An unsupervised statistical method for multivariate identification of abnormal sensors

Abstract

The present invention relates to a method for identifying abnormal sensors measuring characteristics of an individual, the method comprising: a step 50 of collecting curves of the individual characteristics measured by each sensor; a processing step 100 of calculating, for a given sensor and reference curve, a dissimilarity index between that curve and each of the other curves of that sensor; a first iteration step 200 in which the processing step 100 is repeated iteratively for each curve obtained from the same sensor, thereby determining a dissimilarity index for each curve; a second iteration step 300 of performing steps 100 and 200 for another sensor to obtain a table of dissimilarity indices; a step 400 of calculating an anomaly index for each individual from a multivariate statistical processing of the table; and a step 600 of identifying anomaly individuals. a step 700 of identifying an abnormal sensor; Includes:

Description

[Technical Field]

[0001] The present invention relates to the field of identifying sensors, referred to as anomalous sensors. A sensor is considered anomalous if it produces an anomalous curve.

[0002] In this application, the term "curve" refers to any curve generated by a sensor or anything else. These curves may be time or in other forms, for example spectrum.

[0003] Specifically, an abnormal curve is a curve that is statistically different from other curves. Note that the concept of an abnormal curve is fundamentally important in that it constitutes the condition that determines the method to be implemented. Also, the curve can be in various forms, such as amplitude, phase, shape, type (discrete or continuous), univariate (single sensor) and / or multivariate (multiple sensor) differences.

[0004] The field of application of the invention relates to events monitored by a number of sensors, which measure population characteristics from a universe of events. [Background technology]

[0005] Processing of curves, also called functional data, has been done over the past few years, but the identification of anomalous curves has generally been univariate, i.e., existing methods are performed on a sensor-by-sensor basis and attempt to detect anomalous curves for a given sensor.

[0006] Existing methods are not multivariate-compatible because, even when multiple sensors are used, each sensor is processed separately and correlations between sensors are not considered. However, it is often the case that a curve that is not abnormal from a univariate perspective, i.e., when viewed from only one sensor, has a positive correlation with several sensors, while many of the curves have a negative correlation with those same sensors. Those skilled in the art refer to this situation as the "multivariate-derived anomalous curve phenomenon." It will be appreciated that the ability to identify such situations in this specification is of great interest.

[0007] In the past five years, new multivariate equation detection methods for functional data have been proposed in the literature. They are mainly based on two approaches.

[0008] The most common approach is to "reduce" the curve for a given sensor using a variety of different techniques. More specifically, an initial curve characterized by hundreds or thousands of points is transformed into a set of coefficients using projective methods (B-splines, Fourier equations, wavelets, etc.). A subset of these coefficients is then selected to retain only the relevant signal and remove noise. A subset of such coefficients is obtained for each sensor and compiled. Classical statistical techniques can then be used to detect anomalies in this reduced curve.

[0009] However, this method has many problems. First, the selection of a projection method often depends on the type of functional data (e.g., whether it is periodic or not), and is therefore often based on empirical knowledge. Therefore, it is a method with low versatility. Furthermore, in terms of anomaly detection, coefficient selection is an extremely delicate process. If the threshold setting is too "strict," part of the signal, and therefore information about the abnormal state, may be lost.

[0010] Summarizing the curves using one or more simple statistical indices (e.g., mean, standard deviation) There are simpler ways to do this, but they result in a huge loss of information. On the other hand, if you keep all the coefficients, you don't lose any information. However, you can't remove the noise and you may end up with more coefficients than the curve, which is not desirable in statistics.

[0011] Another approach is to directly calculate an index that measures the anomaly of a curve from the concept of depth relative to the majority of data. The idea is that the larger the depth value of a curve, the more similar its state is to the majority of data. The main drawback of this approach is the complexity of the calculations. That is, as the number of sensors increases, even up to about 10, it becomes difficult to apply to real data sets. Summary of the Invention [Problem to be solved by the invention]

[0012] In view of the above, the present invention relates to a statistical method for anomalous curve detection that overcomes the above-mentioned drawbacks. More specifically, the present invention includes an unsupervised method in the sense that it does not rely on a priori knowledge of the type of distribution and / or the type of signal contained in the sensors. Furthermore, the method is based on a two-stage, i.e., inter-curve, inter-sensor, multivariate analysis of the curve conditions, which makes it unique in that all information from all curves is used in the comparison. [Means for solving the problem]

[0013] To this end, the invention proposes a method for identifying, from a plurality of sensors measuring a set of characteristics of a population from a universe of events, at least one sensor, called an anomalous sensor, implemented by computer software executed by a computing unit, the method comprising: collecting, for each sensor, a data curve, called a curve, which represents the characteristics of each individual, measured by that sensor; a processing step of calculating, for a given sensor and for a curve under consideration, called a "reference curve", an index, called a "dissimilarity index", representing the distance between said reference curve and each of the other curves from said sensor; a first iteration in which the above processing steps are repeated iteratively for each curve obtained from the same sensor, thereby determining a dissimilarity index for each curve; a second iteration for performing the processing steps of the first iteration for other sensors to obtain, for each of said sensors, a table of dissimilarity indices of each of said sensor's curves with respect to the other curves of said sensor; calculating an anomaly index for each individual from a multivariate statistical treatment of the whole or part of the table of dissimilarity indices obtained in the second iteration step; Identifying at least one abnormal individual according to the calculated anomaly index; Identifying at least one abnormal sensor by performing a statistical process according to the calculated dissimilarity index for that sensor and the abnormal individual identified in the previous step; A method comprising:

[0014] Advantageously, the information of the curve is aggregated at points while the difference information between the curve and others is maintained.

[0015] In other words, in accordance with the features of the present invention, each curve is characterized by a single representative value. This advantageously allows for extremely minimal data processing during the identification step, while still preserving all information about the state of the curve. Thus, anomalous curves that would not be detected by conventional methods are detected.

[0016] More specifically, the number of curves for each sensor is reduced to a single real number. The usual multivariate anomaly detection methods can be applied to the corrupted data.

[0017] Furthermore, the method of the present invention allows for more accurate detection of abnormal individuals, due to the aggregated measurements of the individual's characteristics by the sensors.

[0018] This aspect also allows for the prevention or reduction of false alarms.

[0019] The present invention advantageously enables the identification of faulty sensors and anomalous populations in a population of events, i.e., anomalous individuals or individuals having state characteristics that do not match the expected or normal state of the event characteristics.

[0020] The method allows for a reduced number of calculations and the processing of a large number of data and therefore a large number of sensors.

[0021] In particular implementations, the present invention may further include one or more of the following features, either alone or in any technically reasonable combination:

[0022] In a particular implementation, the processing step includes the substeps of: subtracting the reference curve from each of the other curves in turn for the given sensor to obtain a difference curve; squaring the difference curve; adding the resulting curves to obtain a single sum curve; and calculating the square root of the average of the sum curves.

[0023] In a particular implementation, the processing step 100 includes the substep of calculating a correlation coefficient between the reference curve and each of the curves generated by the same sensor, and calculating an average of the coefficients.

[0024] In a particular implementation, the processing step 100 includes the substep of calculating a multivariate dissimilarity index by applying an anomaly detection method to the measurements of the curve against the values ​​of each of the other curves.

[0025] In a particular implementation, before process step 100 above, a preparation step is performed to prepare the data before process 500 is performed, in which the curves are all spaced at the same time such that they share the same time, i.e., they all have the same number of points and are necessarily aligned to the same index.

[0026] In a particular implementation, in the method of the present invention, the data is obtained from sensors integrated into the electronic component manufacturing facility and is indicative of a physical parameter.

[0027] In a particular implementation, in the method according to the invention, said data is obtained from sensors installed in the aircraft for carrying out test flights and indicative of physical parameters.

[0028] In a particular implementation, in the method according to the invention, the data is obtained from a sensor measuring a physiological parameter.

[0029] In a particular implementation of the method according to the invention, the data is spectral data.

[0030] In a particular implementation, the method of the present invention is applied to an image, which is characterized by a matrix of pixels, and the sensor(s) measure the color level of gray, blue, red, or green for each pixel.

[0031] The invention also relates to a computer program product comprising program code instructions which, when executed by one or more processors, configure the processor(s) to implement the above-mentioned method.

[0032] The computer program product advantageously enables one or more processors to identify the sensors that generate the anomaly curves.

[0033] The invention also relates to a computer memory storing the above-mentioned computer program product. The invention will be better understood on reading the following description, given by way of non-limiting example and with reference to the following drawings, in which: [Brief explanation of the drawings]

[0034] [Figure 1] 1 is a flowchart of a method according to the present invention. [Figure 2]1 is a table of three measurement periods performed by the sensor. [Figure 3] 3 is a curve corresponding to the measurement period of FIG. 2. [Figure 4] A table of values ​​indicating the distance between each of the curves and all other curves generated by the same sensor, the table being obtained from the first iteration. [Figure 5] 10 is a table of dissimilarity indices of the curves from the two sensor groups, the table being obtained from the second iteration step. DETAILED DESCRIPTION OF THE INVENTION

[0035] The present invention is realized by computer software that is executed by a computing unit such as a processor of a computer.

[0036] This description is presented as a non-limiting example: each feature of an embodiment can be advantageously combined with any other feature of any other embodiment.

[0037] As mentioned above, the present invention relates to a method for identifying a sensor, referred to as an abnormal sensor, from a plurality of sensors. The steps of the method are shown in FIG.

[0038] The sensors measure features of individuals from a population of events. The features may be physical features such as pressure, temperature, brightness, etc. Thus, each individual is characterized by the features measured by the sensors, which may be different from each other. In other words, each sensor may measure a different feature of the individual.

[0039] An individual can be an individual in the statistical sense. It can be an event or part of it.

[0040] There can be a variety of events, including but not limited to a test flight, a part or system manufacturing process, or monitoring of a given space by a monitoring system.

[0041] Thus, the sensors may include pressure, temperature, humidity, brightness, current, displacement, image, etc. The nature of each sensor may differ, as may the type of data that a measurement technology specializes in measuring. For example, a group of pressure sensors may include strain gauge-based sensors, and / or capacitive pressure sensors, and / or piezoresistive pressure sensors, and / or even resonant pressure sensors.

[0042] The sensors can be analog or digital.

[0043] The data from the sensors can relate to the same or different physical parameters. For example, a first sensor can measure pressure and a second sensor can measure temperature. In a preferred embodiment, the sensors measure different physical parameters. In practice, this can provide more information about the manufacturing facility.

[0044] These measurements are typically performed at one or more measurement points per second. At the end of the measurement period, a data curve, called a curve, is obtained for each sensor. This curve may be a function of time. More specifically, each curve may be a function of an individual characteristic, with the ordinate representing the area specific to the parameter measured by the sensor and the abscissa representing the time domain.

[0045] Sensors, in a non-limiting example of the present invention, are integrated into the physical environment.

[0046] A sensor referred to as an anomalous sensor may mean that the sensor is malfunctioning or has failed, or it may mean that it measures a characteristic of a population in a population of events, and that this characteristic is anomalous in that it is not in an expected state or is not the usual state of the characteristic in the event.

[0047] The anomaly may highlight a system failure that has a significant impact on the characterized environment, or may highlight the occurrence of an unexpected event or malfunction in the environment, such as the appearance of a foreign object during space surveillance.

[0048] An event, or an abnormality that is emphasized in an individual measured by a sensor, is called an anomaly and is considered to be abnormal.

[0049] The first step of the method is to collect 50 for each sensor a curve of the data measured by that sensor, each curve representing a characteristic of an individual.

[0050] For example, Figure 2 shows a table of data obtained from three measurement periods C1, C2 and C3 performed by the sensor, and Figure 3 shows the corresponding curves.

[0051] The data collected by the present invention to characterize the curves may be obtained from sensors installed in equipment manufacturing electronic components such as semiconductors. These sensors may provide measurements of physical parameters such as temperature, pressure, humidity, etc. Therefore, identifying one or more sensors that generated an anomaly curve may highlight a high likelihood of one or more defects in a product being manufactured or being manufactured at the manufacturing facility when the sensor(s) generated the anomaly curve. Thus, the present invention may provide a means for performing predictive maintenance. Furthermore, the method may highlight one or more failed sensors.

[0052] Alternatively, the curves may be obtained from sensors installed on an aircraft to measure hundreds of physical parameters for test flights. In this implementation, the sensors may be installed in or around a system or subsystem of the aircraft. In this application, the sensor curves correspond to measurement periods performed during flight. In this implementation, the method may highlight abnormal conditions in the system or subsystem monitored by the sensor. Furthermore, the method may be capable of highlighting one or more faulty sensors.

[0053] Alternatively, the data may be obtained from sensors intended to measure physiological parameters of the subject. In this case, the sensors may be placed on or near the subject. In this application, the set of sensors may include, without limitation, movement sensors or even sensors capable of generating an electroencephalogram. In this implementation, the method may highlight that the subject may be susceptible to a given disease or disorder. Furthermore, the method may include one or more A failed sensor may be highlighted.

[0054] The family of curves may also represent a family of images. These images may be obtained from a CMOS (Complementary Metal Oxide Semiconductor) sensor, a CCD (Charge Coupled Device), which may be included in a camera such as a thermal camera. In fact, an image may be characterized by a matrix of pixels. Each pixel may be characterized by a gray level in the case of a black and white image, or by a green, blue, and red level in the case of a color image. In the case of a black and white image, a sensor measures the gray level. In the case of a color image, three sensors measure the red level, the green level, and the blue level of each pixel. An individual may be the whole image or a part of it. Preferably, each individual is a row or a column of pixels. Thus, for a black and white image, the measurement curve is the gray level corresponding to all or some of the pixels of the image, and for a color image, the curve is the red, blue, or green level corresponding to all or some of the pixels of the image.

[0055] The method may enable detection of abnormal regions in a hyperspectral image. The hyperspectral image may be defined by a matrix of pixels, each characterized by a wavelength. Each sensor is configured to detect a given wavelength. An individual may be all or a portion of the image. Preferably, each individual is a row or column of pixels. Each curve is a wavelength corresponding to all or a portion of the pixels of the hyperspectral image.

[0056] Detecting anomalous images in a collection of images is an advantageous field of application of the present invention, for example to identify "anomalous" elements that recur over time in video or satellite imagery in an unsupervised manner.

[0057] The method includes a curve processing step 100, in which, for a given sensor, all the curves obtained by that sensor are retrieved in order to calculate an index indicating the distance between that curve and each of the other curves generated by this sensor relative to a curve under consideration, called the "reference curve".

[0058] This step is described below in three embodiments, each of which obtains this index in a different way.

[0059] As will be explained below, this index may take the form of a dissimilarity index.

[0060] Advantageously, this dissimilarity index makes it easier to assess abnormal features of the curves.

[0061] In the first embodiment of the present invention, the curve processing step is performed as follows.

[0062] For a given sensor, a curve is selected as described above. This curve is called the "reference curve." This reference curve is characterized by the data in one of columns C1 to C3 of FIG. 2.

[0063] This reference curve is subtracted point by point from each of the other curves in turn. For example, as shown in Figure 2, if C1 is the reference curve, its data at each time point t is subtracted in turn from the data at the same time point t of curves C2 and C3.

[0064] More specifically, for a given sensor, for a set of N processed curves, N so-called "difference curves" are obtained for each reference curve, each of which is the same size as the reference curve. Each difference curve is then squared, again point-wise.

[0065] The squared difference curves are added together to obtain a single sum curve. The final operation is to The lines are averaged and then the square root is taken. At the end of these operations, an index is obtained that indicates the distance between the reference curve and each of the other curves generated by the sensor. This is shown graphically in Figure 3. This index is referred to herein as the "dissimilarity index."

[0066] 3, it can be seen that the dissimilarity index of curve C3 is abnormally high compared to the dissimilarity indexes of curves C1 and C2. Therefore, it can be deduced by the method of the present invention that C3 is in an abnormal state compared to C1 and C2.

[0067] In summary, process step 100 amounts to calculating the square root of the mean of the sum of the deviation values ​​as the dissimilarity index associated with the curves obtained by the sensors.

[0068] Thus, the method of the present invention condenses information about the differences between the reference curve and a group of other curves into a single value, the dissimilarity index.

[0069] In general, the curve processing step according to the first embodiment of the present invention can be formulated as follows: where N is the number of curves, P is the number of sensors,

[0070]

number

[0071] is the nth curve of sensor p, All curves

[0072]

number

[0073] For , the squared pointwise difference is calculated. (C p,n -C p,k ) 2 (Formula 1) Then the pointwise sum of this is calculated.

[0074]

number

[0075] Finally, we obtain the dissimilarity index by calculating the average of all points on the sum curve.

[0076]

number

[0077] These processing operations are then repeated iteratively for each curve n obtained by the same sensor during a first iteration step 200. In other words, each curve obtained by that sensor becomes a reference curve in turn. Thus, at the end of this step, we have a curve n as shown in FIG. Thus, a dissimilarity index is obtained for each curve.

[0078] In the second embodiment of the present invention, the curve processing step 100 is performed as follows.

[0079] For a given sensor and a given reference curve, the correlation coefficient between the reference curve and each of the other curves obtained with that sensor is calculated, which may be obtained by any method known to those skilled in the art, such as the Bravais-Pearson method or the Spearman coefficient method.

[0080] More specifically, for a given sensor, across a group of N processed curves there will be N-1 correlation coefficients for each reference curve.

[0081] The average of the correlation coefficients is then calculated for each of the curves, resulting in an index indicating the distance between the reference curve and each of the other curves obtained by the sensor, which is referred to herein as the "dissimilarity index."

[0082] Similar to process step 100 described for the first embodiment of the present invention, information regarding the deviation of the reference curve from the other curves is advantageously summarized into a single value.

[0083] In general, the curve processing step 100 according to the second embodiment of the present invention can be formulated as follows: where N is the number of curves, P is the number of sensors,

[0084]

number

[0085] and The correlation coefficients of curve n with each of the N curves are calculated. corr[(C n ,C k )],n≠k (Equation 4) The average correlation between a given parameter p and curve n is calculated. φ=average[corr(C n )] (Equation 5)

[0086] These processing operations are then repeated iteratively for each curve obtained by the same sensor during a first iteration step 200. Thus, at the end of this step, dissimilarity indices are obtained for each curve, as shown in Figures 3 and 4.

[0087] In this third exemplary embodiment, the processing step 100 includes calculating a multivariate dissimilarity index obtained by an anomaly detection method known to those skilled in the art, such as the Mahalanobis distance method, applied to the measurements of one curve against the values ​​of each of the other curves, so that the time points become variables in a statistical sense.

[0088] In general, the curve processing step 100 according to the third embodiment of the present invention can be formulated as follows: where N is the number of curves, P is the number of sensors,

[0089]

number

[0090] is a matrix containing the set of curves for the parameter p, All dissimilarity indices for the N curves are calculated multivariately, so that the time point is a variable in the statistical sense. Φ=(φ1,...,φ N )=f[M p ] (Formula 6)

[0091] This step is repeated for each of the sensor's curves. In particular, steps 100 and 200 are combined in this calculation.

[0092] Thus, at the end of this step, we have a dissimilarity index for each curve, as shown in FIG.

[0093] During a second iteration step 300, the curve processing step 100, which includes the first iteration step 200, is performed on the curves from the other sensors to obtain dissimilarity indices for each of the curves for each of these sensors. This second iteration step is shown in FIG.

[0094] The method also includes a step 400 of calculating an anomaly index for each individual. The anomaly index may be specific to each individual. It may be determined from a multivariate statistical process on all or part of the dissimilarity index table obtained by the second iteration step 300. For example, the multivariate statistical process may include applying multivariate statistical algorithms to non-functional data known to those skilled in the art, such as Mahalanobis distance, Hotelling T 2and the like. These multivariate statistical methods generally require processing the dissimilarity index table obtained by the iterative step 300 to derive one or more values ​​in a format comparable to a threshold. The multivariate statistical methods further require determining a statistical threshold in advance and intend to select values ​​above (or below, respectively) the threshold and consider them abnormal, while values ​​below (or above, respectively) the threshold are considered normal. This threshold can be modified depending on the desired detection sensitivity required. Thus, the threshold can be based on a predetermined maximum rate of acceptable abnormal curves.

[0095] The method comprises the step of identifying abnormal individuals 600 from the population of events. This is done from the calculated abnormality indices, which can be compared with the thresholds.

[0096] The method then comprises a step of identifying at least one abnormal sensor by carrying out a statistical process depending on the dissimilarity index calculated for said sensor and the abnormal individuals identified in the previous step.

[0097] To identify sensors that produce abnormal curves, a statistical procedure is carried out: for each sensor, the distance of the dissimilarity index of the curve characterizing the abnormal individual identified from that sensor may be compared with the average of all the dissimilarity indexes of the other curves characterizing other individuals from this sensor. This calculation may allow the calculation of a distance that may be expressed, for example, as a standard deviation from the average.

[0098] Therefore, the longer the distance, the greater the involvement of the sensor in the abnormality of the individual, whereas the shorter the distance, the less the involvement of the sensor in the abnormality of the individual.

[0099] The curves for the sensors most heavily implicated in the anomaly can be plotted and compared with the curves for non-anomaly individuals from the same sensors.

[0100] A more precise cause of the sensor anomaly can then be highlighted, for example, if the anomaly curve from an anomaly sensor is more than a predetermined threshold, the sensor can be deemed to have failed.

[0101] If a sensor is identified as faulty, it is shut down, either manually or automatically, after which it can be restarted or replaced.

[0102] If the number of abnormal curves from an abnormal sensor is below a given threshold, the individual may be deemed defective.

[0103] In order for the algorithms implemented by the present invention to work properly, in these three embodiments of the present invention, the curves all have the same number of points and are necessarily aligned to the same index, in particular the same time domain, or the same spectrum.

[0104] For this purpose, the method according to the invention preferably comprises a preparation step in which the data are prepared before processing, so that they can be directly used by the algorithm, before carrying out the processing step 100, in which the curves are all separated by the same time so that they share the same time, in other words, in this data preparation step the curves are all associated with each other in a manner known per se to those skilled in the art, for example by the dynamic time warping method.

[0105] Advantageously, the present invention provides predictive maintenance so that problems with manufacturing machinery can be anticipated.

[0106] Furthermore, depending on the field of application, it may be possible to detect abnormal flight behaviour by an aircraft, which may reveal, for example, abnormal behaviour of the aircraft's crew, the failure of one or more sensors, etc.

[0107] In another field, the invention makes it possible to predict the appearance of symptoms such as tremors or the onset of obstructive symptoms in patients suffering from Parkinson's disease. In a similar field, these data may be taken from electroencephalograms or electrocardiograms, where the detection of abnormalities may reveal some pathology.

[0108] An example of a computer program product is now described, which includes program code instructions that, when executed by one or more processors, configure the processor(s) to implement one of the methods of any of the above.

[0109] Further, an example of a computer memory is described, which stores the above-mentioned computer program product, and may be, for example, a USB key, a hard drive, or even the cloud.

[0110] More generally, it should be noted that the implementations and embodiments described above are described as non-limiting examples, and thus other variations are possible.

Claims

1. 1. A method implemented by computer software executed by a computing device for identifying at least one sensor, referred to as an anomalous sensor, from a plurality of sensors measuring a set of characteristics of a population from a population of events, the method comprising: a step 50 of collecting, for each of said sensors, data curves, called curves, each characteristic of an individual measured by said sensor, said data curves corresponding to a number of individuals of said population; a processing step 100 of calculating, for a given sensor and for a curve under consideration, called the "reference curve", an index called the "dissimilarity index" that represents the distance between said reference curve and each of the other curves from said sensor; a first iteration 200 in which the processing step 100 is repeated iteratively for each of the curves obtained from the same sensor, thereby determining a dissimilarity index for each of the curves; a second iteration 300 of performing steps 100 and 200 for another sensor to obtain, for each of said sensors, a table of dissimilarity indices of each of said curves of said sensor with respect to the other curves of said sensor; a step 400 of calculating an anomaly index for each of said individuals from a multivariate statistical treatment of all or part of said table of dissimilarity indices obtained by said second iteration step 300; - a step 600 of identifying at least one abnormal individual depending on said calculated abnormality index; a step 700 of identifying at least one abnormal sensor by performing a statistical process depending on the dissimilarity index calculated for said sensor and the abnormal individuals identified in the previous step; A method comprising:

2. The processing step 100 includes: a. determining a difference curve for the given sensor by sequentially subtracting the reference curve from each of the other curves; b) squaring the difference curve; c. adding the resulting curves together to form a single sum curve; d. determining the dissimilarity index as equal to the square root of the mean of the summation curve; The method of claim 1 , comprising:

3. 2. The method of claim 1, wherein said processing step 100 includes the substep of calculating correlation coefficients between said reference curve and each of said curves generated by the same sensor, and calculating an average of said coefficients.

4. The method of claim 1 , wherein the processing step 100 includes the substep of calculating a multivariate dissimilarity index by applying an anomaly detection method to the measurements of the curve relative to the values ​​of each of the other curves.

5. 5. A method according to any one of claims 1 to 4, wherein before the process step 100, a preparation step is carried out to prepare the data before the process 500 is carried out, in which the curves are all separated by the same time so that they share the same time, i.e., they all have the same number of points and are necessarily aligned to the same index.

6. 6. The method of claim 1, wherein the data is obtained from sensors installed in an electronics manufacturing facility and is indicative of physical parameters.

7. 6. The method of claim 1, wherein the data is obtained from sensors installed in an aircraft for carrying out test flights and is indicative of physical parameters.

8. The method of claim 1 , wherein the data is obtained from a sensor that measures a physiological parameter.

9. The method of claim 1 , wherein the data is obtained from a sensor that generates spectral data.

10. 6. The method of claim 1, applied to an image, the image being characterized by a matrix of pixels, and the sensor(s) measuring the color level of gray, blue, red, or green for each of the pixels.

11. 6. The method of claim 1, applied to a hyperspectral image, wherein the image is characterized by a matrix of pixels, each of the pixels being characterized by a wavelength, and each of the sensors detecting a given wavelength.

12. A computer program product comprising program code instructions which, when executed by one or more processors, configure the processor(s) to implement a method according to any one of claims 1 to 10.

13. A computer memory storing the computer program product of claim 11.

Images