EP2391929A1

EP2391929A1 - Method for detecting atypical electronic components

Info

Publication number: EP2391929A1
Application number: EP10702298A
Authority: EP
Inventors: François BERGERET; Anne Ruiz; Carole Soual; Henri Caussinus
Original assignee: SARL IPPON
Current assignee: SARL IPPON
Priority date: 2009-02-02
Filing date: 2010-02-02
Publication date: 2011-12-07
Also published as: FR2941802A1; JP2012516994A; CN102388347A; US20120053877A1; SG174352A1; WO2010086456A1; FR2941802B1

Abstract

The invention relates to a method for detecting atypical electronic components for the quality control of n electronic components at the end of the manufacturing process, said components being subjected to a number p of unit tests providing digital data, said set of n components including electronic components having a response to each of the p unit tests that lies within predetermined limits particular to each of the p tests, wherein the method comprises using the multi-dimensional information of the p dimension responses of said n electronic components. The method uses a generalised principal component analysis for detecting atypical parts in the semiconductor field, or in the fields including modules assembled using electronic components (e.g. an ABS module, a smart card, etc.). The aim of the method is to get close to "zero defect", wherein no part is detected to be substandard by the client.

Description

METHOD FOR DETECTING ATYPICAL ELECTRONIC COMPONENTS

The present invention belongs to the field of quality control of parts and in particular of electronic components.

Background of the Invention and Problem The semiconductor industry produces integrated circuits, referred to as electronic components, which are fabricated on groups of silicon wafers, each wafer comprising several hundred components.

In order to guarantee the operation of these electronic components, a first series of tests called under-probe tests (or probe) is performed on each of the components while they are still part of a wafer.

Each of these tests which consist respectively of an electronic measurement is associated with a specification limit determined, among others, with the customer to whom these electronic components are intended.

The electronic components, whose response to at least one test is not in accordance with the specification of this test of this first series of tests (probe), are then considered as defective, rejected during their separation from the wafer.

On the contrary, the electronic components whose answers are compliant for all the tests are assembled in a case and then again tested by a second series of tests.

As in the first series of tests, specification limits were determined with the customer to whom these electronic components are intended, and the electronic components of which at least one test response is not in accordance with the specification of that test of that second. series of tests are rejected. This second series of tests can be duplicated at several temperatures (-40 ⁰ C, + 90 ° C for example).

Thus by this commonly used method, a component is rejected and therefore not delivered to the customer if at least one test response (belonging to the first test series or the second series of tests) is outside the specification limits associated with the test. this is.

Now it happens that the parts delivered so that have passed all the tests successfully present a latent defect which will be revealed during the implementation of the part within the framework of the application of the customer, from delivery or even later in the final application (an ABS type brake for example).

This quality control, as it is currently practiced in the usual way, is therefore insufficient and some complementary methods are already at work, for example in the components intended for the automotive industry, to minimize these quality problems seen by the customer.

These complementary methods are performed on the electronic components, usually after the first series of tests and / or after the second series of tests and use the distributions of the results to each of these tests to eliminate atypical electronic components called outliers. They are therefore used test by test for each of the tests or for a part of the two series of tests.

For example a method called Part Average Testing (PAT) compares the response of a test of an electronic component to the average distribution of responses of this test of the other electronic components and considers as atypical electronic component, an electronic component admitting a response to a test too distant from the distribution of the answers to this test of the other electronic components. Similarly, a method called Part Average Geographic Testing considers as atypical electronic component, an electronic component surrounded, during testing for example on a silicon wafer, non-compliant components. One then tends to consider that the component surrounded by defective components is probably defective by "geographical" proximity.

Another complementary method consists in creating mathematical models of regression, that is to say of correlation between the results of the components to the various tests, and to consider as atypical and therefore potentially defective, the electronic components whose correlation between two tests is not consistent with the average obtained for the other electronic components.

However, these additional methods, although constituting improvements with respect to prior testing methods, still have disadvantages. Typically, they still allow electronic components with a latent defect to be considered reliable and deliverable to the customer.

This inconvenience is inconvenient, on the one hand, because it forces the manufacturer to send the customer a new batch of replacement parts, and decreases the level of quality perceived by the customer, but, even more so, because some of these components, although of low unit cost, are critical components in the operation of a more complex system, for example a motor controller or an ABS braking system. In this case, a failure of the component can lead to a serious accident, the consequences of which go far beyond the mere financial value of the component. This risk leads manufacturers to choose to reject too many components, including many correct components, because they use univariate methods (PAT ...) or bivariate (regression ...) on a very large number of tests, which deprives them of a few percent of their production, yet does not guarantee them to eliminate all potentially defective components. Although already given a certain performance, these methods are therefore insufficient to reach the zero defect.

Objectives of the invention

The objective of this invention is therefore to provide a method for refining the detection of atypical (and therefore potentially defective) electronic components in a set of electronic components subject to a large number of tests in order to tend towards the zero defect, in accordance with to the requirement for example of the automotive industry.

According to a second objective of this invention, it does not require the realization of new tests on electronic components already tested by conventional methods.

A third object of the invention is to bring back, in the category of components that comply with the specifications and therefore sellable, components that would have been erroneously eliminated (false negative) by the above methods. According to a fourth object of the invention, this may in some cases allow the manufacturers of electronic components to eliminate expensive reliability tests, called "burnin", the parts rejected during this burnin being captured by our invention.

Presentation of the invention

For this purpose, the invention relates to a method for detecting atypical electronic components, intended for the quality control of a set of n electronic components at the end of manufacture, said components being subjected to a number p of unit tests providing digital data, this set of n components consisting of electronic components whose response to each of the p unit tests is contained within predetermined limits called limits of customer specifications, and specific to each of the p tests, using the multidimensional information p-dimensional responses of these n electronic components It is understood that unlike the state of the art that works in one dimension or two dimensions, this process will work in p dimensions and therefore will be able to use all the information of p tests, and consequently identify more atypical components or question some rejected components.

Indeed, for the majority of the atypical components, their latent defect is detectable in the atypical of these electronic components, if one considers all the answers to the tests on all the electronic components to be tested. According to a preferred embodiment, the method of the invention comprises a proposition of a number q less than p of relevant linear combinations of the p tests which comprise an arbitrarily large part of the information present in the p tests.

The use of a Principal Component Analysis considerably reduces the number of working dimensions, while retaining a very significant part of the information present in the initial point cloud, each point corresponding to a result of a test for an electronic component. The extracted information will suffice to characterize a structure of the n electronic components and thus to reveal the atypical electronic components. According to a preferred embodiment, the choice of q linear combinations of the p tests is carried out by establishing a Principal Component Analysis.

Generalized with a choice of metric M adapted to the p tests of the n electronic components.

We choose here to move towards a particular type of Principal Component Analysis, called Generalized Principal Component Analysis, whatever the metric used.

If, for example, the p tests have a common unit of measure, we can use for example the Euclidean metric and perform a Principal Component Analysis with this metric. According to an advantageous embodiment, the metric M is chosen such that

M = W ^{~ ι} (inverse of matrix W) with

W - order square matrix Where exp represents the exponential function and X ₁ column vector associated with an electronic component i among the n electronic components, of dimension p corresponding to p respective responses to each p tests on this electronic component i.

X _n = - VX ₁ vector of empirical averages

'(X ₁ - X _n ) is the transposed vector of (X ₁ - X _n )

Λ n

V _n = - Y] (X ₁ - X _n Y (X ₁ - X _n ), the matrix of the variances and

usual empirical covariances V _n which is a square matrix of order p

V _n ¹ is the inverse matrix of the matrix of usual empirical variances and covariances V _n . β being a real small number.

It is understood that by using such a metric, the problem of centering the data is overcome, since the set of vectors (X ₁ -X _n ) is centered and the problem of the differences of the units is overcome. measure or scale p tests against each other using the standard | x | _ι .

According to a preferred embodiment, the main vectors are chosen equal to the first q eigenvectors associated with the largest eigenvalues among the set of eigenvectors obtained by the Principal Component Analysis, the number q being determined using a criterion previously selected. A method of automatically calculating the number q of the main vectors that will be used to evaluate each component is determined by the method.

Preferentially, this criterion is such that the eigenvalue associated with a principal component is strictly greater than 1 + β.

According to a preferred embodiment, at least one projection is used on a vector subspace generated by a sub-family of the main vectors and least one criterion to identify atypical electronic components.

More particularly in dimension 2, this or these vector subspaces are vectorial planes and the criterion for a vector plane, to identify the atypical components is done by considering the projection of the vectors X ₁ on this vector plane, and by defining a circle ray of confidence r encompassing a so-called "majority" cloud containing by definition the projection of all the typical electronic components, and declaring that an electronic component i is said to be atypical if the projection of X ₁ on the vector plane is in outside the circle of trust. Even more particularly, the radius r of the circle of confidence for a level of significance OC, is defined by the square root of the order quantile

1 -α of a law of χ ² to (2x ^ 1 + β) degrees of freedom.

For a vector ^, the norm of its projection on the vector plane defines a score. The electronic components are then ordered according to this score and eliminated if their score is greater than a threshold previously calculated or chosen.

According to a particular embodiment, the criterion for identifying the atypical electronic components uses the calculation of a score corresponding to its norm for each component, and a statistical limit for this score.

The invention also relates to software implementing the method as described.

Brief description of the figures

The aims and advantages of the invention will be better understood on reading the following description made with reference to the drawings in which: FIG. 1 represents a projection of the vectors characterizing the electronic components and the respective responses to tests on a sub two-dimensional space, generated by the first two main components of the system, in this figure, the atypical components remote from the central point cloud, detected by the method according to the invention, are marked by stars, - Figure 2 illustrates insertion of atypical parts removal steps by the method according to the invention, within the known component control method before delivery to a customer.

Detailed description of an embodiment of the invention The invention is implemented by computer software running on a microcomputer or other standard type computer.

The invention is intended to be used during a manufacturing quality control of the electronic components and this: 1 / at the end of the tests under probe (probe) which consist of several electronic measures and which will be called first series of tests after rejecting the electronic components of which at least one response to at least one test in this first set of tests is outside the specification limits related to that test 2 / and then at the end of the tests that are performed (second set of tests) after assembling the correct electronic components, that is to say the electronic components having passed the tests under a peak and the test of the method of the invention, in a housing.

The method according to the invention can be used after the first series or after the two series of tests, indifferently. It actually uses any number of tests performed on the electronic components considered.

The method according to the invention can also be used for testing electronic modules containing components: ABS modules, airbag, smart cards, etc.

We denote by n the number of electronic components of the current series to study and p the number of tests of the current series.

It is considered that among the n electronic components, the electronic components whose response to at least one test has been outside the specification limits of this test have already been eliminated.

We thus obtain a data table composed of n individuals (the electronic components) and p variables (corresponding to each of the p tests of the current series) for each of the individuals. The values associated with these p variables are quantitative and real numerical data. To each individual i (Ze [1, «]), we associate a vector-individual X ₁ (which will be called by abuse of language in the following of the description individual X ₁ ) of dimension p having for coordinates on each axis i the answer obtained in the index test i.

The aim of the invention is to determine atypical individuals among all individuals X ₁ e IR ^P. To achieve this goal, we will use so-called "revealing projections" techniques. A revealing projection is a projection of the cloud of individuals X ₁ onto a subspace of dimension q (q <p) capable of to highlight a possible particular structure of the distribution of these individuals.

In the case of electronic components, p being a large number

(several hundred tests are typically performed for the validation of an electronic component), it is interesting to look for whether we can define q linear independent combinations (in the sense of the linear algebra) of the p variables which make it possible to restrict the study from the set of individuals Jf ₁ (of dimension p) to a number q substantially smaller than p, without losing information present in the initial variables p, or with a loss of information estimable in proportion to the total information contained in the p variables.

In order to determine these independent linear combinations, a Generalized Principal Component Analysis (ACPG) is performed.

It is recalled that a Principal Component Analysis (PCA) allows to visualize and summarize in a few dimensions (q) the overall structure of the cloud of the individuals instead of to visualize it in dimension p.

Without going into the details of this technique, known in itself, it is recalled that it consists in determining the axes of inertia of a cloud of points (the individuals) in a space of p dimensions (the variables), these axes ( orthogonal by construction) being linear combinations of the axes of origin, but supporting, by definition, a significant part of the inertia of the points of the cloud (here the individuals), ie of the information contained in these individuals .

There are as many axes of inertia as initial axes, but this analysis in main component makes it possible to know the quantity of information present on each one of these axes. By classifying the axes of inertia by the amount of information contained, we obtain the principal axes of inertia, and we generally find that some main axes of inertia contain in fact a considerable part of the total information of the individuals. Typically, a few tens of main axes of inertia comprise more than 99.9% of the total information of the few hundreds of initial axes. Therefore, it is possible to restrict the study of individuals, which should be carried out on p axes or dimensions (a few hundred), to an arbitrary value q of dimensions, according to the share of information that one is ready to not used.

The q linear combinations independent of the original axes (the variables) will then be the main axes (principal components) deduced from the Analysis in Principal Components. In order to refine the q main components, instead of selecting a classical Principal Component Analysis, we choose to use here a Generalized Principal Components Analysis (ACPG), which consists in making a choice of metric (ie mode of calculation of distance between individuals, many distances that can be defined mathematically on the same space), optimized in the method according to the invention, in the particular case of electronic components.

We recall that in the classical Principal Component Analysis (PCA), the metric used M is either the Euclidean metric (M = Id) or the inverse variance metric M = D ₁₇₅₂ (diagonal matrix S = standard deviation). The steps of the process according to the invention are as follows:

Step 1: Creation of n vectors X ₁ of dimension p. This step is supposed to be known, the result files of the n electronic components during the p tests forming an input data of the process. The vectors X ₁ are stored in an ad hoc database. - Step 2: Use the chosen metric. The choice of the metric used in the present process is particularly important. In the preferred implementation, we chose a metric M inspired by the works of H Caussinus and Anne Ruiz-Gazen, and in particular inspired by an article published in the journal of applied statistics, volume 50 n ° 4 (2002) p81 - 94. This metric is adapted to the detection of atypical individuals insofar as it depends on the dispersion of the data, each individual having an influence all the more weak as it is atypical. During the Principal Component Analysis, these atypical individuals will therefore have even more extreme coordinates than with a classical PCA (Euclidean norm) for the different main axes.

This metric is defined by: M = W ^{~ ι} (inverse of a matrix W) in which West defined by:

and is a square matrix of order p.

and X _n = - VX is the vector of empirical averages of

vectors X ₁ ,

'(X ₁ - X _n ) is the transposed vector of (X ₁ - X _n ),

the standard used is defined by '\ x \ _γ i =' XV ^{~ ι} X,

V _n = - Y] (X ₁ - X _n Y (X ₁ - X _n ), the matrix of the variances and

usual empirical covariances V _n which is a square matrix of order p,

V _n ¹ is the inverse matrix of the matrix of usual empirical variances and covariances V _n . exp is the exponential function.

In the formula defining the matrix W, we have introduced a weight function

M || 2

K (x) = Qxp (-x / 2) applied to x = P lX ₁ - X _n I _j with β, a given real small (actually very close to 0: we recommend a value of order of 1 / p but an arbitrary choice can be made with a β between 0.01 and 0.1 - see the works of H. Caussinus and A. Ruiz-Gazen), for each vector X ₁ . We can write using the weight function:

Step 3: Diagonalization of the matrix V _n M where V _n is the matrix of the variances obtained above, and M the metric used also obtained in step 1 (methods of diagonalization of matrix are known to the man of the art, and possibly available in the form of computer libraries), search eigenvalues of this matrix. This step is typical during a Main Component Analysis. Step 4: Calculate the useful dimension q of the projection space. It is recalled that the dimension q determines the number of main axes to which the analysis is reduced, and therefore determines how much information is used among all the information contained in the initial tests. This dimension q (the number of axes) must then be large enough to capture the desired structure (thus being able to determine the atypical individuals, that is to say the electronic components presumably defective) and small enough not to exhibit any artifacts. (false determination of a chip as defective). We recall that if we order the eigenvalues in descending order, the first eigenvectors (in this order) associated with these eigenvalues will then be the main vectors of the system. In this step 4, we choose a criterion that will make it possible to determine the number q of eigenvectors among the eigenvectors that will be sufficient to characterize the atypical individuals of our space of individuals and which will therefore constitute the main vectors of the system.

The projection, obtained by projecting the M-orthogonal individuals on the subset of dimension q thanks to the choice of the metric M, is invariant by affine transformation of the vectors X ₁ . This makes it clear that it only concerns the structure of the cloud of individuals, beyond the various aspects of centering and scale.

The following selection criterion is used: the eigenvectors are kept such that their respective associated eigenvalue is strictly greater than 1 + β.

If we consider a model of atypical values where X ₁ is supposed to be a random vector whose probability law is a mixture of q + \ normal laws (in different proportions) of variable mean: the majority law, and q possibilities of contamination of the average, then some theoretical properties are verified and presented in what follows.

For n large enough, and low proportions associated with q contaminations, the atypical values (ie the likely defective components) stand out at best on the projection subspaces engendered by the q averages (associated with contaminations).

Moreover, for n big, the q largest eigenvalues of V _n M converge to a number strictly greater than 1 + β, the following ones converge to 1 + β. By choice, in the present method described in no way limiting, we therefore do not consider the dimensions for which the eigenvalues are less than 1 + β.

Step 5: Determining the dimension of representation. In order to simplify the representation of the cloud of the points X ₁ , one chooses to make the projections on vectorial planes, thus on spaces in two dimensions. These vectorial planes are thus generated by a choice of two eigenvectors among the q eigenvectors chosen (which are the q principal vectors). The set of eigenvectors of a system forms, in a known way, a free (independent) family in the sense of linear algebra. The q vectors chosen from this set of eigenvectors thus form a free family of this family in the sense of linear algebra. Thus, if for example q is equal to 6 and we denote (Prini, Prin2, Prin3, Prin4, Prin5, Prinθ) these six eigenvectors (principal vectors), we can graph the projections of the X ₁ , on the three vector planes generated respectively by

(Prini, Prin2), (Prin3, Prin4) and (Prin5, Prinθ), the other combinations of these six vectors can also provide additional information. The vector planes to be used for any value of the number q of principal components chosen are likewise determined.

Step 6: Use the criterion for determining atypical individuals. In each of the vector planes defined in step 5, it is decided to define a circle of confidence. The detection of atypical electronic components is then done through this circle of confidence (for a level of significance α fixed) encompassing the majority cloud (2) and outside which are located individuals atypical declared.

Figure 1 thus illustrates a projection on two main axes (Prini, Prin2). In this FIG. 1, two elements (1) move away graphically from the majority cloud (2). In this example illustrated in FIG. 1, only the first two main axes have been retained, ie those associated with the two largest eigenvalues (and which therefore comprise the maximum of information).

The distance between the points on these graphical representations corresponds here to an approximation of the distance of Mahalanobis in the sense of the metric M. The radius of the circle of confidence corresponds to the square root of the quantile of ordrel -α of a law of χ ² to (2x ^ 1 + β) degrees of freedom (this law of the chi-two intervenes under the assumption that the data follows a normal distribution and this circle is a sort of counterpart of a confidence interval).

The value of the level of significance α can be left to the choice of the user of the process of the invention, in general α varies from 1% to 5%.

The atypical individuals determined at the end of the process are then listed in an ad hoc table intended for the operator.

It should be noted that for reasons of cost reduction (assembly price and case price) it is preferable to eliminate the atypical electronic components before assembly in a case and therefore it is advantageous to start the process of the case. invention after the tests under the tip in order to try to detect the most atypical electronic components during this phase of production.

Advantages of the invention

The interest of the method as explained is to go from p continuous variables to q <p principal components, which are linear combinations of the initial variables having the following interesting characteristics:

Ordered according to the information returned: the first principal component is the linear combination of the initial variables having the maximum variance.

Principal components are uncorrelated variables. - They are less sensitive to random fluctuations than the initial variables

It should be noted that this is true only in the case of a Principal Component Analysis using the Euclidean metric (M = id). This is no longer true in the case of a Generalized Principal Component Analysis, using another metric.

Variants of the invention

The scope of the present invention is not limited to the details of the above embodiments considered by way of example, but instead extends to modifications within the scope of those skilled in the art. In a variant, the metric M = is used. The values specific to

consider in order to determine the principal components (eigenvectors associated with these eigenvalues) are therefore in this case the eigenvalues strictly greater than 1. In another variant, any metric M suitable for the types of measurements carried out are used. For example, the metric M may be equal to the identity matrix, if the Euclidean metric is chosen, in the case of p measurements with the same unit of measurement.

Similarly, we can choose a metric M equal to the inverse of the variances when the units of measure are not the same for all the variables. In this case, one diagonalizes during the Main Component Analysis the matrix of correlations.

An alternative way to define atypical individuals is to calculate a score for each point, corresponding to its norm computed with its q selected principal components, and to define a statistical limit by a usual method, known per se, (example: a limit of control) to determine which individuals are out-of-distribution, and therefore atypical, for this score (step 6).

The invention encompasses any generalized PCA method, in the sense of diagonalising a covariance variance matrix estimator relative to another covariances variance matrix estimator, the objective of which is to detect atypical observations.

In particular this includes the diagonalization of any VnM operator, where Vn is the usual empirical covariances variance matrix, and M is the inverse of any robust covariance variance matrix estimator (eg a

M-, an S-, a MM or a tau estimator or the minimum determinant MCD estimator).

This also includes the diagonalization of an operator of the form UnM where Un and the inverse of M are two robust estimators. We note that the classical ACP and the so-called robust ACP are particular cases of generalized PCA but that their primary objective is to detect the structure of the majority of the data, and not any atypical observations. The only atypical observations that are detected on the main axes of a PCA usual or robust are those which are atypical in the directions where the dispersion of the majority of the data is maximum.

Thus, there is an essential difference between conventional or robust PCA and the method according to the invention which lies in the choice of dimension. The usual selection criteria for all these methods are based on the eigenvalues of the diagonalized operators: only the principal components associated with the largest eigenvalues are retained.

But whereas the largest eigenvalues of a classical or robust PCA are associated with projection spaces where the dispersion of the majority of data is maximal, the largest eigenvalues of the generalized PCA are associated with projection spaces. which make it possible to identify as well as possible the atypical individuals.

In the case where the size of the data (the number of variables) is large, the robust covariances variance matrix estimator that is involved in the generalized PCA method is not necessarily invertible. To solve this inversion problem, a pseudo-inverse generalized inverse of Moore-Penrose is used.

To obtain an inverse matrix, it is sufficient to calculate the eigenvalues and eigenvectors of the matrix. In the case of a covariance variance matrix, these eigenvalues are real positive.

The inverse matrix is calculated by taking the inverse of the eigenvalues and keeping the same eigenvectors. If the covariance variances matrix is not invertible (which occurs if the number of variables is large compared to the number of observations), it is because it contains eigenvalues close to 0. Taking a generalized inverse is do not invert the eigenvalues close to 0 but take them equal to 0 in the inverse matrix.

This methodology is recommended as soon as the matrix of covariances is poorly conditioned (small eigenvalues), even if the inverse is computable numerically, to avoid too much instability that would result from the large eigenvalues appearing in the inverse.

3) Other methods of revealing projections

Other variants of the invention may be envisaged, including the methods described below. These methods have never been used in the semiconductor industry with the aim of detecting atypical parts as part of the reliability of electronic chips (zero defect). The generalized PCA, which is the subject of the description given above, is a particular method of revealing projections (see Caussinus and Ruiz-Gazen,

2009). To solve the problem of the large number of dimensions relative to the number of observations, other methods of revealing projections type than the generalized PCA are recommended.

The idea is to look for linear projections of data on a

(possibly two) dimension that highlight atypical observations.

Generalized PCA achieves this goal, but can be complemented by other "Projection Pursuit" methods that consist of:

(i) the definition of a projection index which measures, to some extent, the interest of the projection. In this case, a projection is all the more interesting because it reveals atypical observations. In other words, the projection index will be even higher than the projection will highlight atypical,

(ii) the search for one or more projections that correspond to local maxima of the previously defined projection index. The implementation of this second step goes through an optimization algorithm which can be based on a deterministic method of local optima search or on a heuristic in the case where the index function is not sufficiently regular to use deterministic methods based on the gradient.

Projection indices adapted to the search for atypical values include the Friedman index (1987), but also the index of kurtosis (Pena and Prieto, 2001) and the measure of "outlyingness" of Stahel-Donoho (Stahel , nineteen eighty one ). The first two suggested indices measure the interest of a projection in terms of distance to the normal law. It has been noted that the interesting projections obtained are primarily those that move away from the normal distribution distribution tails and thus are projections that may reveal atypical observations.

For the Stahel-Donoho index, it measures the difference in absolute value from the projection of an observation to the median, standardized by the median absolute deviation ("median absolute deviation" in English) of the projected data. It can be generalized to any standard deviation measure of an observation at the center of the distribution. For example, the median can be replaced by the mean and the median absolute deviation by the standard deviation. In the latter case, we find the measurement used in a standard way in the method PAT ("Part Average Testing") mentioned at the beginning of the text. It should be noted that, unlike the PAT method, which only applies to each of the initial variables, the method recommended in this variant aims to propose a PAT test on all the linear combinations of the initial variables that best reveal atypical individuals. . This last method thus makes it possible to take into account the multidimensional relationships that exist within the data, relationships that are not taken into account in the usual PAT method.

As in the usual PAT method, we can decide to choose a threshold value beyond which an individual is declared atypical based on the maximum rejection threshold that we are willing to tolerate (so-called ^" 3 sigma" rule to adapt to data function and the maximum rejection percentage accepted).

It is also recommended to center and make the spherical data before the index optimization step because it has been found on a practical level that it facilitates the discovery of interesting projections. One way to make spherical data is to compute the usual main components.

The invention also relates to any hybrid method using generalized PCR in connection with the revealing projection methods as recommended above. Thus, the identification of atypical points, obtained by maximizing a projection index, can be used to compute a weighted covariances variance matrix estimator (low weights are assigned to individuals declared atypical at step previous). The Stahel-Donoho estimator (Stahel, 1981) is thus defined from the Stahel-Donoho index. This estimator can then be used as a robust estimator in the generalized PCA.

Claims

1. A method for detecting atypical electronic components, intended for the quality control of a set of n electronic components at the end of manufacture, said components being subjected to a number p of unit tests supplying digital data, this set of n components consisting of electronic components whose response to each of the p unit tests is contained within predetermined limits and specific to each of the p tests, characterized:

in that it uses the multidimensional information of the p-dimensional responses of these n electronic components, in that it comprises a proposition of a number q less than p of relevant linear combinations of the p tests which comprise a arbitrarily large of the information present in the p tests,

in that the choice of the linear combinations of the p tests is carried out by establishing a Generalized Principal Components Analysis, with a choice of metric M adapted to the p tests of the n electronic components,

in that the method is implemented at the end of the tests under a spike and / or at the end of tests carried out after having assembled the correct electronic components, ie the electronic components having passed the tests under a spike.

2. Method according to claim 1, characterized in that the metric M is chosen such that:

M = W ^{~ ι} (inverse of matrix W) with

W - order square matrix p, where exp is the exponential function, and

X ₁ column vector associated with an electronic component i among the n electronic components, of dimension p corresponding to p respective responses to each of the p tests on this electronic component i,

X _n = - VX vector of empirical averages,

'(X ₁ - X _n ) is the transposed vector of (X ₁ - X _n ),

V _n = - Y] (X ₁ - X _n Y (X ₁ - X _n ), the matrix of the variances and

usual empirical covariances V _n which is a square matrix of order p,

V _n ¹ is the inverse matrix of the matrix of usual empirical variances and covariances V _n , β being a small real number.

3. Method according to claim 2 characterized in that β is of the order of

1 / p or arbitrarily chosen between 0.01 and 0.1

4. Method according to claim 3, characterized in that the main vectors are chosen equal to the first q principal vectors associated with the largest eigenvalues among the set of principal vectors obtained by the principal component analysis, the number q being determined using an optimized criterion.

5. Method according to claim 4 characterized in that the criterion is such that the eigenvalue associated with a principal component is strictly greater than 1

⁺ β -

6. Method according to claim 4, characterized in that it uses at least one projection on a vector subspace generated by a sub-family of the main components and at least one criterion for identifying the atypical electronic components.

7. Method according to claim 6, characterized in that:

- this or these vector subspaces are vectorial planes,

the criterion for identifying the atypical components is verified by considering the projection of the vectors X ₁ on each vector plane, and by defining a circle of confidence of a radius r encompassing a so-called "majority" cloud containing by definition the projection of the typical electronic components, and declaring an electronic component i is atypical if the projection of X ₁ on the vector plane is outside the circle of confidence.

8. Method according to claim 7, characterized in that the radius r of the circle of confidence for a level of significance OC, is defined by the square root of the quantile of order 1 -α of a law of χ ² to (2 x -Jl + β) degrees of freedom.

9. The method of claim 6, characterized in that the criterion for identifying the atypical electronic components uses the calculation of a score corresponding to its standard for each component, and a statistical limit for this score.

10. Method according to claim 1, characterized in that it further comprises steps in which:

- we are looking for linear projections of data on one or two dimensions that highlight atypical observations,

- we define a projection index, which measures the interest of the projection, the projection index being all the higher as the projection highlights atypical,

one or more projections are sought which correspond to local maxima of the projection index.