A method for inspecting a part during or after its manufacture.

Abstract

The invention relates to a method for inspecting a part based on a set of characteristics during or after its manufacture. During the method, the correlated characteristics above and below a threshold value are determined. Then, the characteristics of the part that are correlated below the threshold are measured, and the characteristics of the part that are correlated above the threshold but have not yet been measured are predicted. This method greatly reduces the number of measurements required to inspect a part. Figure for the abstract: 1

Description

Title of the invention: Method for inspecting a part during or at the end of its manufacture. Technical field of the invention

[0001] The present invention relates to the control of manufacturing conformity of parts.

[0002] It can in particular find application for aircraft engine parts, for aircraft landing gear parts and more broadly in the aeronautical field, although parts manufactured for other transport sectors such as the automotive industry may also be concerned.

[0003] Other fields such as energy, for example turbines for thermal power plants, or for the manufacturing industry, for example machine tool parts, may also be concerned. Technological background

[0004] The manufacturing processes for parts, particularly parts for aircraft engines, are complex. Indeed, the finished part must meet a very large number of specifications. These specifications can be of various kinds, for example, regarding the shape of the part, its thickness in certain areas, etc. Verification of these specifications ensures the conformity of the part to the specifications, the correct assembly of the engine, and consequently, the reliability and performance of the engine.

[0005] To meet all these specifications, numerous characteristics are defined by the design offices for each engine part. For each characteristic, a tolerance range is defined to assess its conformity. A characteristic whose measurement falls within its tolerance range is considered compliant. A part is compliant if all its characteristics are compliant. To produce compliant parts, manufacturing generally involves several, often numerous, steps. For example, for a turbine blade (which is typically about ten centimeters long), more than twenty manufacturing or inspection operations are necessary. Indeed, it is important to ensure that production runs smoothly in order to detect any potential problems as early as possible and avoid having parts with non-compliant characteristics.

[0006] More generally, since the parts are complex and often made of expensive materials such as titanium, it is imperative not to continue the manufacturing process of a part which is known to be non-compliant once finished and which will then have to be reworked or even rejected.

[0007] The data collected during these inspections are currently used solely to ensure the conformity of the parts. However, all these inspections are costly. In the aerospace industry, the manufacture of complex parts requires high-tech machining and measuring equipment, which is particularly expensive. Therefore, the higher the production rates, the more machines are needed to produce and measure the parts. Consequently, the time and cost associated with part inspection can represent more than 30% of the total manufacturing time and cost.

[0008] In order to reduce the number of machines required for the same production rate and quality, certain optimizations of the control processes are sometimes possible.

[0009] Thus, a first approach consists of relying on industry expertise, i.e., on the physical knowledge of manufacturing processes held by experts. This approach has limitations because the physics of the various machining stages is sometimes poorly understood. For example, in aeronautics, for certain parts, several hundred characteristics must be verified after very diverse manufacturing stages: casting, chemical treatment, painting, shot blasting, turning, milling, etc. Thus, it is sometimes impossible to have industry knowledge of the entire part. Moreover, it sometimes happens that industry experts have different opinions on the effects of a manufacturing operation. Finally, even for a group of experts who have mastered all the production stages, listing all the dependencies between the different characteristics is a tedious, if not impossible, task.

[0010] Another approach consists of using statistical methods to analyze the collected data. This is, for example, what is proposed in document FR 3 063 153 A1. However, this type of method (purely statistical), while sometimes useful, does not always provide significant improvements.

[0011] There is therefore a need to optimize the production process of such parts, in particular to reduce the inspection time of said parts.

[0012] Also, an objective of the invention is to propose a solution to improve the current situation.

[0013] In particular, an objective of the invention is to propose a solution enabling the identification of a possible link between characteristics of the part, defined in the plan, at the same stages or at distinct stages of the production of the part in order to reduce the number of measurements to be taken on the part during its control. Summary of the invention

[0014] To this end, a method for inspecting a part is proposed, comprising the following steps: A) provide a part which must be checked on a plurality of characteristics, during or at the end of its manufacture; B) determine dependency values ​​between said part characteristics, according to the following computer-implemented steps: (b) choose a correlation law estimator between said characteristics from a sample of measurements on said plurality of characteristics relating to N > 30 parts all manufactured in accordance with said part provided in step A) such that all the parts follow the same statistical law, B2) estimate pairwise dependency values ​​D kl between said characteristics Y k, K / of the part, with 1 < k < d and 1 < 1 < d with the estimator chosen in step Bl), B3) organize the characteristics Y k, Y / of the part in the form of K blocks Bj with 2 < j < K < d, each block Bj being defined on the basis of the closest dependency values ​​D kl obtained at the end of step B2) and to which is finally associated a common dependency value D(Bj), said common dependency value D(Bj) being greater than or equal to a predefined threshold T such that 0 < T < 1 for Kl block(s) and strictly less than said threshold T for the remaining block; C) Check the part supplied in step A) for the aforementioned characteristics according to the following steps: C1) measure one or more, but not all, of the part's characteristics that are associated with the K1 block(s) having a dependence value greater than or equal to said threshold T and measure all the part's characteristics that are associated with the remaining block having a dependence value strictly less than said threshold T, C2) predict the characteristics of said part for the characteristics that were not measured in step C1).

[0015] Thus, thanks to the method according to the invention, the number of measurements to be taken on the part during its inspection can be reduced, as the unmeasured characteristics can be predicted. This results in a time saving for carrying out the inspection of the part.

[0016] [The method according to the invention may comprise one or more of the steps below, taken individually or in combination with each other: - the correlation law estimator chosen in step Bl) is a binary law, that is to say for which D kl is 0 or 1; - the correlation law estimator chosen in step B1) to determine the dependence values ​​Dkl is a covariance estimator [Maths 1] = Cov(YY[) where by definition [Maths 2] Cov ( Yj Yj ) = E [ YkY[ ] - E [ Yk ] where E denotes the function for calculating the average over the N pieces; - the correlation law estimator chosen in step B1) to determine the dependence values ​​D is a Pearson estimator [Maths 3] Dki = R(Xk> where [Maths 4] Ca^X^X^ with by definition [Maths 5] RIXk, À / = I , r ^(XX^X^) Cov(Yk Y / ) =E[ Y^Yj] -E[ Yk] in which designates the function for calculating the average over the N pieces; - the number K of blocks to complete step B3) is determined by a user; - the number K of blocks to perform step B3) is determined by implementing the following sub-steps: for each possible value of K, determine for each block Bj such that D(Bj) > T, the number n of features that are dependent on another feature; choose K as the smallest value that maximizes the number n; - the number K of blocks to perform step B3) is determined statistically; - the process implements a hierarchical ascending classification to, during step B3), define each block Bj on the basis of the closest dependency values ​​D obtained at the end of step B2) and associate it with a common dependency value D(Bj); - the process implements a stochastic block method with degree correction to, during step B3), define each block Bj on the basis of the closest dependence values ​​D obtained at the end of step B2) and associate it with a common dependence value D(Bj); - the threshold T is such that 0.5 < T < 1; - the part is an aircraft part; - the aircraft part is an aircraft engine blade; - the aircraft part is a rotating part of the rotor type for an aircraft engine. Brief description of the figures

[0017] The invention will be better understood with the aid of the following description, given solely by way of example and made with reference to the accompanying drawings in which: - [Fig.1] is a diagram representing a method according to the invention; - Fig. 2 is an example of linear correlation between two distinct characteristics; - Fig. 3 is an example of a non-linear correlation between two distinct features; - Fig. 4 represents a graph showing dependencies between the characteristics of a part, before (left) and after (right) permutation of blocks to the closest data, in accordance with a step according to the invention; - Figure 5 shows a fan blade for an aircraft engine and its division into numerous sections in which specific characteristics are to be controlled. Detailed description of the invention

[0018] Fig. 1 is a diagram representing a method 100 according to the invention.

[0019] In the process 100 of inspecting a part during or after its manufacture, it firstly, it is appropriate to A) provide a part which must be checked on a plurality of characteristics, during or at the end of its manufacture.

[0020] Then, in step B), it is necessary to determine dependency values ​​between said part characteristics. This is carried out according to the computer-implemented steps B1), B2), and B3).

[0021] First, step B1) consists of choosing an estimator of the correlation between said characteristics from a sample of measurements on said plurality of characteristics relating to N > 30 parts, all manufactured according to said part provided in step A), such that all the parts follow the same statistical distribution. From a practical point of view, the fact that the parts follow the same statistical distribution can be achieved by identical manufacturing of the parts, but not only that. More broadly, one can predict similar, but not necessarily identical, manufacturing of the different parts, based, for example, on a cause-and-effect diagram (also called an Ishikawa diagram or 5M diagram).

[0022] The correlation estimator chosen in step B1) can be based on a binary law, for example, a test of independence in which the elements of the dependence matrix are 0 (no dependence) or 1 (proven dependence). In this case, it is simply a matter of determining whether two characteristics are dependent on each other or not.

[0023] It is however possible to determine more precisely the level of dependence between two characteristics.

[0024] Thus, alternatively, the correlation estimator chosen in step B1) to determine the dependence values ​​D can be a covariance estimator [Maths 6] = Cov ( Y^ Y[ ) where by definition [Maths 7] ^[YM- of average over the N pieces. The choice of covariance as estimator is well suited for linear dependencies between characteristics (cf. [Fig.2] - a first characteristic on the abscissa and a second characteristic, distinct from the first characteristic, on the ordinate).

[0025] According to another variant, the correlation estimator chosen in step B1) to determine the dependence values ​​D can be a correlation estimator of gr ]£fyj in which E denotes the calculation function Cov (Y. Y A Pearson [Maths 8] = Xt) where [Maths 9] R (A Xi) - "j............-.........; The choice of the Pearson correlation estimator is well suited for linear dependencies between features (see [Fig.2]).

[0026] From a practical point of view, it should be noted that the estimator, due to measurement uncertainty, does not allow the identification of independence between two features by a calculation leading exactly to D kl = 0. In the literature (for example, Adaptive Thresholding for Sparse Covariance Matrix Estimation, Cai and Liu, 2011, arXiv: 1102.2237vl), a classic way to resolve this difficulty is to consider a slightly modified version of the empirical correlation by setting a threshold hj. In this case, the dependence value D kt is written [Maths 10] Du = R(Xb XZ)°Û:

[0027] Xi) ^x^ Xù Xi) Xt) | > 4 / d

[0028] This is the mathematical formulation to say that if the correlation is smaller in absolute value than the threshold Akl, then the correlation is estimated to be 0. Otherwise, it is modified proportionally to the threshold Akt. This threshold is typically a positive real number, which is proportional to [Maths 13] [Math.sll],where [Mathsl2] where d is the number of features measured on the N pieces of the sample.

[0029] There are many other correlations on which to build an estimator that can be used within the scope of the invention. For example, it is possible to consider other types of correlation, such as the correlation of distances (Measuring and testing dependence by correlation of distances, Szekely et al., The Annals of Statistics, 35(6), pp. 2769–2794, 2007), to detect a nonlinear dependence (see [Fig. 3]). It is also possible to consider a correlation to detect a monotonic dependence, such as Spearman's rank correlation. Other correlation coefficients are more universally applicable.

[0030] Then, step B2) consists of estimating pairwise dependence values ​​D between said characteristics Yk, Y / of the part, with l <k<detl<l<d avec l’estimateur choisi à l’étape Bl).

[0031] Finally, step B3) consists of organizing the characteristics Y k, Y / of the part in the form of K blocks Bj with 2 < j < K < d, each block Bj being defined on the basis of the closest dependency values ​​D kt obtained at the end of step B2) and to which is finally associated a common dependency value D(Bj), said value of dependency D(Bj) common being greater than or equal to a predefined threshold T such that 0 < T < 1 for Kl block(s), and strictly less than said threshold T for the remaining block.

[0032] During step B3), there are different points to determine, namely the number K of blocks to consider, the way to aggregate the dependency values ​​between features of the part in the same block, which are ultimately intended to be associated with a common dependency value, and the threshold value T beyond which the block thus aggregated will be kept or not.

[0033] As regards the number K of blocks to be taken into consideration, various ways of proceeding can be envisaged.

[0034] Thus, the number K of blocks to carry out step B3) can be determined by a user, for example by a business expert (e.g., manufacturing manager, production engineer or other) or a big data expert (“data scientist” according to Anglo-Saxon terminology).

[0035] According to one embodiment, the number K of blocks required to perform step B3) is determined statistically. For example, one possible statistical method is to consider a penalty term that decreases the likelihood value as the number K of classes increases. See Charles Bouveyron et al., "Model-based clustering and classification for data science: with applications in R," Cambridge University Press, 2019. According to another example, and advantageously, a method based on statistical tests can be used. This method consists of testing, for different values ​​of K, whether the data are realistic with respect to the model with K blocks. See Wang et al., "Likelihood-based model selection for stochastic block models," Annals of Statistics, 45 (2), pp. 500-528 (2017), the contents of which are available in the R software library called "randnet."In addition to allowing the selection of an appropriate number of blocks, this method ensures that the statistical model is consistent with the data.

[0036] According to an alternative embodiment, the number K of blocks to perform step B3) is determined by a method specific to the invention, by implementing the following sub-steps: - for each possible value of K, determine for each block Bj such that D(Bj) > T, the number n of features that are dependent on another feature; - choose K as the smallest value that maximizes the number n.

[0037] Regarding the method of aggregating dependency values ​​between part features within a single block, it is possible to implement hierarchical ascending classification (HAC). Within the framework of the invention, this approach makes it possible to define each block Bj based on the closest dependency values ​​D obtained at the end of step B2) and to associate it with a common dependency value D(Bj). Hierarchical ascending classification is notably presented by Christian Paroissin, “Practicing data science with R: arranging, visualizing, analyzing and presenting data”, Practicing data science with R, 2021, p. 1-282.

[0038] Alternatively, and for the same purpose, a method for estimating a so-called block stochastic model with degree correction can be implemented.

[0039] We first present the stochastic block method.

[0040] This method is a random graph model. A random graph is a The set (V, E, P) such that V = {1, ..., d] is called the set of nodes (here the set of nodes corresponds to the d characteristics of the part to be inspected), [Maths 14] Ec Vx V is called the set of edges (here, constructed from combinations of characteristics of the part), and [Maths 15] P = {Pki} a set of probability distributions. A realization of a random graph is then a matrix IV such that the value VF pi of the matrix was randomly drawn according to the probability distribution Pk4. Thus, applied to the invention, VF is a matrix whose terms are likely to represent the dependence values ​​D obtained at the end of step B2).

[0041] Among random graphs, the particularity of the block stochastic model is that there exist permutations of the graph's nodes such that the matrix IV resembles a block matrix once its rows and columns have been permuted (see [Fig. 4], left before permutation and right after permutation). In mathematical terms, this amounts to saying that the matrix IV can be obtained in the following way: - a vector p called the "latent partition" is drawn at random such that the kth coefficient pk of the vector is drawn independently of the others on {1, ..., K], - once this vector has been drawn, a collection of probability distributions [Maths 16] (Pk'j'} is defined such that [Maths 17] Wik — is drawn according to [Maths 18] Ppkp;. The observed value therefore depends only on the class of nodes k and 1 in p.

[0042] Thus, if the matrix IV can be put in this form, applying an estimation method adapted to this kind of model makes it possible to find p (which is unknown in practice) and we can then estimate a dependence value by looking at the values ​​of each of the groups (likely to be associated with the blocks Bj).

[0043] Within the framework of the invention, the correlation between features (unknown) is provided by the estimator chosen in step Bl1), for example the Pearson correlation estimator [Maths 19] R or, possibly, its modified version [Maths 20] r.

[0044] With its estimators, there is a degree of heterogeneity to be taken into account within the blocks (whereas theoretically, two given variables will always have exactly the same dependence).

[0045] This is where the stochastic block model with degree correction becomes interesting.

[0046] Indeed, this model differs from the block stochastic model by incorporating a parameter k, for each node i which represents its degree of heterogeneity. Typically, the model with degree correction is then written as: - a vector p called "latent partition" and a sequence of degrees [Maths 21] (^ ) which are drawn at random, - once the vector p and the sequence kt have been drawn, a family of distribution P& with mean # is defined, and the coefficient k,l of the graph IV is drawn according to a distribution with mean [Maths 22] kpkkp!6pk,pr. The observed value therefore depends only on the class of the nodes k and l.

[0047] To find the latent partition associated with the data, it is necessary to be able to recover the model parameters. The most common approach for this is to maximize a function called the likelihood. This is a classic approach in the literature, which amounts to choosing the model that maximizes the probability of observing the matrix [Maths 23]. See the book by Charles Bouveyron et al., "Model-based clustering and classification for data science: with applications in R," Cambridge University Press, 2019. Then, by determining W, we can estimate a dependence value D(Bj) by looking at the values ​​of each of the blocks Bj.

[0048] Finally, with regard to the threshold, it will advantageously be chosen such that 0.5 < T < 1. The choice of this threshold is particularly well suited with the Pearson correlation.

[0049] Once step B) has been carried out, it is then possible, during step C), to check the part supplied in step A) against the aforementioned characteristics. This can be carried out according to the following steps C1) and C2).

[0050] Step Cl) consists of measuring one or more, but not all, of the part's characteristics that are associated with the K-1 block(s) having a dependence value greater than or equal to said threshold T and measuring all the part's characteristics that are associated with the remaining block having a dependence value strictly less than said threshold T. In the case of the remaining block, the dependence is indeed not sufficient to establish links between these characteristics.

[0051] As for step C2), it consists of predicting the characteristics of said part for the characteristics that were not measured in step Cl). This prediction can therefore only be associated with characteristics belonging to the block(s) having a dependence value greater than or equal to the threshold T. Step C2) is advantageously implemented by computer.

[0052] An example of the application of the invention corresponds to the control of an aircraft engine blade at the end of its manufacture (final control).

[0053] Each motor contains about twenty blades. Thus, a blade is a part that is produced in very large quantities. Therefore, it is important to optimize as much as possible. Its manufacturing process and one way of doing things is to reduce measurement time during the final inspection of the part.

[0054] A distinctive feature of a blade is that it is divided into several sections (see [Fig. 5]). We can consider 28 sections in our example. On each of these sections, approximately twenty characteristics are checked. Their number may vary depending on the section, but no characteristic is specific to only one section. In total, there are therefore several hundred characteristics to check. These characteristics may represent angles, flatness measurements, position measurements, etc.

[0055] Due to the manufacturing process of the blades, experts expect dependencies between the same characteristic measured at different sections. However, these dependencies may either remain the same throughout the entire blade or depend on the section's position on the blade; that is, two characteristics will be more dependent if the associated sections are close together on the part. In cases where this is not exactly what is observed, experts expect to observe up to three different behaviors (a first behavior for the blade's root, a second behavior for the blade's center, and a third behavior for the blade's tip). However, since the behavior can vary from one characteristic to another due to the complex nature of the part and its manufacturing process, it can be difficult to detect the characteristics that are highly dependent on each other.

[0056] On an aircraft engine blade, measurements are typically taken on N = 1500 parts whose manufacturing process follows the same statistical law. Here, since we expect to have about sixty blocks (around twenty characteristics, 3 distinct behaviors per characteristic), the number K of blocks (step B3) can be determined by subject matter experts, so in this case, K = 60, for example. Incidentally, it should be noted for the purposes of this exercise that if we define the number K of blocks automatically (i.e., not imposed by a user, for example, a statistical method such as the one proposed in Bouveyron et al. or in Wang et al., or alternatively, a method specific to the invention described above), we also find about sixty blocks. Thus, the method produces a number of blocks that is relevant to the physics of the part.Furthermore, if we look at the blocks in detail, for almost each of them, the characteristics inside correspond to the same characteristic which is repeated over several sections.

[0057] Furthermore, if we look at how the sections for the same feature are distributed, we can see that if they are in different blocks, then these blocks maintain the geometry of the part in the sense that there are no blocks with a feature whose section is not adjacent to another section in the same block. For example, if a feature is present in sections one through six, and that they are distributed over two blocks, we can have sections 1, 2 and 3 (which follow each other) in one block, and sections 4, 5 and 6 in the other, but not, for example, sections 1, 3 and 5 (which do not follow each other) in one block and sections 2, 4 and 6 in the other block.

[0058] Here, and to complete the conditions for implementing the method according to the invention, a Pearson correlation estimator for the dependence (step B1) was chosen, based on a sample of N = 1500 blades. With regard to step B3), if the number K of blocks is, for example, K = 60 (see above), the likelihood maximization method of a stochastic block model with degree correction was used to aggregate the data, and finally, a threshold T = 0.5 was chosen.

[0059] After implementing the invention, 11 sections of the blade (out of the 28 sections defined for the blade) were measured, allowing the remaining 17 sections to be predicted. If these are predicted to be compliant, the part is accepted; otherwise, the remaining 17 sections are measured to assess the part's conformity. In the case of a compliant prediction, the inspection time for a blade is reduced from 14 minutes to 10 minutes, representing a time saving of almost 30%. Of course, if the prediction is not compliant, measurements must then be taken on the remaining 17 sections, resulting in a loss of part inspection time equal to the computation time associated with implementing the method. However, statistically, this is a much rarer situation than one resulting in the prediction of part conformity, and overall, the time saved in inspection is substantial.

[0060] The invention thus makes it possible, based on production data, to transmit information on the physics of the manufacturing process and therefore of the part. Furthermore, thanks to step B2), it is easy to select from these blocks those that correspond to strong dependencies between characteristics. These blocks can then be studied in more detail to determine whether it is possible to optimize the part measurement process.

[0061] Another example of application relates to the manufacture of an aircraft engine rotor.

[0062] Manufacturing a rotor involves dozens of operations, including machining and inspection. In particular, the part is inspected at various points. One way to optimize production is to reduce the number of inspections. To do this, it is necessary to understand the influence of the manufacturing operations performed between inspections. However, this understanding of the impact of the different operations can be difficult to obtain and may even lead to differing opinions among experts. By implementing the invention, it is possible to detect blocks of dependent characteristics based on measurements (approximately one hundred different characteristics are measured on the part in total). This is particularly relevant for parts of this type that are produced in large quantities. For blades, which are less complex, we generally have around a hundred observations. For this application, there are no blocks known a priori by the subject matter experts. Indeed, we don't know if there are dependencies between the measurements of the same characteristic that is checked at different stages of manufacturing. Furthermore, it is possible that only some of the characteristics are measured during intermediate checks. Therefore, in this application, it is necessary to automatically determine the number K of blocks. By applying this, we discover that the value of certain characteristics hardly changes at different stages of manufacturing. We can therefore consider the need to maintain both checks during manufacturing. Moreover, and more surprisingly, the method according to the invention makes it possible to highlight dependencies between different characteristics at different stages of manufacturing.These dependencies, which are not the first that industry experts think of, might not have been identified without the contribution of the invention.

Claims

1. Demands Method (100) for inspecting a part (P) comprising the following steps: A. provide a part that must be checked on a plurality of characteristics, during or at the end of its manufacture; B. Determine dependency values ​​between said part characteristics, according to the following computer-implemented steps: B1) choose a correlation distribution estimator between said characteristics from a sample of measurements on said plurality of characteristics covering N > 30 parts all manufactured according to said part provided in step A) such that all parts follow the same statistical distribution, B2) estimate pairwise dependence values ​​Dkl between said characteristics Yk, Y, of the part, with l <k<6?etl<l< d avec l’estimateur choisi à l’étape Bl), B3) organize the characteristics Y k, Y / of the part in the form of K blocks Bj with 2 < j < K < d, each block Bj being defined on the basis of the closest dependency values ​​D kl obtained at the end of step B2) and to which is finally associated a common dependency value D(Bj), said common dependency value D(Bj) being greater than or equal to a predefined threshold T such that 0 < T < 1 for Kl block(s) and strictly less than said threshold T for the remaining block; A. Check the part supplied in step A) for the aforementioned characteristics according to the following steps: Cl) measure one or more, but not all, characteristics of the part that are associated with the Kl block(s) having a dependence value greater than or equal to said threshold T and measure all the characteristics of the part that are associated with the remaining block having a dependence value strictly less than said threshold T, C2) predict the characteristics of said part for the characteristics which were not measured in step Cl).

2. Method (100) according to claim 1, wherein the correlation law estimator chosen in step Bl) is a binary law, i.e. for which D kl is 0 or 1.

3. Method (100) according to claim 1, wherein the correlation law estimator chosen in step B1) to determine the dependence values ​​D is a covariance estimator = Cov ( Yk, Yi ) where by definition Cw(Yh Yi) = E[YkYl] - E[in the 9 which Edesiëne the mean calculation function over the N pieces.

4. Method (100) according to claim 1, wherein the correlation law estimator chosen in step B1) to determine the dependence values ​​D is a Pearson estimator Dki -RiX^ Xt) where CoAx^x,) with by definition Xi) — । ^CwX^X^Co^X / Xi) Cov(Yfr Y / ) = e [ y^ ] - e [ Yk ] Xyp in leouel E denotes the calculation function of the average over the N pieces.

5. Method (100) according to any one of the preceding claims, wherein the number K of blocks to perform step B3) is determined by a user.

6. A method (100) according to any one of claims 1 to 4, wherein the number K of blocks to carry out step B3) is determined by implementing the following substeps: - for each possible value of K, determine for each block Bj such that D(Bj) > T, the number n of features that are dependent on another feature; - choose K as the smallest value that maximizes the number n.

7. A method according to any one of claims 1 to 4, wherein the number K of blocks to carry out step B3) is determined statistically.

8. A method (100) according to any one of the preceding claims, wherein a hierarchical ascending classification (HAC) is implemented to, during step B3), define each block Bj on the basis of the closest dependency values ​​D obtained at the end of step B2) and associate it with a common dependency value D(Bj).

9. A method (100) according to any one of claims 1 to 7, wherein a stochastic block method with degree correction is implemented to, in step B3), define each block Bj on the base of closest dependency values ​​D obtained at the end of step B2) and associate it with a common dependency value D(Bj).

10. Method (100) according to any one of the preceding claims, wherein the threshold T is such that 0.5 < T < 1.

11. Method (100) according to any one of the preceding claims, wherein the part is an aircraft part.

12. Method (100) according to the preceding claim, wherein the aircraft part is an aircraft engine blade.

13. Method (100) according to claim 11, wherein the aircraft part is a rotating part of the rotor type for an aircraft engine.

Images