Method for determining the abnormality level of an individual, especially for statistical detection of abnormal individuals in multivariate situations

Abstract

The invention relates in particular to a method for determining the abnormality level of individuals in order to statistically detect abnormal individuals in a pre-collected data set obtained from measurements of individual parameters acquired by a plurality of measurement systems, characterized in that the method comprises: a step 100 of pre-processing the data; a step 200 of determining a multivariate abnormality index, said index being based on the pre-processed data and transformed by a function f so that it is between 0 and 1 for the set of measurements relating to each individual; and a step 300 of identifying abnormal individuals.

Description

[Technical Field]

[0001] The present invention relates to the field of quality control, such as the quality control of electronic components or products from the pharmaceutical industry, for example vaccines. In particular, the invention relates to a method for determining the level of abnormality in an individual, which method makes it possible to statistically detect abnormal individuals, in particular in multivariate situations, i.e. situations in which individuals are characterized by several tests or measurements. [Background technology]

[0002] In this application, the term "individual" refers to any part or batch produced in small, medium or large series, such as by industrial production means.

[0003] In particular, an outlier is an individual that successfully passes all tests but may be non-compliant because it is a statistically abnormal individual. Indeed, experience with electronic components or vaccines shows that even in parts or batches that successfully pass quality control tests, statistical anomalies may reveal potential quality or reliability issues.

[0004] The invention can be advantageously applied in the semiconductor industry sector, which produces integrated circuits, also called "electronic components", which are manufactured on silicon wafer batches, each wafer comprising several hundred electronic components.

[0005] To ensure the operation of these electronic components, a series of tests are performed on each of the components while they are still part of the wafer.

[0006] One or two specification limit(s) are associated with each of these tests.

[0007] Electronic components whose response to at least one test does not comply with the standard for this test, i.e., whose response is outside the specification limits, are considered defective and are rejected when separating the electronic components from the wafer.

[0008] This traditional quality control can be complemented by statistical methods of detecting anomalies in order to minimize quality problems perceived by customers, for example in components intended for the automotive industry. In this case, the key is to reject good parts (which pass all inspections) because they may prove to be statistically abnormal and therefore defective to the customer. This zero-defect goal uses so-called "univariate" or "bivariate" methods.

[0009] For example, a univariate method called PAT (Part Average Testing) compares the response of a test on an electronic component to the average distribution of responses of this test on other electronic components, and considers electronic components whose responses are too far from the response distribution of other electronic components to be abnormal electronic components.

[0010] Generally speaking, this method is not satisfactory in practice, as it often results in a large number of false alarms, since a single value in just one of the p parameters of an individual, more than k standard deviations away from the mean, is enough to be considered abnormal. The more the number of parameters p increases, the more likely it is that the individual is abnormal in at least one test.

[0011] Therefore, using such univariate methods (such as PAT) or bivariate methods (simple regression between one test and another and the detection of outliers based on these regressions) for a large number of tests can lead manufacturers to include many suitable components and eliminate too many, which deprives a small percentage of production of suitable components, while still not guaranteeing to the manufacturer that all potentially defective components have been eliminated.

[0012] In fact, these methods are applied independently to each measurement in a test, and therefore the false alarm rate increases exponentially with the number of measurements. To limit this number of false alarms (and therefore the costs associated with their detection), statistical limits are often chosen quite generously, with the risk of being insufficient.

[0013] Thus, these methods present a risk that in a customer context, electronic components with latent defects that may become apparent during use of the part may be deemed reliable and deliverable to the customer.

[0014] On the one hand, this drawback is inconvenient because, if a potential defect is identified, the manufacturer is forced to send new replacement parts to the customer, lowering the quality level perceived by the customer, but also because some of these components, despite their low unit cost, are critical components in the operation of more complex systems, such as engine controllers or ABS braking systems, where failure of the component can result in serious accidents, the consequences of which far exceed the simple monetary value of the component.

[0015] Therefore, although these methods already feature some achievements, they are not sufficient to achieve zero defects.

[0016] Therefore, to meet this need, several solutions have been proposed using multivariate statistical methods, in particular the principal component analysis method or the Mahalanobis distance and Hotelling's T 2 This is being developed through the implementation of the following:

[0017] However, these solutions only partially meet expectations in industrial manufacturing situations. Indeed, principal component analysis methods are not specifically designed to detect outliers, and therefore often have difficulty accomplishing such a task.

[0018] More specifically, the Mahalanobis distance and Hotelling's T 2is a multivariate distance related to the similarity of individuals and can therefore be used to identify anomalous individuals. However, these methods, assuming high standards for the desired quality, become problematic when anomalies are due to only a subset and not the majority of measurements. In fact, this is the reality in industrial manufacturing situations.

[0019] There are other multivariate methods, such as the one known to those skilled in the art by the English acronym "LOF" which stands for "local outlier factor". However, this method has some major drawbacks: the lack of clear statistical rules for identifying abnormal individuals, the need for optimization of parameters specific to the algorithm (number of neighbors), and the need for increasingly large amounts of computational power as the amount of data increases, which are significant limitations. In particular, these drawbacks make it impossible to use this method in real time on a production line.

[0020] Second, some methods are based on knowledge of the defects to be detected. These are called "supervised" methods, and force manufacturers to recalibrate the method for each new product and / or the addition of inspections in the inspection coverage improvement process.

[0021] Many methods involve assigning an abnormality index to each individual without actually identifying the abnormal individuals. Specifically, these methods sort or classify individuals from most to least abnormal, but cannot guarantee that the initial individuals are actually abnormal.

[0022] Finally, the greatest limitation of these methods is that they are generally no longer applicable when the number of measurements exceeds the number of individuals or when there is a correlation between the variables, as is common in industrial or pharmaceutical situations. Indeed, analytical data are often obtained from a large number of measurements carried out by a certain number of tests, typically hundreds or even thousands of tests, for each electronic component or individual. Summary of the Invention [Problem to be solved by the invention]

[0023] An object of the present invention is to overcome the above-mentioned drawbacks. [Means for solving the problem]

[0024] To this end, the present invention relates to a method for determining the level of abnormality in an individual, which method in particular allows for the statistical detection of abnormal individuals within a previously collected data set, the data set being obtained from measurements of individual parameters carried out by a plurality of measurement systems, preferably the individual being an electronic component.

[0025] The method is: - pre-processing the data; - determining a multivariate anomaly index, said index being transformed from the pre-processed data by a function f so that it is between 0 and 1 for all measurements for each individual; - identifying abnormal individuals; Includes:

[0026] In particular embodiments, the invention also incorporates the following features, implemented either individually or in any technically effective combination:

[0027] In some implementations of the invention, the preprocessing step is performed by a standardization step in which, for each measured parameter j, each individual x(i,j) is centered by the empirical mean μ of the set of values ​​of parameter j and divided by the empirical standard deviation σ.

[0028] In some implementations of the invention, the pre-processing step is performed by a robust standardization step that implements robust statistical measures.

[0029] In some implementations of the present invention, after the determining step, a step of identifying and selecting a subset of individuals having an anomaly index of zero is performed.

[0030] In some implementations of the present invention, during the step of determining the multivariate anomaly index, for each electronic component i, the absolute deviations from the mean for each variable p are summed, with each difference exceeding a reference value k.

[0031] In some implementations of the present invention, the following transformations:

[0032]

number

[0033] is applied to the raw multivariate anomaly index, and Φ is the distribution function of the inverse Gaussian law.

[0034] In some implementations of the invention, the step of determining the multivariate abnormality index comprises the sub-step of determining the abnormality index of individuals forming a subset called the "reference subset" and the sub-step of determining the multivariate abnormality index of a new individual, and the step of identifying abnormal individuals subsequently applies to the new individual the abnormality index of the individuals of the reference subset, and then integrates said new individual into the reference subset if the new individual is deemed not abnormal, or excludes the new individual if the new individual is deemed abnormal.

[0035] In some implementations of the present invention, the step of identifying abnormal individuals comprises: -Determined individual abnormality index z i a substep of sorting in ascending order; - two consecutive anomaly indices z i measuring the deviation between - determining populations that have statistically similar anomaly indices; - identifying abnormal individuals, more specifically populations, based on a defined maximum rate of acceptable abnormal individuals; Includes:

[0036] In some implementations of the present invention, during the step of identifying abnormal individuals, it is determined whether the group of individuals whose individuals have the highest abnormality index contains a number of individuals that is lower than a predetermined maximum threshold of acceptable abnormal individuals as a function of a predetermined maximum rate. If this is the case, proceed to determine successively in descending order a set of groups that includes the first group and the next group(s), such that the sum of the individuals of the set groups is lower than a predetermined threshold.

[0037] In another aspect, the present invention also relates to the use of the method described above for detecting an aberrant vaccine batch in a sample of vaccine batches, which means that the individual is a vaccine batch.

[0038] According to another aspect, the invention also relates to the use of the method described above for detecting anomalous electronic components in a sample of electronic components, which means that the individual is an electronic component.

[0039] According to another aspect, the invention also relates to the use of the method described above for detecting anomalous measurements obtained from data emitted by sensors installed on production or measurement equipment.

[0040] Therefore, in the context of the present invention, the term "individual" also extends to data representing measurements.

[0041] These features advantageously allow the present invention to be integrated into predictive maintenance applications, where anomalies in measurements may be an indication of potential defects in the production or measurement equipment from which the data originates.

[0042] The present invention also relates to a processor, the processor comprising: - preprocessing previously collected data relating to the characteristics of the electronic components obtained from measurements of the parameters of the electronic components performed by a plurality of measurement systems; - determining a multivariate anomaly index, said index being transformed by a function f from the pre-processed data so that it is between 0 and 1 for all measurements relating to each electronic component; -Identifying anomalous electronic components within a sample of electronic components It is configured as follows.

[0043] The processor thus enables identification of anomalous electronic components within a sample of electronic components.

[0044] The invention also relates to a computer program product comprising program code instructions which, when executed by one or more processors, cause the processor(s) to perform the methods described above.

[0045] Advantageously, the computer program product enables a processor to identify anomalous electronic components within a sample of electronic components.

[0046] The invention will be better understood on reading the following description, given by way of non-limiting example and with reference to the drawings in which: [Brief explanation of the drawings]

[0047] [Figure 1] 1 is a flow chart illustrating steps of a method according to the present invention. [Figure 2] 1 is a graph representing a sample of individuals sorted in ascending order according to anomaly index. DETAILED DESCRIPTION OF THE INVENTION

[0048] In these drawings, like reference numerals refer to the same or similar elements relative to one another. Moreover, for clarity, the drawings are not to scale unless otherwise noted.

[0049] The present invention is implemented by computer software, which is executed by a computer calculator such as a processor.

[0050] In particular, the method according to the invention for determining the abnormality level of an individual, which allows for the statistical detection of abnormal individuals, is applied to a data set, which is previously collected and characterized by one or more measurements carried out by a measurement system on a number of individuals.

[0051] In this application, a measurement system may be an electrical tester, a biological measurement system, a physical measurement sensor, or any more general tool adapted to perform measurements.

[0052] For example, this data set can be organized in the form of a data table, where the rows of the data table represent n individuals and the columns of the data table represent p variables corresponding to measurements performed on the individuals. These measurements can be numeric or binary.

[0053] Advantageously, the data set may contain a different number of measurements for each individual without impairing the proper operation of the method according to the invention. Furthermore, the number of measurements may be greater than or equal to the number of individuals, or vice versa, without affecting the operation of the method according to the invention.

[0054] The data set is obtained from a sample of at least three individuals and the number of variables can range from one to several thousand, the only limitation being the computational power of the computer means intended to implement the invention.

[0055] As a non-limiting example of application, the individual may be an electronic component, a vaccine batch, etc. Preferably, the individual is an electronic component.

[0056] Furthermore, measurements may be carried out on parameters or physical properties of industrially produced individuals, for example at the end of production, in the context of quality control of said individuals.

[0057] At the end of production, quality control of electronic components involves several steps. First, a control operation, also called wafer inspection, is performed by a fine needle on a wafer where several electronic components are grouped together. The equipment simulates the operation of the components and measures their physical properties and parameters during operation. The inspection can be performed under various environmental conditions, such as low temperature, ambient temperature, or high temperature.

[0058] Thus, the number of tests performed on a component can vary from 10 to several thousand. A single wafer can contain hundreds to thousands of electronic components.

[0059] Another control action is usually performed after the assembly stage and is called final inspection.

[0060] The physical properties and parameters can be current, voltage, frequency, delay, as well as acceleration or brightness.

[0061] The method aims to identify abnormal electronic components during each of the management operations of said electronic components.

[0062] As shown in the flow chart of FIG. 1, the method sequentially comprises: a step 100 of pre-processing the data; - a step 200 of determining a multivariate anomaly index, said index being transformed by a function f so that it is between 0 and 1 for all measurements relating to each individual; a step 300 for identifying abnormal individuals; Includes:

[0063] Data pre-processing allows data to be obtained independent of the units or scale of measurement of the data.

[0064] In the following sentences, x1, , x n where n is the number of observations characterized by p quantitative variables. More specifically, in the application of the method according to the invention, x i represents all of the measurements performed on part i.

[0065] In this pre-processing step 100, new data y i , j, the initial data x i , we try to standardize j.

[0066] In one embodiment of the present invention, the pre-processing step 100 may be performed by a non-robust standardization step that can be characterized as described below.

[0067] For each measurement parameter j, for each individual x i , j is the empirical mean μ of all values ​​of parameter j j Center by and divide by the empirical standard deviation σj:

[0068]

number

[0069] In another embodiment of the present invention, the pre-processing step 100 may be performed by a robust standardization step.

[0070] More specifically, the robust standardization step may implement a robust statistical measure such as the median or truncated mean as the location parameter, i.e., the empirical mean μ j , and instead of the interquartile range or absolute median deviation for the scale parameter, i.e., instead of the empirical standard deviation σ j in the formulae above, robust statistical measures may be implemented.

[0071] In particular, a non-robust standardization step is preferable to a robust standardization step in the presence of data that do not follow a normal distribution or when these data have rather extreme values: indeed, non-robust estimates such as the mean are sensitive to the presence of extreme values, unlike the median.

[0072] A step 200 of determining a multivariate anomaly index is then performed for all measurements for each individual.

[0073] More specifically, in this step, the pre-processed data y i、j From, we seek to determine the raw multivariate anomaly index for each part or individual i.

[0074] To this end, for each individual i, we add up the absolute deviation from the mean for each variable p for any deviation above a reference value k, where:

[0075]

number

[0076]

number

[0077]

number

[0078] The raw anomaly index with values ​​in [0, 1] undergoes a transformation f that estimates the raw anomaly index between [0, 1]. i This gives the anomaly index z i are comparable to each other. z i =f(exponent i )(Formula 4)

[0079] As an alternative to the previous conversion, the following conversion to the raw anomaly index:

[0080]

number

[0081] It is possible to apply

[0082] where Φ is the inverse Gaussian law distribution function with mean and variance equal to 1.

[0083] Advantageously, the step 200 of determining a multivariate anomaly index allows for the generation of an index that aggregates univariate anomalies, even for low univariate anomalies.

[0084] More specifically, the anomaly index makes it possible to univariately identify, on the one hand, weak anomalies across multiple parameters, and, on the other hand, strong anomalies across one or several parameters, as possible anomalous individuals.

[0085] Note that an individual with an anomaly index of zero does not have any anomalies, and the closer an individual has an anomaly index of one, the more likely the individual is to be anomaly.

[0086] It is understood herein that an individual with a non-zero abnormality index can be either non-abnormal or abnormal.

[0087] Thus, the present invention advantageously makes it possible to detect non-abnormal individuals to the extent that the anomaly index of an individual may be zero. If all individuals have an anomaly index of zero, there are no abnormal individuals within the analyzed sample of individuals.

[0088] Thus, after decision step 200, it is possible to identify and select a subset of individuals that are indeed not abnormal by identifying and selecting individuals with an abnormality index of zero.

[0089] This subset is referred to herein as the "reference subset." The selection of this "best" individuals, which are the best in the sense that they have no abnormalities, is not exactly symmetrical to the procedure that consists in eliminating all abnormal individuals, since the latter procedure statistically assesses the level of abnormality, while the selection of the reference subset goes further by eliminating individuals with slight, but not statistically proven, abnormalities.

[0090] The selection of a reference subset is particularly advantageous as it isolates non-anomalous parts and allows these parts to be used, which is essential in certain industries, for example the space industry or large space programs, where only parts without any anomalies are sent into space.

[0091] In summary, the method according to the invention allows for the discrimination between abnormal and non-abnormal individuals.

[0092] Furthermore, the multivariate anomaly index can be calculated even when there are more variables than observations, unlike most multivariate detection methods.

[0093] This specificity is especially advantageous to the extent that too many variables compared to the number of observations can generate noise that can prevent or complicate the identification of abnormal individuals.

[0094] Finally, these features advantageously allow the method to be applied to data sets containing missing data.

[0095] The normalization of the anomaly index between 0 and 1 at the end of step 200 of determining said index offers great industrial advantages, especially when traceability of parts is not guaranteed during post-production quality control.

[0096] Indeed, thanks to the method according to the invention, in a particular embodiment of the invention, the step 200 of determining the multivariate anomaly index comprises a sub-step of determining the anomaly index for individuals forming a subset, the so-called "reference subset", and a sub-step of determining the anomaly index for new individuals.

[0097] The step 300 of identifying abnormal individuals is then applied to the new individual with respect to the abnormality index of the individuals of the reference subset.

[0098] In this case, the new individual is directly rejected if it is deemed abnormal.

[0099] Alternatively, the step 200 of determining the multivariate anomaly index may comprise the substeps of determining the anomaly indexes of the individuals forming the reference subset and then determining limits for the anomaly indexes of said reference subset based on the anomaly indexes of the individuals, for example these limits being defined according to the standard deviation of the reference subset of the type μ±kσ, where k=3, for example μ and σ being determined according to the values ​​of the reference subset.

[0100] Thereafter, the substep of determining the abnormality index of the new individual is carried out in the same manner as previously described.

[0101] It is thus possible to achieve a "first in, first out" type of rotation, known to those skilled in the art by the acronym "FIFO" which stands for "first in first out".

[0102] Therefore, in industrial production processes, it is possible to continuously and integrally reject any abnormal parts.

[0103] This dynamic implementation mode of the present invention is particularly advantageous in cases where the traceability of the produced individuals cannot be physically guaranteed during production and where it is not possible to remove abnormal individuals detected by conventional methods of detecting abnormal individuals, so-called "post-processing" methods.

[0104] This post-processing method involves statistically identifying abnormal individuals from individual samples, then physically identifying the abnormal individuals and eliminating them. Therefore, this post-processing method requires traceability of each produced individual and processing of the individual sample. The present invention overcomes this constraint and makes it possible to eliminate individuals on the spot.

[0105] Note that thanks to the normalization of the anomaly index, the anomaly index of a new individual observed can be compared to the anomaly index of other individuals, even if the new individual comes from a new batch, which is not necessarily the case with known statistical detection methods.

[0106] The step 300 of identifying abnormal individuals preferably includes the following steps: - As shown in the graph in Figure 2, the determined individual abnormality index z i in ascending order, i.e.: z (1) ≦ ≦ z (n) (Formula 6) - two consecutive anomaly indices z i and the substep of measuring the deviation between: w i =Z (i) -Z (i-1) (Formula 7) - identifying individuals that are abnormal compared to all individuals; The difference w i has the following distribution law: F(w)=1-(1-w) n Following (Equation 8), the average

[0107]

number

[0108] It has.

[0109] This last substep is carried out by identifying populations with statistically similar anomaly indices, according to the deviation between two consecutive anomaly indices.

[0110] In the graph of Figure 2, populations are separated by dashed horizontal lines.

[0111] More specifically, deviations between two consecutive anomaly indices that exceed an α percentile rank are identified:

[0112]

number

[0113] In the embodiment depicted in FIG. 2, the percentile value selected is α=90%.

[0114] These deviations are calculated because, in theory, the difference in deviations follows the distribution law already explained in Equation 7, and the method can be based on a theoretical threshold for distinguishing populations.

[0115] Thus, comparisons between individuals are advantageously performed so as not to identify individuals as abnormal if they are statistically abnormal but not significantly more so than other individuals.

[0116] Thanks to this identification operation, a group of individuals having substantially similar or the same level of abnormality is determined. In this way, all individuals forming the group are considered to be abnormal or non-abnormal. Therefore, the reliability of detecting abnormal individuals is significantly increased to the extent that the method guarantees that two individuals having very similar abnormality indices will both be identified as abnormal or non-abnormal. It is also therefore possible to go beyond the detection of abnormal individuals, since it is possible to generate a subset of individuals having similar abnormality indices.

[0117] Advantageously, therefore, an approach of the method of the invention is to identify abnormal populations.

[0118] A group of individuals having substantially similar or the same level of abnormality is determined, and the abnormality index z of the determined individuals is calculated. i are sorted in ascending order, and the groups are actually sorted in ascending order according to the abnormality level of each individual, as shown in Figure 2.

[0119] The next step in the method is the identification of the abnormal population according to a predetermined maximum percentage of acceptable abnormal individuals.

[0120] In the embodiment depicted in Figure 2, the maximum percentage of acceptable outliers is 10%. In the graph representing 57 individuals, the maximum threshold of acceptable outliers is 5.

[0121] More specifically, during this step, it is determined whether the group of individuals whose individuals have the highest abnormality index, referred to as the "first group", contains a number of individuals below a predetermined maximum threshold of allowable abnormal individuals.

[0122] If this is the case, then all of the individuals in this group are identified as abnormal; otherwise, all of the individuals in this group are considered non-abnormal.

[0123] In the case where an individual in the first group is identified as abnormal, the system proceeds to identify the group whose individual(s) have the next lowest abnormality index, and each successive group, the individuals in the identified group(s) are identified as abnormal as long as the sum of the individuals in that group, or the sum of all of the groups, is below the maximum threshold.

[0124] In other words, if an individual in the first group is identified as abnormal, we want to know whether the individuals in the group whose individual has the next lowest abnormality index to that of the first group are also abnormal. This group is called the "second group."

[0125] In the example depicted in FIG. 2, the first group contains unique individuals, so the search for anomalous individuals is extended to the next group, ie the second group.

[0126] The populations of the first and second groups are then added together.

[0127] If the sum of the individuals in the first and second groups is less than or equal to the maximum threshold, then all of the individuals in the second group are also identified as abnormal, otherwise all of the individuals in the second group are considered non-abnormal.

[0128] In the example represented in Figure 2, the second group contains 3 individuals, the sum of the individuals in the first and second groups is equal to 4, and the search for anomalous individuals is extended to the next group, i.e., the third group. The individuals in the first and second groups are considered anomalous.

[0129] These operations are repeated until a group whose individuals are not abnormal is identified, i.e., until the sum of the individuals of the examined group exceeds a maximum threshold value. From this group, all subsequent groups, i.e., groups having individuals whose abnormality index is lower than that of the examined group, are therefore composed of non-abnormal individuals.

[0130] In the example represented in Figure 2, the sum of the individuals from the first group to the third group is equal to 5, since the third group contains a unique individual. The individuals from the first group to the third group are considered abnormal, and the individuals in the other groups are considered non-abnormal.

[0131] Advantageously, thanks to these operations, the false alarm rate, i.e., the rate of errors in identifying abnormal individuals, is substantially reduced, thereby eliminating the need to impose strict abnormal individual rates that must be met.

[0132] After detection, the abnormal individuals can either be finally identified thanks to physical traceability or directly rejected thanks to the dynamic implementation mode.

[0133] Alternatively, in another implementation mode of the method according to the invention, the step 300 of identifying anomalous individuals may consist in applying statistical methods to the anomaly index in order to detect anomalous individuals.

[0134] For example, these limits can be calculated as limits of the type μ±kσ, where k=3, and μ and σ are determined according to the sample values. Then, if the production process of the individuals allows traceability of the produced individuals, these abnormal individuals are excluded from production.

[0135] In this case, the present invention is able to revert to the univariate case, to the extent that there is only one abnormality index per individual, but by eliminating or significantly limiting false alarms compared to prior art statistical detection methods.

[0136] Next, an example of a processor will be described. - preprocessing previously collected data relating to the characteristics of the electronic components obtained from measurements of the parameters of the electronic components performed by a plurality of measurement systems; - determining a multivariate anomaly index, said index being transformed from the pre-processed data by a function f so that it is between 0 and 1 for all measurements relating to each electronic component; -Identifying anomalous electronic components within a sample of electronic components It is configured as follows.

[0137] The processor may contain several computational cores and have a clock frequency of several gigahertz.

[0138] The processor may be integrated into the computer and connected to a compatible motherboard.

[0139] Also, some processors may be configured to handle all of the calculations necessary to perform the various steps in parallel, allowing for increased computing power and therefore accelerating the completion of the various steps, allowing for more data to be processed, and limiting overheating of the processor.

[0140] An example of a computer program product is described below, comprising program code instructions that, when executed by one or more processors, configure the processor(s) to perform any of the methods in any mode. The program code instructions may be coded in, for example, Python or C++.

[0141] More generally, it should be noted that the implementations and embodiments of the invention considered above are described as non-limiting examples, and that other variations are therefore possible.

Claims

1. 1. A method for identifying abnormal electronic components, the method being implemented by a computer used to detect abnormal electronic components in a sample of electronic components, and in particular enabling statistical detection of abnormal components in a pre-collected data set on parameters or physical properties of the electronic components, the data set being obtained from measurements of the parameters or physical properties of the electronic components performed by a plurality of measurement systems, the method comprising: a step 100 of pre-processing said data; - a step 200 of determining a multivariate anomaly index, in which, from the pre-processed data, the absolute deviations from the mean for each variable p are summed for each electronic component i, each deviation exceeding a reference value k, to determine a raw multivariate anomaly index, and the raw multivariate anomaly index is transformed by a function f to be between 0 and 1 for all measurements on each electronic component, to obtain the multivariate anomaly index; a step 300 of identifying said abnormal electronic component; A method comprising:

2. 2. The method of claim 1, wherein the pre-processing step 100 is performed by a standardization step in which, for each measurement parameter j, each electronic component x(i,j) is centered by the mean μ of the set of values ​​of the parameter j and divided by the standard deviation σ.

3. The method of claim 1 , wherein the pre-processing step 100 is performed by a robust standardization step implementing a robust statistical index.

4. The method according to any one of claims 1 to 3, wherein after said determining step 200, a step of identifying and selecting a subset of electronic components having an anomaly index of zero is performed.

5. The following conversions: [Equation 1] 3. The method of claim 1 , wherein:

6. 6. The method according to claim 1, wherein the step 200 of determining multivariate anomaly indices comprises the sub-steps of determining anomaly indices of electronic components forming a subset called a "reference subset" and determining a multivariate anomaly indices of a new electronic component, and wherein the step 300 of identifying an anomalous electronic component subsequently applies to the new electronic component the anomaly indices of the electronic components of the reference subset, and then integrates the new electronic component into the reference subset if the new electronic component is deemed not anomalous, or excludes the new electronic component if the new electronic component is deemed anomalous.

7. The step 300 of identifying a malfunctioning electronic component comprises: the determined anomaly index z of said electronic component; ｉ a substep of sorting in ascending order; - two consecutive anomaly indices z ｉ measuring the deviation between - separating the selected electronic components into groups based on the measured deviations; - identifying, from the group of separated electronic components, the abnormal electronic components based on a predetermined maximum allowable rate of abnormal electronic components; 7. The method of claim 1, comprising:

8. 8. The method of claim 7, wherein during the step 300 of identifying the abnormal electronic components, it is determined whether the group of electronic components in which the electronic components have the highest abnormality index contains a number of electronic components that is lower than a predetermined maximum threshold of allowable abnormal electronic components as a function of the predetermined maximum rate, and if it is determined that the number of electronic components is lower than the maximum threshold, it proceeds to determine successively in descending order a set of groups comprising a first group and a next group(s), such that the sum of the electronic components of the groups in the set is lower than the maximum threshold.

9. A processor, the processor comprising: - pre-processing previously collected data relating to the properties of electronic components, obtained from measurements of the parameters of said electronic components carried out by a plurality of measurement systems; - determining a multivariate anomaly index, said multivariate anomaly index being determined by summing, for each electronic component i, the absolute deviations from the mean for each variable p, each deviation being above a reference value k, to determine a raw multivariate anomaly index from the pre-processed data, and transforming said raw multivariate anomaly index by a function f to be between 0 and 1 for all measurements for each electronic component; - identifying an abnormal electronic component within a sample of electronic components; a processor configured to:

10. 9. A computer program comprising program code instructions which, when executed by one or more processors, configure the processor(s) to perform the method of any one of claims 1 to 8.

Images