Method for non-destructive testing of parts made of composite material

Abstract

The invention relates to a non-destructive testing method for a woven composite material part, the fault volume of which has been acquired in advance and divided into a plurality of visual areas, characterized in that the method comprises the following steps: dividing the visual areas of the fault volume into three-dimensional sub-volumes; extracting a descriptor vector, and projecting the extracted descriptor vector so as to compress the descriptor vector into a dimensionality reduction descriptor vector; converting the dimensionality reduction descriptor vector into scalar values representing the likelihood that the sub-volume belongs to a health category, aggregating the scalar values at each visualization region, and estimating the membership of the health category; and detecting the existence of defects in the woven composite part according to the obtained membership estimated value of the health category of the visual region of the fault volume.

Description

Technical Field

[0001] The present invention relates to nondestructive testing of parts made of composite materials ( The present invention relates to the field of non-destructive testing (NDT), in particular the non-destructive testing of braided composite blades in the aeronautical field. More specifically, the present invention relates to the use of tomographic images of parts to inspect such parts. Background Art

[0002] Composite materials are commonly used for their high strength and light weight. Typically, such composite materials consist of reinforcements (such as carbon fiber reinforcements) and a resin matrix into which the reinforcements are embedded. Especially in the aerospace sector, safety and reliability are major concerns, so manufactured parts must undergo various types of analysis and inspection operations to verify their correct performance and defect-free nature.

[0003] In fact, depending on the location of the defect in the material and especially its proximity or nature (fiber volume fraction variation, warp / weft ratio variation), detecting the presence of a defect may call into question the integrity and reliability of the part.

[0004] A common nondestructive testing (CND) method involves acquiring cross-sectional images of the part to be inspected and analyzing them by experts. These experts typically examine multiple dimensional cross-sections of the part, selected to scan the part in three dimensions.

[0005] Understandably, this type of "manual" inspection is unsatisfactory. Due to the large size of the parts, it is tedious and can take many hours. Furthermore, this type of inspection is closely tied to operator expertise: variability within the same operator or between operators can negatively impact inspection quality.

[0006] Therefore, it is necessary to make the quality of nondestructive testing more reliable by avoiding these variabilities as much as possible to improve repeatability and increase yield.

[0007] Document FR 3050826 A1 proposes implementing a semi-supervised training technique based on a method called the “mean teacher” method, in order to free operators from the “manual” inspection step of parts made of composite materials. However, the use of this training technique is not satisfactory due to the small amount of data annotated by qualified operators and the insufficient quality of the available data. In addition, the technique proposed in this document only performs a classification based on a binary type, which assigns a “non-defective” (“healthy”) value or a “defective” (“abnormal”) value. However, this classification does not allow reliable detection that minimizes false alarms (i.e. healthy areas are detected as defective). In fact, some defective areas may be classified as “non-defective”, or many “non-defective” areas may be classified as having “defects”. In addition, no reliable results can be obtained in terms of detecting new or previously unobserved and / or recorded weaving defects.

[0008] Other training techniques are also known. In particular, unsupervised training methods are able to classify unlabeled data by identifying similarities. However, the quality of available images and the high similarity between the properties of healthy and defective materials make these techniques difficult to implement satisfactorily, particularly in the field of aviation and braided composite blades. Summary of the Invention

[0009] The general purpose of the present invention is to propose a method for detecting a set of defects or possible defects in a woven fabric and with a limited amount of data in a fast and robust manner.

[0010] To this end, according to one aspect of the present invention, a method for nondestructive testing of a woven composite part is provided, wherein a tomographic volume of the woven composite part has been previously acquired and divided into a plurality of visualization regions, characterized in that the method comprises the following steps:

[0011] - pre-processing the tomographic volume, the pre-processing comprising: dividing the visualization area of ​​the tomographic volume into three-dimensional sub-volumes;

[0012] - processing each subvolume, the processing comprising: for each subvolume, on the one hand, extracting a descriptor vector, the extraction being performed by performing image analysis using a pre-trained representation model, the image analysis decomposing each subvolume into weights of a plurality of descriptors applied to the representation model, the descriptor vector being composed of the decomposed weights; and on the other hand, projecting the extracted descriptor vector into a pre-trained dimensionality reduction model in order to compress the descriptor vector into a reduced dimensionality descriptor vector having a lower dimensionality;

[0013] - post-processing the reduced-dimensionality descriptor vectors, the post-processing comprising: on the one hand, converting the reduced-dimensionality descriptor vectors of each subvolume into a scalar value representing the likelihood that the subvolume belongs to the healthy class using a pre-trained likelihood estimation model; and on the other hand, aggregating the scalar values ​​by calculating the average of the scalar values ​​in each visualization region, the result of which is a membership estimate of the healthy class for each visualization region of the tomographic volume; and

[0014] - Checking the presence of defects in woven composite parts based on membership estimates of healthy classes of the obtained visualization areas of the tomographic volume.

[0015] The method combines, among other things, the creation of a representation model and a likelihood estimation model. This combination enables the detection and localization of a set of anomalies, including those that have never been observed or are not available in the training base.

[0016] Furthermore, the method provides more robust localization results than existing techniques by reducing the number of "false positives" (i.e., areas incorrectly classified as healthy).

[0017] Furthermore, the method improves detection by simplifying it and making it faster.

[0018] Advantageously but optionally, the described method comprises at least one of the following features, used alone or in combination:

[0019] - A step of creating a representation model, allowing the extraction of descriptor vectors, comprising the steps of: selecting reference subvolumes, the reference subvolumes forming a reference tomographic volume, annotated as either healthy or defective; performing a diversification process on the reference subvolumes, said diversification process implementing a multiplication or geometric transformation of one or more reference subvolumes; and training a neural network using the diversified reference subvolumes to obtain a representation model of the reference subvolumes, the training implementing an error function and an optimization algorithm. This makes it possible, in particular, to increase the representation of subvolumes containing defects and to perform non-destructive testing even in the absence of annotated subvolumes.

[0020] - a step of creating a dimensionality reduction model capable of projecting the extracted descriptor vectors, the step of creating the dimensionality reduction model comprising: extracting a reference descriptor vector from the reference subvolume using the representation model; and performing a computational construction of the dimensionality reduction space, said computational construction performing an orthogonalization of the first covariance matrix of the extracted reference descriptor vectors. This makes it possible, in particular, to compress the descriptor vectors, thereby minimizing the effect known as the "curse of dimensionality".

[0021] - A step of creating a likelihood estimation model capable of transforming the reduced-dimensionality descriptor vector, said creation implementing the steps of: projecting the extracted reference descriptor vector labeled as healthy into a reduced-dimensionality space, so as to compress the extracted reference descriptor vector labeled as healthy into a reduced-dimensionality reference descriptor vector of the healthy type; and calculating a second covariance matrix of the reduced-dimensionality reference descriptor vector of the healthy type, the second covariance matrix representing the domain of the healthy class. This, in particular, enables a correct measurement of the distance between the scalar value and the domain of the healthy class, thereby measuring normality.

[0022] - Projecting the descriptor vectors into the reduced-dimensional descriptor vectors Implementing a principal component analysis method This makes it possible, among other things, to directly project each new descriptor vector into the reduced-dimensional space without the need for additional steps.

[0023] - The reduced-dimensional descriptor vector of each subvolume is converted into a scalar value and the Mahalanobis distance is calculated. This makes it possible to minimize the influence of the noisiest components.

[0024] - Training the neural network uses a semi-supervised method called "average teacher" using two neural networks. The weights of one neural network are defined as the exponential moving average of the weights of the other neural network. This makes it possible to obtain descriptor vectors of dimension 2048 instead of binary responses.

[0025] -Neural network training uses a "focal loss" or cross-entropy error function. This makes it possible to penalize poor predictions, especially for imbalanced datasets.

[0026] - Creating the representation model includes grouping the reference subvolumes obtained by the diversification process into subsets, each subset including a reference subvolume of the healthy type and a reference subvolume of the defective type. This makes it possible to ensure that the network does not overfit to the subvolumes of the healthy type.

[0027] - Creating the representation model involves randomly sampling reference subvolumes that are selected and labeled as healthy and evenly distributed in the tomographic volume. This makes it possible, in particular, to rebalance the selected subvolumes according to their type.

[0028] The pre-processing comprises an initial correction step which implements an adjustment of the format and the compressibility of the slice volume.

[0029] - The pre-processing includes a step of flattening the slice volume in order to process a volume without curvature, this curvature being independent of the local direction of the braid.

[0030] - Preprocessing includes a step of global normalization of the gray levels of the adjusted tomographic volume in order to increase the contrast of the tomographic volume.

[0031] The division of the visualization area consists of tiling into sub-volumes of size 100×100×D voxels, D being the dimension of the thickness of the slice volume. This makes it possible, in particular, to improve the statistical processing performed subsequently.

[0032] The visualization area is divided so that only 50% overlap is achieved between each pair of adjacent sub-volumes of the same visualization area in order to minimize the leakage rate. This makes it possible to eliminate bias in particular when a defect is located between two sub-volumes.

[0033] - Pre-processing includes a local contrast correction step implementing a histogram normalization of each divided sub-volume.

[0034] The local contrast correction step is implemented by means of a CLAHE (Contrast Limited Adaptive Histogram Equalization) type algorithm. This makes it possible in particular to increase the local contrast while limiting noise amplification.

[0035] - Checking for the presence of defects in woven composite parts Implementing the comparison of membership estimates of healthy classes with predefined thresholds.

[0036] According to another aspect, a computer program product is provided, comprising code instructions for performing a non-destructive testing method when the code instructions are executed on a processing device of a computing unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Other features, objects and advantages will become apparent from the following description, which is intended to be illustrative and non-limiting and which needs to be read in conjunction with the accompanying drawings, in which:

[0038] Figure 1a 、 Figure 1b and Figure 1c Examples of weaving defects are shown.

[0039] Figure 2 The steps of a method for non-destructive testing of a braided composite material part according to an embodiment of the present invention are shown.

[0040] Figure 3 An example of a view of a woven composite part is shown with visualization areas indicated on the view.

[0041] Figure 4 The division of the tomographic volume visualization area into sub-volumes according to an embodiment of the present invention is shown.

[0042] Figure 5 The steps of training a detection model according to an embodiment of the present invention are shown.

[0043] Figure 6 The steps of a method for analyzing the quality of a part are schematically shown.

[0044] Similar elements have the same reference numerals throughout the drawings. DETAILED DESCRIPTION

[0045] General Information

[0046] The implemented non-destructive testing (CND) method can be applied to different manufacturing steps of the braided composite part 1 to be tested (for example a fan blade or turbine blade). Depending on the part, the composite material can be different (carbon matrix resin, ceramic matrix resin, etc.).

[0047] The purpose of nondestructive testing of part 1 is to detect weaving or injection defects d in the part, which usually includes:

[0048] Looping: The carbon warp or weft strands do not follow their theoretical trajectory and form loops within the part ( Figure 1a );

[0049] Missing: Carbon warp or weft strands are missing ( Figure 1b );

[0050] flambage, with or without resin buildup ( Figure 1c ).

[0051] Nondestructive testing methods

[0052] Figure 2 The steps of a method for non-destructive testing of a braided composite material part 1 are shown. A three-dimensional image (tomographic volume) of the part is previously obtained by reconstruction from a plurality of two-dimensional cross sections of the part, which were acquired by X-ray radiography.

[0053] For example Figure 3 As schematically shown in the figure, the tomographic volume 2 obtained from the woven composite part 1 is pre-divided into multiple visualization areas 3 (volumes), typically dozens of visualization areas 3, for example 60 visualization areas 3, each area having a size of 250×250×D voxels, where D represents the thickness of the part represented by the tomographic volume 2 (expressed in the number of voxels).

[0054] Each visualized region 3 of the woven part is first subjected to a pre-processing E1 in order to be divided into three-dimensional sub-volumes 4 .

[0055] The sub-volumes 4 then undergo a process E2 which makes it possible to obtain a reduced-dimensionality descriptor vector 6 for each sub-volume 4 .

[0056] These reduced-dimensionality descriptor vectors 6 are themselves subjected to a post-processing E3 in order to obtain a membership estimate 8 to the healthy class for each visualization region 3 of the tomographic volume 2 .

[0057] Finally, the method comprises a detection E4 taking into account the estimated membership 8 of the healthy class for each visualization area 3 .

[0058] These different steps E1 to E4 are described in detail below.

[0059] The preprocessing step E1 itself comprises several sub-steps. First, the slice volume 2 is corrected (step E10) so that it can be read. Correction E10 comprises adjusting the format and compression of the slice volume 2. This step can be followed by a three-dimensional flattening E11 of the slice volume 2. Flattening removes the curvature of the slice volume 2 that is not related to the weave itself. In other words, the slice volume 2 is projected three-dimensionally onto a plane in order to remove the curvature that is not related to the local direction of the weave, so that this curvature is not considered a weave defect d, but rather a fixed and sought-after constraint in the design. This step allows only the curvature related to the local direction of the weave to be retained.

[0060] Furthermore, preprocessing E1 includes a global grayscale normalization E12 of the slice volume 2. This normalization E12 implements a homogenization to reduce acquisition artifacts, so that the grayscale of each voxel in the slice volume 2 is proportional to the material density of the part at that point. Normalization E12 removes extreme values ​​from the grayscale distribution to improve the overall contrast of the slice volume 2. The extreme values ​​are defined, for example, by the 5% percentiles at each end of the distribution (in other words, the 5% of the highest values ​​and the 5% of the lowest values). Normalization E12 thus makes the subsequent steps more reliable.

[0061] like Figure 4 As shown, during the preprocessing E1, the visualization areas of the tomographic volume 2 are divided (step E13). The division E13 performs the following tiling, i.e. tiling each visualization area 3 of the tomographic volume 2 into three-dimensional elementary sub-volumes 4. The sub-volumes 4 have, for example, a size of 100×100×D voxels, so as to correspond to the representative dimensions of the weaving pattern (the elementary unit of the weaving from a material / mechanical point of view). To this end, each sub-volume 4 must meet two conditions: on the one hand, the sub-volume 4 must be larger than the characteristic dimensions of the constituent parts (here, for example, the strands) so that their number is sufficient to allow a correct statistical representation of the volume under investigation; on the other hand, the sub-volume 4 must be smaller than the characteristic dimensions of the visualization areas 3 in order to be able to apply statistical processing in each of these visualization areas 3. Furthermore, the determination of the sub-volumes 4 depends on the weaving pattern used in the composite part. Preferably, the sub-volumes 4 are oriented according to the orientation of the part 1.

[0062] Partitioning step E13 performs an overlap only between pairs of adjacent subvolumes 4 within the same visualization region 3 to minimize leakage. The overlap between each pair of adjacent subvolumes 4 can range from 20% to 80%, preferably 50%. This 50% overlap ensures that every voxel in the tomographic volume 2 is contained within both subvolumes 4. The purpose of this overlap is to eliminate errors in the event of a defect d between two subvolumes 4. Furthermore, this partitioning E13 is limited to overlaps between subvolumes 4 "within the visualization region 3." Therefore, partitioning E13 does not perform an overlap of subvolumes 4 from different visualization regions 3. This ensures that, throughout the remainder of the method, the reduced-dimensionality descriptor vector 6 for each subvolume 4 is associated with the portion of the tomographic volume 2 that is completely contained within a single visualization region 3. Consequently, the healthy class membership estimates 8 for each visualization region 3 can be aggregated by visualization region 3 using simple statistical operations, without the need to develop algorithms specifically for the distribution of membership estimates 8.

[0063] After the division, step E14 is performed: a local correction of the contrast of each subvolume 4 is performed. Local correction E14 implements a histogram normalization algorithm of the "Contrast Limited Adaptive Histogram Equalization" (CLAHE) type. Furthermore, local correction E14 performs a simple "whitening" normalization of the subvolume 4, giving it zero mean and unit variance. This local correction E14 is more robust than a global correction, thus increasing the reliability of subsequent processing. Furthermore, the CLAHE method limits the amplification of noise generated by the histogram normalization.

[0064] Next comes the processing step E2, which is performed on each pre-processed subvolume 4 and comprises a number of consecutive steps. First, a processing step is performed to extract E21 a vector for each subvolume 4, which is referred to as a descriptor vector 5. Extraction E21 is performed by image analysis using a pre-trained representation model 9. The representation model 9 is obtained through training, one embodiment of which is described in detail below. The image analysis decomposes each subvolume 4 into the weights of a plurality of descriptors applied to the representation model. These descriptor weights thus form a descriptor vector 5.

[0065] The descriptor vectors 5 extracted from each subvolume 4 by extraction E21 are then subjected to projection E22. This step implements a pre-trained dimensionality reduction model 10, which makes it possible to compress each descriptor vector 5 extracted from the subvolume 4 into a reduced-dimensionality descriptor vector 6. To this end, projection E22 implements a parameterized method, such as principal component analysis. This analysis makes it possible to obtain the main dimensions of each descriptor vector 5 by simply projecting into the reduced-dimensionality space obtained using the dimensionality reduction model 10 (without the need for additional steps). The main dimensions of the descriptor vectors 5 synthesize important information and form a reduced-dimensionality descriptor vector 6. The dimensionality of the reduced-dimensionality descriptor vector 6 is therefore lower than that of the extracted descriptor vector 5. Projection E22 makes it possible to minimize the significant increase in the amount of data (also known as the "curse of dimensionality" effect), which would isolate and disperse the data, rendering the method ineffective.

[0066] The processing E2 is followed by a post-processing step E3 of the reduced descriptor vectors 6. The post-processing E3 first consists in converting E31 each reduced descriptor vector 6 into a scalar value 7. The scalar value 7 represents the likelihood of membership of the subvolume 4 from which the reduced descriptor vector 6 is extracted to the healthy class. The conversion E31 implements the covariance matrix 12 (called the health matrix) obtained by means of a pre-trained likelihood estimation model 11. The conversion E31 is achieved by calculating the Mahalanobis distance taking into account the health covariance matrix 12. The Mahalanobis distance gives a smaller weight to the components with the greatest degree of dispersion, which makes it possible to minimize the influence of the components with the greatest noise (components with the greatest variance). Optionally, the calculated distance is normalized to obtain a distance within a fixed range (for example between 0 and 1).

[0067] Post-processing E3 then comprises aggregating E32 the scalar values ​​7 per visualization region 3. Aggregation E32 performs a mean calculation of the scalar values ​​7 in each visualization region 3. This calculation results in an estimated membership value 8 of the health class for each visualization region 3 of the slice volume 2.

[0068] This post-processing E3 enables a detection E4 of the possible presence of defects d on the woven composite part 1 whose tomographic volume 2 has been analyzed. This detection E4 considers the estimated membership values ​​8 of the healthy category of the visualization areas 3 of the tomographic volume 2, obtained by post-processing E3, to determine whether one of the visualization areas 3 contains a defect d. This detection step E4 can be performed in various ways, for example by a qualified inspector. In this step, the inspector checks for visualization areas 3 in the tomographic volume 2 whose estimated membership values ​​8 of the healthy category are below a fixed threshold. This detection verifies the valid presence of defects d in these visualization areas 3, and only in these visualization areas, thus saving time.

[0069] Thanks to the pre-trained detection model, the implementation of the steps of this non-destructive testing method is automated.

[0070] Detection model training

[0071] Figure 5 The steps of a possible implementation of the training of the detection model are shown. The detection model uses three sub-models: a representation model 9 implemented in the extraction E21 described above, a dimensionality reduction model 10 implemented in the projection E22 described above, and a likelihood estimation model 11 implemented in the transformation E31 described above.

[0072] To perform this training, the computing unit is provided with a dataset A0 comprising one or more slice volumes 2r, referred to as reference volumes. The reference slice volumes 2r are preprocessed (step E1) in the same manner as the slice volume 2 during the nondestructive testing method to obtain reference subvolumes 4r. The dataset thus provided (step A0) and preprocessed (step E1), i.e., the collection of reference slice subvolumes 4r, constitutes the database for creating the A1 representation model 9.

[0073] Creation A1 of the representation model 9 implements multiple steps. This creation A1 is based on a training technique known as supervised training. The representation model 9 is capable of extracting reference descriptor vectors 5r for the reference subvolume 4r. Creation A1 first implements selection A10 of the reference subvolume 4r. The reference subvolumes 4r are annotated by a qualified inspector who indicates the location of the area corresponding to the defect d. The inspector assigns a "no defect" ("healthy") value or a "defect" ("abnormal") value to each reference subvolume 4r. Selection A10 is followed by sorting (tri) A11, which retains all reference subvolumes 4r of unhealthy type and randomly samples reference subvolumes 4r of healthy type so that these reference subvolumes are evenly distributed in the visualization area 3 of the reference tomographic volume 2r. This sorting A11 can improve the balance of the categories in the database. According to one embodiment, the ratio between the reference subvolumes 4r of healthy type and the reference subvolumes of unhealthy type in the database is set to 3:1.

[0074] The creation A1 of the representation model 9 also includes a diversification step A12 of the selected reference subvolumes 4r to increase the diversity of the unhealthy reference subvolumes 4r. The diversification step A12 applies geometric transformations to the multiple reference subvolumes 4r, thereby multiplying these reference subvolumes 4r. The applied geometric transformations may include, for example, a mirror effect, random rotation, affine deformation, or random segmentation. Only the mirror effect is applied to all healthy reference subvolumes 4r, while all geometric transformations are applied to the unhealthy reference subvolumes 4r. These transformations can artificially create diversity in the reference subvolumes 4r and further increase the representativeness and diversity of the unhealthy reference subvolumes 4r. This step, in particular, makes it possible to compensate for a small number of annotated reference subvolumes 4r.

[0075] Following the diversification process A12, a step A13 of grouping the multiplied different reference subvolumes 4r is performed by grouping the healthy and unhealthy reference subvolumes 4r into a plurality of subsets (or "batches"). Thus, the healthy subvolumes 4 are inevitably grouped into subsets containing the unhealthy reference subvolumes 4r. The ratio of healthy to unhealthy reference subvolumes 4r in the different subsets is preferably kept constant. Consequently, the representation model 9 is not distorted by local overrepresentation of the healthy subvolumes 4.

[0076] These steps are followed by training the neural network A14. The neural network receives as input the reference subvolumes 4r, grouped into multiple subsets, and is trained to provide a representation model 9. The training A14 is preferably deep training. Training A14 employs a semi-supervised technique known as the "average teacher" method (Tarvainen, 2017) (Wang, 2021). This technique ensures effective training even in the absence of reference subvolumes 4r. This method uses two networks, referred to as a "teacher" and a "student." The weights of the "teacher" network are defined as the exponential moving average (EMA) of the weights of the "student" network. That is, after each "training" step, the parameters of the "teacher" network are slightly updated based on the parameters of the "student" network. The "student" network is optimized according to a chosen focal error function combined with a consistency loss function that penalizes any deviations in the responses of the "student" network from those of the "teacher." The "average teacher" method is particularly effective in promoting the preservation and refinement of information throughout the network. Once training is complete, the weights of the "student" network are fixed for use in testing (or inference) mode in non-destructive testing methods. The training A14 implements a focal error function and an optimization algorithm to improve its performance.

[0077] According to one embodiment, the neural network adopts the following types of architecture: ResNet 10 (He), ResNet 18 (He), ResNet 50, DenseNet 121 or ViT, preferably ResNet 50.

[0078] In an alternative embodiment, training A14 implements a supervision technique.

[0079] Training A14 is combined with an optimization algorithm that tends to converge slowly to promote information distillation and stabilization of the neural network parameters. In the method proposed here, instead of retaining all the weights of the neural network, the last classification layer is removed, so that the model output corresponds to the output of the second-to-last layer (or "average pooling"), which serves as a descriptor extractor.

[0080] The implemented focal error function is preferably a “focal loss” type function (Lin., 2017) or a cross entropy type function, which are particularly suitable when the dataset is imbalanced. The current formula shows an example:

[0081] focal loss(x,y)=-α c (1-p c ) γ log(p c )·1{y≠ignore label},

[0082] in,

[0083]

[0084] where p corresponds to the output of the neural network, α corresponds to a weight coefficient fixed at 0.25, and γ corresponds to a penalty factor (or modulation factor) that depends on the difficulty of a prediction belonging to one of these categories.

[0085] The optimization algorithm can be, for example, the Adam algorithm (Kingma, 2014), SGD, or RMSProp.

[0086] Thus, instead of having a model providing a binary response for each given reference subvolume 4, it provides a descriptor vector 5 having, for example, a length of 2048. The training A14 of the neural network makes it possible to obtain a model as representation model 9 implemented in the non-destructive testing method.

[0087] Training the detection model then involves the creation (A2) of a dimensionality reduction model 10. This creation (A2) allows for the projection (E22) of the descriptor vectors 5 to reduce the dimensionality of the descriptor vectors 5 and retain only the information most relevant to the detection of the defect d. The creation (A2) of the dimensionality reduction model 10 first involves extracting (E21) reference descriptor vectors 5r from the reference subvolume 4r, regardless of their labeling. This creation (A2) also includes the implementation of a reduced-dimensionality space construction calculation (A21). This calculation (A21) performs an orthogonalization process on the covariance matrix of the reference descriptor vectors 5r extracted using the principal component analysis method described above. Thus, only a certain number of eigenvectors are selected from the total number of eigenvectors generated by the orthogonalization process to reduce the dimensionality of these extracted reference descriptor vectors 5r. For example, only the first 50 of the 2048 eigenvectors of each extracted reference descriptor vector 5r may be retained, as these eigenvectors represent 99% of the total variance. The dimensionality reduction model 10 thus enables the creation of a reduced-dimensionality space for compressing the extracted descriptor vectors 5.

[0088] Finally, a likelihood estimation model 11 is created A3. This model is adapted to allow the transformation E31 of the reduced-dimensionality descriptor vectors 6 of each subvolume 4 into scalar values ​​7. To this end, this creation A3 implements the previously described projection E22 of the extracted reference descriptor vectors 5r of the labeled healthy type in the reduced-dimensionality model 10, in order to obtain reduced-dimensionality reference descriptor vectors 6r of the healthy type. This creation then performs a calculation A31 of the health covariance matrix 12 of these reduced-dimensionality reference descriptor vectors 6r of the healthy type. This health covariance matrix 12 is the likelihood estimation model 11 that allows the transformation E31. In other words, the health covariance matrix 12 represents the domain of the healthy class subvolume 4.

[0089] Method evaluation

[0090] In order to be able to evaluate the overall behavior of the method in various scenarios, the metric of the area under the curve (or AUC) of the receiver operating characteristic (or ROC) curve was used. The ROC curve is a tool commonly used in binary detection situations and is able to characterize how the good detection rate varies with the false alarm rate. This measure allows to illustrate the predictive power of the method, as it highlights its quality in distinguishing abnormal samples.

[0091] For a perfect estimator, the AUC of the ROC curve will be close to 1, while for a "random" estimator, the AUC of the ROC curve will be 0.5.

[0092] Furthermore, the ROC curve and its AUC can directly illustrate industry standards, such as minimizing the number of false alarms while maximizing the correct location of the indication. In other words, the goal is to avoid presenting the inspector with only healthy areas to inspect, but rather present the inspector with a complete set of problematic areas.

[0093] Therefore, the proposed method was implemented on a set of 31 fan blades, 23 of which contained weaving anomalies as annotated by qualified inspectors using associated binary masks. These parts were used to evaluate the performance of the model during training.

[0094] Subsequently, a database of 28 parts was created, including 16 parts with anomalies that were also annotated by qualified inspectors. These parts were processed using the proposed method in test mode (or "inference"), and the results were used to plot ROC curves.

[0095] The results obtained demonstrate the exceptional performance of the proposed method, achieving an AUC of 96.06% for identifying defect d. This was achieved by adhering to industrial constraints, including a good detection rate above 80% and a false alarm rate below 30% for multiple operating points of the method. Furthermore, the method achieved a remarkably high AUC of 99.46% for defect d types not present in the database initially provided for training (the first 31 blades). Therefore, these results also demonstrate the proposed method's ability to generalize to new types of defect d.

[0096] Integration into the quality analysis process

[0097] The nondestructive testing method M4 of the braided composite material part 1 can be implemented in a quality analysis process. The quality analysis process includes multiple steps, some of which are Figure 6 Schematically illustrated in FIG. A preliminary acquisition step M1 is performed using an imaging system (e.g., a tomography device). This step enables a three-dimensional image of the component 1 to be obtained, in other words, a tomographic volume 2. This tomographic volume 2 is transmitted to a module that implements the non-destructive testing method as described above.

[0098] The quality analysis process may also include an analysis M2 of the conformity of the slice volume 2 with the shape and dimensions of the manufactured part, and an analysis M3 of the porosity and the presence of foreign matter within the volume of the part. For example, the analysis M3 may be implemented by a macro program to automatically calculate the porosity of each part, as presented, for example, in document FR 3 046 844 A1.

[0099] A computing unit is provided for performing computer processing operations. This computing unit implements the quality analysis process, particularly the steps of acquisition M1, consistency analysis M2, porosity and foreign matter presence analysis M3, and detection method M4. To this end, the computing unit comprises a number of modules. The computing unit stores the detection model described above and implemented in detection method M4. The computing unit includes data processing means and memory. Furthermore, such a unit may include a module adapted to perform the step of displaying the results of the analysis and detection M5.

Claims

1. A method for non-destructive testing of a braided composite material part (1), wherein: A previously acquired and reconstructed tomographic volume (2) is divided into a plurality of visualization regions (3), wherein the method comprises the following steps: - pre-processing (E1) the tomographic volume (2), the pre-processing comprising: dividing (E13) the visualization area (3) of the tomographic volume (2) into three-dimensional sub-volumes (4); - performing a process (E2) on each sub-volume (4), said process comprising: for each sub-volume (4), On the one hand, extracting (E21) a descriptor vector (5), said extraction being carried out by means of an image analysis performed by a pre-trained representation model (9), said image analysis decomposing each subvolume (4) into weights of a plurality of descriptors applied to said representation model (9), said descriptor vector (5) being composed of the weights of this decomposition, and On the other hand, the extracted descriptor vector (5) is projected (E22) into a pre-trained dimensionality reduction model (10) so as to compress the descriptor vector (5) into a reduced dimensionality descriptor vector (6) having a low dimension; - performing post-processing (E3) on the reduced-dimensionality descriptor vector (6), the post-processing comprising: On the one hand, the reduced-dimensional descriptor vector (6) of each subvolume (4) is converted (E31) into a scalar value (7) by a pre-trained likelihood estimation model (11), wherein the scalar value (7) represents the likelihood that the subvolume belongs to the healthy class, and On the other hand, aggregation (E32) is performed by calculating the average value of the scalar values ​​(7) in each of the visualization areas (3), the result of this calculation being an estimated value (8) of membership in the healthy class for each visualization area (3) of the tomographic volume (2); and - detecting (E4) the presence of a defect (d) in the braided composite material part (1) based on the estimated value (8) of membership of the healthy class of the obtained visualization area (3) of the tomographic volume (2).

2. The method according to claim 1, comprising: A step of creating (A1) said representation model (9), said representation model being capable of allowing said extraction (E21) of the descriptor vector (5), and said creation of said representation model comprising the steps of: - selecting (A10) a reference subvolume (4r), said reference subvolume forming a reference slice volume (2r), annotated as a healthy type or a defective type; - performing a diversification process (A12) on the reference subvolumes (4r), the diversification process (A12) implementing a multiplication or geometric transformation of one or more reference subvolumes (4r); and - training (A14) a neural network using the diversified reference subvolumes (4r) to obtain a representation model (9) of the reference subvolume (4r), said training implementing an error function and an optimization algorithm.

3. The method according to claim 2, comprising: A step of creating (A2) said dimensionality reduction model (10), said dimensionality reduction model being able to allow said projection (E22) of the extracted descriptor vector (5), said creating of said dimensionality reduction model comprising the implementation of the following steps: - extracting ( E21 ) a reference descriptor vector ( 5r ) from said reference subvolume ( 4r ) by means of said representation model ( 9 ); and A construction calculation (A21) of a reduced-dimensional space, said construction calculation (A21) implementing an orthogonalization process of the first covariance matrix of the extracted reference descriptor vector (5r).

4. The method according to claim 3, comprising: A step of creating (A3) said likelihood estimation model (11), said likelihood estimation model being capable of allowing the transformation (E31) of the reduced-dimensional descriptor vector (6), said training being performed by: - Projecting (E22) the extracted reference descriptor vector (5r) labeled as healthy type into the reduced dimensionality space in order to compress the extracted reference descriptor vector labeled as healthy type into a reduced dimensionality reference descriptor vector (6r) of healthy type; as well as - Calculating (A31) a second covariance matrix (12) of the reduced-dimensional reference descriptor vector (6r) of the health type, said second covariance matrix representing the domain of the health category.

5. The method according to any one of claims 1 to 4, wherein The descriptor vector (5) is projected (E22) into a reduced-dimensional descriptor vector (6) by performing a principal component analysis (PCA) method.

6. The method according to any one of claims 2 to 4, wherein: Creating (A1) the representation model (9) comprises grouping (A13) the reference subvolumes (4r) obtained by the diversification process (A12) into subsets, each subset comprising a reference subvolume (4r) of a healthy type and a defective type.

7. The method according to any one of claims 1 to 6, wherein The pre-processing (E1) comprises a step of globally normalizing (E12) the grayscale level of the adjusted tomographic volume (3) to improve the contrast.

8. The method according to any one of claims 1 to 7, wherein The division (E13) of the visualization region (3) is such that overlap is achieved only between each pair of adjacent sub-volumes (4) of the same visualization region (3) in order to minimize the leakage rate.

9. The method according to any one of claims 1 to 8, wherein The pre-processing (E1) comprises: a local correction step (E14) of the contrast, which performs a histogram normalization of each divided sub-volume (4).

10. The method according to any one of claims 1 to 9, wherein Detecting (E4) the presence of a defect (d) in the woven composite material part (1) involves comparing the estimated value (8) of membership of the healthy class with a predefined threshold value.

11. A computer program product comprising code instructions for executing the non-destructive testing method according to any one of claims 1 to 10 when the code instructions are executed on a processing device of a computing unit.

Images