grayscale), but the control screens only allow the display of 8 bit images. The operator must then constantly adjust the contrast of the image to distinguish small irregularities in gray levels in the image. This phenomenon is accentuated for parts of variable thickness (such as blade parts), the gray levels being correlated with this variation in thickness. The operator must therefore scan the entire image while adjusting the contrast of the image. It is a relatively long and tedious operation, and a source of error.

Finally, after having manually detected all the indications of the part, the part is either accepted and mounted on the engine, or is then rejected according to the derogation process.

To allow the automation of the control, there are automatic control methods implementing supervised statistical learning in order to learn the signature of a sound part and the various indications (anomalies, structural defects). These methods offer known detection performance. We thus know the patent application FR.1661101. We also know the publication W02015 / 033044 addressing the characterization of a piece of woven composite material. However, the statistical learning carried out upstream requires the establishment of an annotated database, complete, and difficult to obtain in practice.

Indeed, the annotation of the database is:

- costly in time: operators must annotate exhaustively and precisely a very large set of images. In addition, a consensus must be reached by an expert to validate the annotated database;

- costly in money: many people are involved in the creation of this annotated database (operators, project manager, etc.);

- subject to human factors: annotating images is a tedious manual task and therefore, errors can occur. This inter and intra operator variability can disrupt statistical learning;

- a crucial step for statistical learning which relies exclusively on these annotations to learn the signatures of indications.

There is therefore a need to improve the automation of non-destructive testing methods, making them more economical and quick to implement.

PRESENTATION OF THE INVENTION

The invention implements non-destructive testing, by X-rays generating 2D images of the parts to be checked, and by an automatic analysis method using statistical methods of supervised deep learning. The invention simplifies the design of a database implemented in the context of statistical methods of supervised deep learning.

The invention provides a method of non-destructive testing of a metal part for aeronautics implementing a radiographic image acquisition system, comprising a step of learning statistical indications including the following sub-steps:

- acquire a plurality of real 2D images of a test piece likely to contain potential indications,

- from a 3D reference model of the metal part, modify said 3D model to generate at least one indication given in a given area of said model,

- simulate the operating conditions for acquiring said plurality of real 2D images and generate at least one simulated 2D image from said modified 3D model,

- adding the simulated 2D image to a database of simulated images of the metal part, as well as at least one element for identifying an identification given in said database;

- from said database and from the plurality of real 2D images, implementing a statistical analysis by supervised adversary learning to generate a statistical model of indications in a metal part;

a step for calculating the probabilities of the presence of indications in the document to be checked, comprising the following substeps:

- acquire at least a 2D image of a metal part;

- calculate the probabilities of the presence of indications in said 2D image from the statistical model; and determine if the metal part (10) is characteristic of a healthy part.

The invention makes it possible to implement a method of statistical analysis of supervised deep learning requiring much less parts manually annotated by an operator, which significantly reduces the additional workload in the factories to annotate the parts.

The invention also makes it possible to accelerate the deployment of automatic non-destructive checks based on statistical learning.

Advantageously, the invention according to the method makes the design of an annotated database more reliable, since the indications present in the simulated parts are perfectly controlled.

The invention also makes it possible to obtain an exhaustive database, in number, in type and in position of the indications. Indeed, it is possible to generate indications in areas where there are very few, and also to generate very infrequent types of indications.

The invention may also include the following characteristics:

the element for identifying an indication given in said database is a bit mask of a given indication;

advantageously, obtaining a binary mask makes it possible to have a very precise identification of the indications in a room;

the statistical learning step also comprises the sub-step consisting in estimating parameters of the 3D fitting of said test piece (PI) during the acquisition of said plurality of real 2D images; and in which the substep of simulating the operating conditions for acquiring the plurality of real 2D images is carried out on the basis of said estimated parameters of the 3D fitting of said test piece (PI) to determine registration parameters of a 3D model of the metal part (10);

the method comprises a step of validating the statistical model comprising the following sub-steps:

o acquire at least one 2D image of a reference part, o assign to said image at least one label relating to the presence of indications as a function of a visual inspection of the image, o calculate the probabilities of presence of indications in the 2D image from the statistical model (MS) and determine at least one label, o compare the determined label (s) with the label (s) assigned, and calculate a correlation indicator, o if the indicator is less than a given value, continue the learning step, o otherwise validate the statistical model.

- after the acquisition of at least one 2D image of a test piece, and / or of a reference piece, and / or of a reference piece, said acquired 2D image undergoes processing by means of treatments adapted to attenuate artifacts and / or variabilities of the 2D image;

- After generation of a simulated 2D image from said modified 3D model, said acquired 2D image undergoes processing by processing means adapted to attenuate artifacts and / or variabilities of the 2D image;

- statistical analysis is obtained by deep learning based on neural networks;

- deep learning based on neural networks is configured to perform an adversary learning algorithm;

- statistical analysis by supervised learning includes:

o for at least one simulated image and at least one real image of the plurality of real 2D images, automatic extraction of a signature from said images, determining a plurality of p descriptors, a descriptor being the result of an image analysis , o automatic classification of the signature of the simulated image making it possible to assign to said image, a label representative of the presence of an indication in the part, o back-propagation of the error of said classification of the label to minimize a cost function of the classification of the signature making it possible to attribute a label to the image, o automatic classification of the signature of the real or simulated image in a domain of the simulated image, or of the real image, o reverse propagation of the reverse the error of said classification of the domain to maximize a cost function of the classification of the signature of the thumbnail in a domain, so that the extraction of a signature is invariant according to the domain.

- the learning algorithm implemented for backpropagation is the stochastic gradient descent method.

The invention also relates to a calculation unit, comprising processing means and a memory, said unit being configured to implement a method according to one of the preceding characteristics.

The invention also relates to a computer program product, configured to be executed by a calculation unit as described above and comprising code instructions for the execution of a method according to one of the preceding characteristics.

PRESENTATION OF THE FIGURES

Other characteristics, objects and advantages of the invention will emerge from the description which follows, which is purely illustrative and not limiting, and which should be read with reference to the appended drawings, in which:

- Figure 1, already presented, illustrates a 2D image acquisition system by an X-ray generator;

FIG. 2 illustrates certain steps of an embodiment in accordance with the invention, as well as the links between a CND method according to the invention and a database generation method for implementing a CND;

- Figure 3 illustrates the various devices and products involved in the context of the invention;

- Figure 4 illustrates a calculation of the position of the flaps of a collimator from a signal according to the method of the invention;

- Figure 5 illustrates a scene modeling the acquisition of a part to be checked according to an embodiment according to an embodiment according to the invention;

- Figures 6A to 6D illustrate 3D models of indications contained in the parts to be checked according to the embodiment according to the invention;

- Figure 7 illustrates the identification of indications in an image annotated manually according to the state of the art, and in a simulated image according to embodiment according to the invention; and

- Figure 8 illustrates an adversary learning method according to the state of the art.

DETAILED DESCRIPTION

The CND can be applied at different stages of manufacturing a part 10 to be checked. The part 10 is preferably made of metallic material.

The parts on which the methods described can be implemented are preferably intended for aeronautics. It is a high pressure turbine blade for example.

FIG. 2 represents certain steps of an embodiment in accordance with the invention.

In order to be able to carry out a non-destructive check inside the part 10, a preliminary acquisition step using an imaging system 20, such as an X-ray generator, and calculation means 30 (see Figure 3A) is performed. The X-ray generator emits a beam passing through the object to be explored before being analyzed, after attenuation, by a detection system. This step provides a 2D image of part 10.

This 2D image is called a projection. A set of projections is acquired without moving, and the images can be averaged in order to reduce the acquisition noise. For example, in one embodiment, the generator 21 and the detector 22 being fixed, an articulated arm (example: typical robot of the manufacturer Kuka®) automatically grasps the piece 10 and places it between the generator 21 and the detector 22 to a position defined in advance in a range of acquisitions.

A computing unit 30 is provided for carrying out the computer processing. It includes processing means 31 such as a microprocessor and a memory 32.

With reference to FIG. 3B, the method implemented for automatically detecting indications requires a database (BDD) 36. BDD 36 is stored on at least one server 35, and a calculation unit 37, which can be the same calculation unit that for the acquisition of the 2D image, is used. The calculation unit 37 also includes data processing means 38 and a memory 39. The server 35 can be this calculation unit 37 and the BDD 36 can be stored in said memory 39.

El: Offline learning

In a first phase E1, the method includes supervised learning of the part 10 from an annotated database obtained by digital simulation.

To achieve this goal, in a first step E10 of phase E1, the method performs statistical learning, from 2D images obtained by simulation and by X-ray acquisition.

Thus, a sub-step EU of E10, comprises the acquisition of at least one real 2D image of at least one real part PI (representative of part 10) likely to contain potential indications, using the system imagery 20, such as an X-ray generator, and computing means 30.

Numerical simulation

The method also has a 3D triangular mesh of a part to be checked 10, that is to say a set of 3D coordinates of the vertices of the triangles which form the mesh and the connectivity between the vertices which defines the topology of the mesh. . This mesh is recoverable from the design of the geometry of the part made on the CATIA® software for example.

Thereafter, in order to simulate an exhaustive annotated database, the calculation unit 37 simulates indications in the part to generate a defective part.

Thus, there are several factors that come into play to generate a part containing simulated indications, such as for example in a non-exhaustive manner:

- the number of simulated indications per part;

- the typology of the indications (oxide / shrinkage / inclusion ...);

- the location of the indications in the room (in the cooling circuit / at the interfaces / in the material);

- the morphology of the indications (spherical / linear / convex / textured ...);

- the orientation of the indications; and

- the chemical composition of the indications (air / material + or dense / same material as the part, etc.).

Consequently, to insert indications into the mesh of the 3D model of the part 10, the calculation unit 37 randomly generates a computer model of each indication that we wish to insert in the dawn (a mesh of the indication). Different methods of automatic shape generation can be used depending on the types of indications to simulate.

The indications follow a basic morphology and their insertion into the material follows a given behavior law, allowing the automatic generation of a representative digital part (in size / morphology / orientation / position / ...).

A non-exhaustive list of the types of simulated indications includes: blowing / inclusion / residue / shrinkage / texture / crack / oxide / core breakage ...

For example :

- The blows 61, illustrated in FIG. 6A, are gas bubbles of spherical shape. Also, the blows are simulated by objects of spherical shape of a certain size. Their chemical composition is air.

- The inclusions 62, illustrated in FIG. 6B, have a more random form and they are simulated by generating binary objects based on a random walk algorithm. Initially, a sphere is inserted into the 3D volume, then a random vector of size equal to radius is drawn and corresponds to the center of the next sphere to be inserted in the image. This process is repeated a random number of times.

- The cracks 63, illustrated in FIG. 6C, are indications which have a cylindrical shape of length (preferably greater than the diameter of the cylinder, for example by at least a factor of 10). The orientation of these cracks is random in space.

- The residues 64, illustrated in FIG. 6D, have a much more complex shape. In order to automatically generate these complex random shapes, structured noise is used: Perlin noise. This texture noise is used in texture synthesis and parameters allow the complexity of the generated objects to be controlled. The cube opposite is an embodiment of Perlin's noise. In order to extract an object from this 3D volume, a Gaussian weighting centered in the volume is applied making it possible to limit the object (the standard deviation of the filter is lower along an axis of the volume in order to obtain objects having a slightly more "surface" shape). The result of this filter is binarized (random threshold) in order to obtain a set of binary objects. An object (a related component) is extracted from this set and isolated and then meshed.

Thus, from this 3D model, in a sub-step E12 of E10, the calculation unit 37 performs a simulation of the firing of the X-rays on said 3D model in order to generate a 2D image from the 3D triangular mesh d 'a part to be inspected 10 comprising indications (referenced 10a, cf. FIG. 5).

For this purpose, the computing unit 37 can use Beer Lambert's law which describes the attenuation of X-rays in matter, to reproduce the real image digitally:

I = I ₀ .

With I the flux passing through the part 10, Io the initial flux, μ the density of the material crossed, xo the position of the tube and Xmax the position of the detector. As the photon beam passes through the object, it is attenuated as a function of the thickness dx passed through and the attenuation coefficient μ. This attenuation coefficient depends on the energy E of the photon and the atomic number Z of the structure encountered at depth x.

In order to obtain realistic simulations, the computing unit 37 can also simulate by Monte Carlo the attenuation of photons by Compton effect (the photons change trajectory and can thus interact on an area of the detector for which they are not not representative of the materials crossed).

To realistically simulate the radiography, the calculation unit 37 can use complete software suites. For example :

- CIVA: commercial software developed by CEA; and

- GATE / GEANT4: open source software.

The configuration of the radiography simulation includes the characterization of certain parameters of the simulated acquisition system, such as for example the parameters relating to:

- Filters: thickness, material, position ...

- The collimator, thickness, material, position ...

- The part 10a to be radiographed: position in the acquisition system, material, etc.

- The source of X-rays: radius, opening angle, spectrum, intensity ...

- The detector: size, pitch, position, materials, type, gain ...

In the simulation of the acquisition, the computing unit 37 can also characterize the elements of the scene. Thus, said unit can characterize the material of the part 10 from its chemical composition, the proportion of each atom and its density.

Consequently, the calculation unit 37 defines a new alloy, in the simulation, having a spectrum of own absorptions as well as a specific density. The emission spectrum of the tube is experimentally characterized with a dosimeter and the characteristics of the RX shot are similar to the experimental conditions (voltage / intensity / exposure time, etc.).

The simulation, for example by Monte Carlo, of the photon attenuation by Compton effect is inserted by calculating the different interactions between electrons and matter, making it possible to know the impact of X-rays on the detector.

As previously seen, in the geometry of the image acquisition scene of a part to be inspected 10, an articulated robot can grasp the part 10 and automatically place it between the source 21 and the detector

22. To reproduce in simulation the precise positioning of the part between the source 21 and the detector 22, the calculation unit 37 can thus apply a 3D / 2D registration algorithm to allow matching between the 2D image obtained and the 3D position of the part. The registration algorithm provides the translations and rotations to be applied to the 3D model of the part (CAD) in order to match it with the projection obtained.

Also, in the geometry of the simulated acquisition scene, the X-ray source is placed at the origin of the coordinate system (0, 0, 0) and the center of the detector is placed at the position (0, 0, "distance detector tube ”). The positioning of the part 10 is given by the 3D / 2D registration taking into account the parameters of the acquisition system 20 so as to generate a simulated 2D image similar to the real 2D image acquired by the acquisition system. picture 20.

The geometry of the scene can also include a collimator. The latter, which can be controlled, is placed between the X-ray source 21 and the part 10 in order to reduce the artefacts linked to the scattered radiation by masking part of said part 10. The collimator has four components (preferably comprising lead, a 5mm thickness).

From a real 2D image, the positions of these four panes are automatically deduced from the masked areas on said image. Thus, with reference to FIG. 4, this shows a profile line along the x axis of the 2D image, this signal can be analyzed in order to detect the position of the flaps on the image. A simple geometric calculation makes it possible to deduce therefrom the position 41 of the flaps in 3D. The same process is carried out along the y axis to determine the position of the other two flaps.

Also, in the geometry of the simulated acquisition scene, the calculation unit 37 can insert a copper filter of the same size, and at the same position as in the real assembly of the acquisition system 20.

FIG. 5 illustrates a scene 50 in three dimensions of the acquisition of a view of a part to be checked. Said scene 50 generates the projection 53, corresponding to the simulated 2D image, and includes 3D models corresponding to the ray source 21, to a lead collimator 54, to a copper filter 55, to the part to be checked 10a and to detector 22.

Said simulated image 53 is calculated by launching rays between each point of the detector 52 and the center of the tube of the source 51. For each ray, the calculation unit 37 calculates the thickness crossed for each material, and calculates its response with Beer Lambert's law.

All the parameters of the simulation can be adjusted so that the image resulting from the simulation is as close as possible to a real image.

Consequently, by having complete knowledge of each indication, the calculation unit 37 deduces their position on the simulated image, which makes it possible to obtain a very precise simulated and annotated database 36.

Usually, in a base manually annotated by an operator, the annotation of the indications is represented in FIG. 7 by an enclosing box 71.

In the context of the method according to the invention, the database can also include the binary mask of the indications.

Such a mask, referenced 72, indicates the pixels belonging to an indication and the pixels belonging to the background.

Consequently, the characterization of the simulated database according to the method is reliable is very precise (the binary mask 72 of the indications is available and no longer only the bounding box 71).

In addition, the number of simulated parts 10a and of annotated indications is therefore advantageously potentially infinite.

In addition, the design of the simulated database makes it very easy to generate a very large number of images. This is particularly advantageous in the context of statistical learning methods, which require a lot of data to be able to obtain suitable detection and false alarm performance.

Preferably, the simulated database is constituted as follows for a given view of the part to be inspected:

- around 5000 indications distributed uniformly in all areas where these indications may appear; and

- on average 10 indications per image.

To improve the robustness of learning, the simulated images can be boosted (application of small realistic geometric transformations). Thus for a given simulated image, a hundred other simulated images are generated. This boosting of the BDD 36 makes it possible to increase the variability and thus limit the over-learning.

pretreatment

Thereafter, the images from the EU and E12 sub-steps are pretreated in a pre-processing E13 sub-step. Thus, pretreatments are preferably carried out to improve the quality of the analyzes, following the actual or simulated acquisitions of the 2D image of the part 10.

Indeed, during this acquisition, artefacts and variabilities can appear (defective pixels, noise generated by the diffused flux, diffraction ...). Preprocessings can therefore possibly be applied to the 2D image to remove and / or attenuate these artefacts / variabilities.

Thus, the calculation unit 30 or 37 can implement filtering of the noise of the image on the 2D image by using, for example, a median filter, a Gaussian filter (class of low-pass filters).

Also, the computing unit 30 or 37 can apply a normalization step to standardize the 2D images so that they are all comparable. Several standardization methods can be used independently or coupled together:

- A registration of images applied to 2D images in order to put them all in a common frame of reference in order to absorb the variability of the acquisition.

- A calculation of the median value and the MAD (Median Absolute Difference) on the pixels corresponding to part 10 for each 2D image and fixed at identical predetermined values for all 2D images.

- A calculation of the local contrast: the part 10 does not always have a constant thickness (this is never the case for a blade for example). Also, the local contrast is calculated for all the pixels of the 2D image (subtraction of the low frequencies of the standardized image) (for example by bilateral filter [1], anisotropic scattering [2], total variation filter [3], Edge-avoiding wavelets [4]).

In addition, a mask of the acquired or simulated part is generated from the 2D image, that is to say that a segmentation between two classes of pixels, air and material, is carried out. This segmentation can be achieved by automatically calculating the optimal threshold allowing the two classes to be separated.

An example of a method consists in calculating the histogram of the gray levels of the image and in maximizing the interclass variance of this histogram in order to define the threshold which separates the pixels of air and those of matter (see reference [5 ]). Other automatic thresholding methods are possible (see reference [6]).

These methods make it possible to carry out the analysis of the part only on regions of interest, relating to the areas of material of the acquired or simulated part.

Supervised statistical learning

Subsequently, in a sub-step E14 of E10, the calculation unit 37 carries out supervised statistical learning, with two types of input data: the set of simulated 2D images with their indications located in said images , and the set of real 2D images from real PI parts.

Statistical learning is mainly supported by simulated 2D images, the number of simulated 2D images preferably being significantly greater than the number of real 2D images. In addition, the indications are in the simulated 2D images.

Also, it is necessary to ensure that a statistical model generated by said learning is as efficient on real 2D images as it is on simulated 2D images.

Indeed, even if visually, the simulated images are close to the real images, they are not exactly identical.

Also, the calculation unit 37 can use adversary learning in order to transfer learning from simulated 2D images to real 2D images ("domain adversarial transfer").

FIG. 8 illustrates such an adversary learning method within the framework of a deep learning architecture by a neural network.

In a known manner, to analyze several characteristics of an image, such a neural network stacks up processing layers (layers of neurons), each layer analyzing a characteristic of the image. All of the strata thus stacked forming a treatment layer.

All the outputs of a processing layer make it possible to reconstruct an intermediate image, which will serve as the basis for the next processing layer.

Learning rules vary the efficiency of signal transmission from one neuron to another, we speak of "synaptic weight", said weights being modulated.

In the context of an adversary learning method according to the method, the network 80 admits two inputs 81: "source" data (simulated 2D images) and "target" data (real 2D images).

The source data are associated with the ground truth (the labels: position of the indications in the image), while the target data have no label.

From the source data, in a step called “feed forward” going from the first layer to the last processing layer, a first set of layers 82 extracts a signature 82a from the simulated 2D images presented to the neural network.

A signature is a set of numerical values, which are conveniently defined in the form of a vector. These numerical values are obtained using descriptors, which are for example convolution operations applied to the thumbnail. The descriptors constitute analyzes of the NDG gray levels of the thumbnail.

For each real 2D image, the method thus preferably comprises a plurality of descriptors.

Then, still in the so-called “feed forward” step, a second set of layers 83 allows, from an extracted signature 82a, the prediction of at least one corresponding label 83a.

Then, for each simulated 2D image, the label 82a predicted in the “feed forward” step is compared to the label (position of the indication) associated with said 2D image simulated in the simulated and annotated BDD 36. If these labels differ , the calculation unit 37 determines an error. Statistical learning being carried out by backpropagation of the error made, the calculation unit 37 calculates the derivative of this error to determine how to change the weights of the different layers of neurons so as to reduce the error.

The error is thus back-propagated in the network layer by layer, from the last to the first layer to update the weights of each layer.

The algorithms allowing to learn on deep architectures are complex (and are based on many "hyper parameters" which must be chosen with care so as not to get stuck in poorly performing local minima. The learning algorithm put implemented for backpropagation is for example by an optimization method or stochastic gradient descent.

In a particular embodiment, the neural network 80 can for example be a convolutional neural network.

A convolution is a mathematical operation which makes it possible to modify the initial image by a convolution kernel by filtering it.

It is therefore these convolution kernels that are learned during statistical learning when updating weights.

A convolution kernel will analyze a characteristic of the input image. A non-linear operation is applied to each step (at each convolution): sigmoid, re-read, tanh, ... allowing the extraction of visual characteristics (side, angle, end detection ...).

The convolution kernel can be for example a matrix of size 3x3.

To analyze several characteristics, the neural network 80 stacks strata of independent convolution nuclei, each stratum analyzing a characteristic of the image. All of the strata thus stacked form the convolutional treatment layer. The number of treatment layers defines the depth of the convolution layer.

A pooling layer allows a step of sub-sampling the image, thus reducing the quantity of parameters and calculation, allowing large gains in computing power. The pooling layer also allows a multi-scale approach, the next convolution layer processing a larger part of the initial image.

It is therefore frequent to periodically insert a pooling layer between two successive convolutional layers of an architecture of deep statistical methods.

Following the sequences of convolutional and pooling layers, the architecture comprises at least one fully connected layer FC, which is a layer of the perceptron type. The neurons in a fully connected layer have connections to all of the outputs of the previous layer (as seen regularly in regular neural networks). Thus, this layer makes it possible to classify the signature of an image and therefore to assign at least one label to the image of said signature.

Referring again to FIG. 8, adversary learning is enabled by another branch of the network, comprising a third set of processing layers 84.

This branch of the network is learned to try to distinguish between a source image and a target image. Thus, it will classify the image in a domain 84a. A first domain corresponding to real images, and a second domain corresponding to simulated images.

The weights of the neurons of the network branch 84 are modulated so that the latter is unable to distinguish between the simulated images and the real images, with the aim that the network 80 gives the same detection performance on the source data and target data, so on simulated 2D images and real 2D images.

To achieve this goal, the learning algorithm will therefore minimize a cost function of classification of the signature allowing the image to be assigned at least one label and maximize a cost function of classification of the signature of the image. in a domain.

Consequently, this means that the learning algorithm somewhat penalizes the main task (prediction of labels) in order to impose that the “descriptor extraction” layers 82 have the same behavior for target images and source images. The “descriptor extraction” layers 82 will therefore be invariant to the image domain.

In a particular embodiment where the learning algorithm implemented for backpropagation is the stochastic gradient descent method.

In the network branch 84, the output gradient will be calculated, inverted and injected into the “descriptor extraction” layers 82.

Opponent learning is complex to master since we simultaneously learn the weights of 2 branches (83 and 84) which do not have the same objectives, or even contradictory objectives. To carry out adversary learning, the learning algorithm preferably imposes a very weak constraint for the first iterations then accentuate this constraint as the learning progresses.

This substep E14 of statistical learning can conclude step E10, the output of which is a statistical model MS.

Performance validation

To validate the statistical model MS obtained in step E10 of statistical learning, it is necessary to measure the performance in detection and in false alarms on real data.

Thus, in a sub-step E21 of the step E20 of validation of the statistical model, from test pieces P2, the calculation unit 30 proceeds as in the sub-step EU, to acquire 2D image by X-rays. Also, as in sub-step E12, in a sub-step E22 of step E20, the calculation unit 30 performs pretreatments on the previously acquired images.

Then, in a sub-step E24, taking as input the statistical model determined by step E10 and the preprocessed data coming from E22. The model is therefore applied to this new image and the process obtains as an output the probabilities of the presence of indications in the form of labels.

In a sub-step E23 of E20, the 2D images acquired in E21 are manually annotated by an operator, according to a visual inspection, so as to assign at least one label to the test piece P2.

Then, in a sub-step E25, the calculating unit 37 compares the label predictions of the automatic algorithms to the manual annotations, labels assigned by the operators. Correlation indicators (for example a detection percentage, a number of false alarms, a confusion matrix, etc.) are therefore calculated to see if the correlation and the performances are in line with expectations.

Consequently, advantageously, the step E20 of validation of the performances, is a step whose cost is considerably reduced compared to an approach without simulation: in fact, the number of images to be annotated is low and will only be used for validate the performance of algorithms (statistical learning essentially based on simulation data).

In the event that performance is insufficient (measured by the indicators), the proposed method returns to the previous step E10 so as to refine the statistical model MS.

Online prediction E2

In a second phase E2, the calculating unit 37 classifies online all of the parts to be inspected P3 from a production line.

Thus, in a step E31 of E2, on the basis of the parts to be checked, the calculation unit 30 proceeds as in the sub-step EU, to acquire 2D image by X-rays.

Also, as in the sub-step E12, in a step E32 of E2, the calculation unit 30 carries out preprocessings on the previously acquired images.

Thus, from 2D images of the part to be checked, the probabilities of the presence of indications are calculated in a step E33 using the statistical model MS resulting from the phase E1.

As a result, areas that differ too much from a healthy room are brought up in the room control report. A map of the indications present in the image of a part to be checked is therefore generated. If necessary, a visual inspection, called consolidation, is carried out by an operator to validate the detection.

In addition, following this visual inspection, the data generated by the process can complete the database 36, so that BDD 39 is enriched as the parts are analyzed. Preferably, these data only complete the database if a difference between the result obtained and the consolidation is observed. Thus, the part to be tested becomes a P2 test part.

The embodiments described are implemented in the context of metal parts associated with a non-destructive X-ray control means. However, this method can be deployed according to other embodiments, for example with control means such as by :

- ultrasound: the process described above can similarly simulate this control means and generate parts with simulated indications; and

- digital tomography: digital tomography is based on digital radiography (RX acquisition of multiple 360 ° views around the room with reconstruction of the 3D image of the interior of the room).

References [1] K. Dabov, A. Foi, V. Katkovnik Image and video denoising by sparse 3D transform-domain collaborative fiitering.

[2] Nobuyuki Otsu, A threshold selection method from gray-level histograms ”, IEEE Trans. Sys., Man., Cyber., Vol. 9, 1979, p. 62-66 (DOI 10.1109 / TSMC.1979.4310076).

[3] M. Sezgin and B. Sankur, Survey over image thresholding techniques and quantitative performance evaluation, Journal of Electronic Imaging, vol. 13, n ° l, 2003, p. 146-165 (DOI 10.1117 / 1.1631315) [4] Carlo Tomasi and Roberto Manduchi, Bilateral fiitering for gray and coior images in Computer Vision, 1998. Sixth International Conférence on. IEEE, 1998, pp. 839-846.

[5] Pietro Perona and Jitendra Malik (November 1987). Scaie-space and edge detection using anisotropic diffusion. Proceedings of IEEE Computer Society Workshop on Computer Vision, pp. 16-22.

[6] Rudin, L. I .; Osher, S .; Fatemi, E. (1992). Nonlinear total variation based noise removai aigorithms. Physica D 60: 259-268.

Claims (11)

claims 1. A method of non-destructive testing of a metal part (10) for aeronautics implementing a radiographic image acquisition system (20), comprising a step of statistical learning of indications (E10) comprising the sub -following steps : - (EU) acquire a plurality of real 2D images of a real part (PI) likely to contain potential indications; - (E12) from a reference 3D model of the metal part (10), modifying said 3D model to generate at least one indication given in a given area of said model; - simulate the operating conditions for acquiring the plurality of real 2D images and generate at least one simulated 2D image from said modified 3D model; - adding the simulated 2D image to a database of simulated images (39) of the metal part (10), as well as at least one element for identifying the indication given in said database (39) ; and - (E14) from said database and the plurality of real 2D images, implementing a statistical analysis by supervised adversary learning to generate a statistical model (MS) of indications in a metal part (10); a step of calculating the probabilities of the presence of indications (E2) in the metal part to be checked (10) comprising the following sub-steps: - (E31) acquiring at least a 2D image of said metallic part (10); - (E33) calculating probabilities of the presence of indications in said 2D image from the statistical model (MS); and determine if the metal part (10) is characteristic of a healthy part. 2. Method of non-destructive testing of a metal part (10) for aeronautics according to the preceding claim wherein the element for identifying a given indication in said database is a binary mask of said given indication. 3. A method of non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims in which the statistical learning step (E10) also comprises the sub-step consisting in: estimating parameters of the 3D fitting of said test piece (PI) during the acquisition of said plurality of real 2D images; and in which the substep of simulating the operating conditions for acquiring the plurality of real 2D images is carried out from said estimated parameters of the 3D fitting of said test piece (PI) in order to determine registration parameters d '' a 3D model of the metal part (10). 4. A method of non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims in which after the statistical learning step (E10), the method comprises a step of validating the statistical model ( E20) comprising the following sub-steps: - (E21) acquire at least a 2D image of a reference part (P2), - (E23) assign to said 2D image at least one label relating to the presence of indications as a function of a visual inspection of the 2D image, - (E24) calculate the probabilities of the presence of indications in the 2D image from the statistical model (MS) and determine at least one label, - compare the determined label (s) with the assigned label (s), and calculate a correlation indicator, - if the indicator is less than a given value, continue the statistical learning step (E10), - otherwise validate the statistical model (MS). 5. Method for non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims, in which after the acquisition of at least a 2D image of a test part (PI), and / or a reference piece (P2), and / or a reference piece (P3), said acquired 2D image undergoes processing (E13 / E22 / E32) by processing means adapted to attenuate artefacts and / or variability of the 2D image. 6. A method of non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims wherein after generation of a simulated 2D image from said modified 3D model, said acquired 2D image is processed by processing means adapted to attenuate artifacts and / or variabilities of the 2D image. 7. A method of non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims wherein the statistical analysis is obtained by deep learning based on neural networks. 8. A method of non-destructive testing of a metal part (10) for aeronautics according to the preceding claim wherein the deep learning based on neural networks is configured to perform an adversary learning algorithm. 9. A method of non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims in which the statistical analysis by supervised learning comprises: for at least one simulated image and at least one real image of the plurality of real 2D images, the automatic extraction of a signature from said images, determining a plurality of p descriptors, a descriptor being the result of an analysis d 'picture ; the automatic classification of the signature of the simulated image making it possible to attribute to said image a label representative of the presence of indications in the part; the back-propagation of the error of said classification of the label to minimize a cost function of the classification of the signature making it possible to attribute a label to the image; the automatic classification of the signature of the real or simulated image in a domain of the simulated image, or of the real image; and - backpropagation of the inverse of the error of said classification of the domain to maximize a cost function of the classification of the signature of the thumbnail in a domain; so that the extraction of a signature is invariant depending on the domain. 10. A method of non-destructive testing of a metal part (10) for aeronautics according to one of the preceding claims in which the learning algorithm used for backpropagation is the stochastic gradient descent method. 11. computer program product, intended to be executed by processing means (31) of a calculation unit (30), said calculation unit further comprising a memory (32), and configured to implement process steps according to any one of the preceding claims.