Method for detecting a defect in a structure using high-frequency mechanical waves

Abstract

Method for detecting a defect in a structure using high-frequency mechanical waves. This process comprises: - the acquisition (216) of first signals S1,i measured by a first sensor and of second signals S2,i measured, in parallel, by a second sensor, then - the identification (222), by a statistical analysis method, in each signal S1,i, of the variations of this signal S1,i that are correlated with variations of the signal S2,i and, in each signal S2,i, of the variations of this signal S2,i that are correlated with variations of the signal S1,i, then - the elimination (230), in each of the signals S1,i and S2,i, of the variations of this signal identified as being correlated with variations of the other signal to obtain signals Sc1,i and Sc2,i corrected for environmental variations common to the first and second sensors, then - the detection (250) of faults from the corrected signals Sc1,i and Sc2,i. Fig. 2

Description

Title of the invention: Method for detecting a defect in a structure using high-frequency mechanical waves

[0001] The invention relates to a method and a system for detecting a defect in a structure using high-frequency mechanical waves.

[0002] The invention applies in particular, but not exclusively, to the field of non-destructive testing and structural health monitoring.

[0003] Known methods for detecting a defect in a structure use sensors, fixed to the structure, which measure ultrasonic waves. Then, during the operation of the structure, a defect is detected if the ultrasonic wave measured by one of these sensors deviates significantly from a reference ultrasonic wave measured in the absence of a defect. The reference ultrasonic wave is measured, for example, when the structure is put into service at a time when no defect is present.

[0004] During the operation of the structure, the environmental conditions under which it is operated may vary. For example, if the structure is a portion of an aircraft fuselage, the temperature, pressure, humidity, sensor wear, and other environmental conditions may vary significantly during its operation. Thus, during the operation of the structure, sensor measurements are generally taken under environmental conditions that differ from those under which the reference ultrasonic waves were measured. However, variations in environmental conditions alter sensor measurements even in the absence of a defect in the structure. Therefore, if no action is taken, variations in environmental conditions can cause numerous false detections of a defect and also mask the occurrence of a defect.

[0005] To solve this problem, methods have been developed to compensate for the effect of variations in environmental conditions on sensor measurements. Many of these methods focus solely on compensating for the effects of temperature variations and use restrictive assumptions to operate, thus limiting their scope. An example of such a method is described in the following article: CROXFORD, Anthony et al.: “Efficient temperature compensation strategies for guided wave structural health monitoring,” Ultrasonics, 2010, vol. 50, no. 4-5, pp. 517-528. Other methods are based on learning the effects of variations in environmental conditions on sensor measurements. These methods are complex because many parameters must be considered over wide ranges, and The effect of certain parameters, such as sensor aging, is difficult to simulate or predict. Such a method using machine learning is described in the following article: LE BOURDAIS Florian et al.: “Machine-learning based temperature compensation for guided wave imaging in structural health monitoring”, Proceedings of the 1st International Symposium on NDT in aerospace, Paris-Saclay, 2019.

[0006] Finally, other imaging methods are insensitive to these effects either through a calibration process or through a statistical analysis of the simultaneously acquired signals. In both cases, a large number of sensors is required to statistically extract data from the measurements that are independent of variations in environmental conditions. Examples of these methods can be found in the following articles:

[0007] - DRUET, Tom et al. : “Autocalibration method for guided wave tomography with undersampled data”, Wave Motion, 2019, vol. 89, p. 265-283, and

[0008] - MESNIL, Olivier et al. : “Self-referenced robust guided wave based defect detection: Application to woven composite parts of complex shape”, Mechanical Systems and Signal Processing, 2023, vol. 188, p. 109948.

[0009] The invention aims to provide a method for detecting a defect in a structure using high-frequency mechanical waves in which the effects of variations in environmental conditions are compensated without necessarily using a large number of sensors.

[0010] The invention therefore relates to a method for detecting a defect in a structure using high-frequency mechanical waves, this method comprising the following steps:

[0011] - the instrumentation of the structure by fixing on this structure a first and a second high-frequency mechanical wave sensors in locations where:

[0012] - the first and second sensors are both sensitive to defects likely to appear in the structure,

[0013] - the variations in the signals measured by the first and second sensors which are caused by a defect in the structure, are not correlated with each other, and

[0014] - the first and second sensors are subjected to the same variations in conditions environmental factors that could alter the signals measured by these first and second sensors,

[0015] - during the operation of the structure, the acquisition, by an electronic computer tronic, during several successive measurement periods, of first signals Sij measured by the first sensor and of second signals S2> measured, in parallel, by the second sensor, each measured signal Skji being made up of a temporal sequence of P samples Sk,ij, where:

[0016] - the index i is a sequence number that identifies the measurement period during in which the Skji signal was measured,

[0017] - the index k is an identifier of the sensor that measured this time sequence of P samples, this index k being equal to one to identify the first sensor and to two to identify the second sensor, and

[0018] - the index j is a sequence number that identifies the position of the sample within of a time sequence of P samples, then

[0019] - the construction of Scu and Sc2>i signals corrected for environmental variations common to the first and second sensors, each signal Sck>i being made up of a temporal sequence of P samples Sck>ijj,

[0020] - fault detection from the corrected Scu and Sc2>i signals,

[0021] wherein the construction of the Scu and Sc2>i signals comprises:

[0022] - the identification, by a statistical analysis method, in each Ski signal, of the variations of this Ski signal that are correlated with variations of the S2ji signal and, within each S2ji signal, variations of this S2ji signal that are correlated with variations of the Sij signal, then

[0023] - the elimination, in each signal Sij, of the variations of this signal Sij identified as being correlated to variations of the S2ji signal and, in each S2ji signal, variations of this S2ji signal identified as being correlated to variations of the Sij signal, to obtain the Scij and Sc2ji signals.

[0024] Embodiments of this process may include one or more of the following features:

[0025] 1) The statistical analysis method implemented is the mCCA (multiset) method Canonical Correlation Analysis).

[0026] 2) The identification, in each Skji signal, of the variations of this Skji signal which are correlated with variations in the Sk ji signal, includes for k equal to one and k' equal to two and for k equal to two and k' equal to one:

[0027] - the projection of a measure matrix Sk from a measure space to a latent space to obtain a matrix ek containing M rows and Q columns, where:

[0028] - the matrix Sk is a matrix of M rows and P columns containing in each cell Sk[i,j] the sample Skjij,

[0029] - M is the number of measurement periods,

[0030] - Q is a predefined positive integer less than P,

[0031] - the matrix ek is the matrix which contains, in columns, the Q canonical variables ek>m of the mCCA method, where m is an integer index that varies from 1 to Q, the correlation between each canonical variable ek>m and the canonical variable ek >m being given by the canonical correlation pkjk >m of the mCCA method,

[0032] - the selection in the matrix ek of a limited number of canonical variables ek>m for in which the canonical correlations pk>k->m are the highest, the canonical variables ek>m thus selected identifying the variations of the Skji signals which are correlated with variations of the Sk ji signal.

[0033] 3) The elimination, in each Skji signal, of the variations of this Skji signal identified as being correlated with variations in the Sk ji signal, includes for k equal to one and for k equal to two:

[0034] - the replacement in the matrix ek of each canonical variable ek>m that has not been selected by a column of zeros to obtain a modified eek matrix, then

[0035] - the projection of the modified matrix eek into the measure space by applying the inverse projection of the projection used to project the Sk matrix into latent space and thus obtain a Sek matrix of measures, then

[0036] - the subtraction of the matrix Sek from the matrix Sk to obtain a matrix Sck containing in each row i of this matrix Sck the P samples of the corrected signal Sck>i.

[0037] 4) The elimination, in each Skji signal, of the variations of this Skji signal identified as being correlated with variations in the Sk ji signal, includes for k equal to one and for k equal to two:

[0038] - the replacement, in the matrix ek, of each canonical variable ek>m that has been se selected, by a column of zeros to obtain a modified eck matrix, then

[0039] - the projection of the modified matrix eck into the measure space by applying the inverse projection of the projection used to project the Sk matrix into the latent space and thus obtain a Sck matrix of measurements containing in each row i of this matrix the P samples of the corrected signal Sck>i.

[0040] 5) The process also includes:

[0041] - for each sensor Ck, the construction of Scck>i signals corrected for variations in periodic environmental signals of predefined period T from the corrected signals Sck>i, the construction of these signals Scck>i comprising:

[0042] - the determination, from a Sck matrix of measures and by implementing a dimensionality reduction method, of a Vk matrix of R principal components, where:

[0043] - the Sck matrix is ​​a matrix of M rows and P columns containing, in each cell Sck[i,j], the sample Sckjijj of the signal Sck>i,

[0044] - M is the number of measurement periods,

[0045] - R is a predefined positive integer less than P, then

[0046] - the projection, using the matrix Vk, of the matrix Sck of measures from a space of measures to a latent space to obtain a matrix yk containing M rows and R columns, then

[0047] - filtering the columns of the matrix yk to obtain a filtered matrix yfk, the filtering each column, eliminating from that column periodic frequency variations equal to 1 / T,

[0048] - the projection of the filtered yfk matrix into measure space by applying the inverse projection of the projection used to project the Sck matrix into the latent space and thus obtain a Scck matrix of measurements containing in each row i of this matrix the P samples of the corrected signal Scck>i, then

[0049] - fault detection from the corrected signals Scu and Sc2>i, includes the fault detection from corrected Scck>i signals.

[0050] 6) The filtering of the columns of the matrix yk involves, for each column of this matrix yk:

[0051] - the construction of the power spectrum of the values ​​contained in this column, Then

[0052] - the replacement of this column by a column of zeros when the power The cumulative power of the frequencies that, in this constructed power spectrum, are above a threshold ST, is xb times greater than the cumulative power of the frequencies that, in this constructed power spectrum, are below this threshold ST, where:

[0053] - the ST threshold is between 0.3*T and T, and

[0054] - xb is greater than one.

[0055] 7) The dimensionality reduction method implemented is the method principal component analysis.

[0056] 8) The process comprises:

[0057] - the measurement by the first and second sensors, respectively, of a first signal If a reference signal and a second reference signal S2jref are present in the absence of defects in the structure, each reference signal being made up of a time sequence of P samples Skjrefj, then

[0058] - recording the Si>ref and S2>ref signals, then

[0059] - during the fault detection step, a fault is detected from the deviations between a signal obtained from the corrected signal Scu and the recorded reference signal Si>ref and, in parallel, from the differences between a signal obtained from the corrected signal Sc2>i and the pre-recorded reference signal S2>ref.

[0060] The invention also relates to an information storage medium, readable by a microprocessor, comprising instructions executable by this microprocessor, in which this medium includes non-transient instructions for the execution of the above detection method, when these instructions are executed by the microprocessor.

[0061] The invention also relates to a system for detecting a defect in a structure using high-frequency mechanical waves, this system comprising:

[0062] - a first and a second high-frequency mechanical wave sensors fixed on the structure in locations where:

[0063] - the first and second sensors are both sensitive to defects likely to appear in the structure,

[0064] - the variations in the signals measured by the first and second sensors which are caused by a defect in the structure, are not correlated with each other, and

[0065] - the first and second sensors are subjected to the same variations in conditions environmental factors that could alter the signals measured by these first and second sensors,

[0066] - an electronic computer (30) programmed to execute the following steps:

[0067] - during the operation of the structure, acquire, over several periods of successive measurements of the first signals Su measured by the first sensor and the second signals S2ji measured, in parallel, by the second sensor, each measured signal S k>i being made up of a temporal sequence of P samples Skjij, where:

[0068] - the index i is a sequence number that identifies the measurement period during in which the Skji signal was measured,

[0069] - the index k is an identifier of the sensor that measured this time sequence of P samples, this index k being equal to one to identify the first sensor and to two to identify the second sensor, and

[0070] - the index j is a sequence number that identifies the position of the sample within of a time sequence of P samples, then

[0071] - construct Scu and Sc2>i signals corrected for environmental variations common to the first and second sensors, each signal Sck>i being made up of a temporal sequence of P samples Sck>ijj,

[0072] - detect faults from the corrected Scu and Sc2>i signals,

[0073] wherein the electronic computer is also configured to, during the construction of the Scu and Sc2>i signals:

[0074] - to identify, by a statistical analysis method, in each signal S^, va variations of this Ski signal that are correlated with variations of the S2ji signal and, in each S2ji signal, variations of this S2ji signal that are correlated with variations of the S signal then

[0075] - eliminate, in each signal Su, the variations of this signal Si identified as being correlated to variations of the S2ji signal and, in each S2ji signal, the variations of this S2ji signal identified as being correlated to variations of the signal, to obtain the Scij and Sc2ji signals.

[0076] The invention will be better understood upon reading the following description, given solely by way of non-limiting example and made with reference to the drawings in which:

[0077] - Fig. 1 is a schematic illustration of the architecture of an ins- structure instrumented including a fault detection system in this structure,

[0078] - [Fig.2] is a flowchart of a first operating method of the instrumented structure of the [Fig.1],

[0079] - Figures 3 and 4 are graphs illustrating moments when defects are detected in two different cases,

[0080] - [Fig. 5] is a flowchart of a second operating method of the instrumented structure of the [Fig.l].

[0081] In this description, the terminology, conventions, and definitions of the terms used in this text are introduced in Chapter I. Detailed examples of embodiments are then described in Chapter II with reference to the figures. Variants of these embodiments are presented in Chapter III. Finally, the advantages of the different embodiments are specified in Chapter IV.

[0082] Chapter I: Definitions, terminology and conventions:

[0083] In the figures, the same references are used to designate the same elements.

[0084] In the remainder of this description, the well-known characteristics and functions of a person skilled in the art are not described in detail.

[0085] The symbol “*” denotes scalar multiplication.

[0086] The symbol “.” denotes vector or matrix multiplication.

[0087] The symbol “T” denotes the transposed operation.

[0088] The inverse of a matrix A, denoted A4, is the matrix such that AA 1 = I, where I is the identity matrix.

[0089] The correlation of two variables X and Y is defined by the following relation: Cor(X,Y) = Cov(X,Y) / [Var(X)*Var(Y)]0'5, where:

[0090] - Cor(X,Y) is the correlation between variables X and Y,

[0091] - Cov(X,Y) is the covariance between the variables X and Y,

[0092] - Var(X) is the variance of the variable X, and

[0093] - Var(Y) is the variance of the variable Y.

[0094] The term "high-frequency mechanical wave" or simply "mechanical wave" refers to a mechanical wave whose fundamental frequency is greater than 1 kHz and, preferably, greater than 10 kHz. Typically, it is an acoustic wave.

[0095] An ultrasonic wave is an acoustic wave whose frequency is greater than 20 kHz.

[0096] The expression "variations in environmental conditions" refers to any variation in the external environment of the structure that is likely to affect the measurement of the sensors, regardless of whether or not there is a defect in that structure. By way of example, the following events are considered as variations in environmental conditions:

[0097] - a variation in the temperature of the external environment in which the structure is immersed,

[0098] - a variation in the pressure of this external environment,

[0099] - a variation in the humidity of the external environment,

[0100] - the aging of sensors over time.

[0101] Chapter II: Example of an embodiment

[0102] Figure 1 represents an instrumented structure 2 comprising:

[0103] - a structure 6, and

[0104] - a system 8 for detecting a defect in the structure 6.

[0105] Structure 6 is a structure in which a defect, capable of altering the propagation of ultrasonic waves in that structure, may appear. The defect detectable using the system described here is, for example, a crack or a microcrack. It may also be a defect such as a trace of corrosion or a local modification of the structure's porosity.

[0106] Structure 6 is a mechanical part. By way of illustration, structure 6 is a thin structure. For example, here, the thin structure 6 has an outer face and an inner face separated from each other by the thickness e6 of the thin structure 6. The thickness e6 is small enough that the outer and inner faces guide the propagation of an elastic wave or a Lamb wave in the thin structure in directions parallel to these outer and inner faces. For this purpose, typically, the thickness e6 is ten or one hundred times smaller than a length and / or a width of the thin structure 6. Here, the thin structure 6 is a composite panel constituting the fuselage of an aircraft.

[0107] For example, the thin structure 6 is made of laminated composite materials, that is, by stacking, in a direction perpendicular to the outer face, a succession of layers, each made of a respective material. As explained in Chapter III on variants, the lessons learned in this particular case can be easily applied to many other possible structures.

[0108] For the sake of simplicity in [Fig. 1], the thin structure 6 is represented as a simple rectangle. However, in reality, the shape of the thin structure 6 is more complex. In particular, in the specific case of a composite panel of an aircraft fuselage, the thin structure 6 typically has rounded curves.

[0109] System 8 makes it possible, in particular, to detect the appearance of a defect by measuring the ultrasonic waves propagating in the structure 6. To this end, system 8 comprises:

[0110] - Ck sensors each capable of measuring ultrasonic waves propagating in structure 6,

[0111] - a transmitter 12 capable of emitting a predefined ultrasonic wave which propagates in structure 6, and

[0112] - a monitoring unit 14 which monitors the occurrence of a defect in the structure 6 from the measurements of the Ck sensors.

[0113] The index k is an identifier of the sensor Ck. Here, system 8 is described in the specific case where it comprises only two sensors Ci and C2. Thus, in this particular case, the index k takes either the value one or the value two. In this text, the reference Ck therefore designates both sensor Ci and sensor C2.

[0114] For example, here, each sensor Ck is a piezoelectric sensor fixed without any degrees of freedom to structure 6. More precisely, the sensors Ck are arranged at respective locations on structure 6 where they are simultaneously subjected to the same variations in environmental conditions. However, these locations are chosen so that the traces generated by a defect in the signal measured by sensor Ci are not correlated with the traces generated by the same defect in the signal measured by sensor C2. In general, this condition is satisfied as soon as sensors Ci and C2 are located at sufficiently distant locations. For example, sensors Ci and C2 are separated by a distance greater than 10 cm, 30 cm, or 1 m.

[0115] The transmitter 12 is controlled by the unit 14 to emit, typically at regular intervals, a predefined ultrasonic wave that propagates through the structure 6 until it reaches each of the sensors Ck. The ultrasonic wave generated by the transmitter 12 is typically an elastic wave. In the case of a thin structure, this elastic wave is, for example, a Lamb wave. For example, the transmitter 12 is a piezoelectric actuator fixed, without any degrees of freedom, to a face of the structure 6.

[0116] The monitoring unit 14 acquires measurements from each of the sensors Ck and, based on these measurements, detects the occurrence of a fault if such a fault appears in the structure 6. Thus, the unit 14 makes it possible to monitor the health status of the structure 6 and to inform a maintenance operator. To this end, the unit 14 comprises an electronic computer 30 and a human-machine interface 32 connected to the computer 30.

[0117] The calculator 30 includes a programmable microprocessor 34 and a memory 36. The memory 36 contains the instructions and data necessary for the execution of the process of [Fig.2] or 5, when these instructions are executed by the microprocessor 34.

[0118] The human / machine interface 32 is capable of communicating, in a way directly intelligible to a human being, the results of the implementation of the detection process of [Fig.2] or 5. For example, the interface 32 includes a screen.

[0119] The operation of system 8 will now be described with reference to the process in [Fig.2],

[0120] During an instrumentation step 200, the Ck sensors are fixed on the structure 6 each at a respective location.

[0121] Then, in step 202, a reference signal Skjref is recorded for each of the sensors Ck. For example, in the absence of a fault in structure 6, an ultrasonic wave is measured by each of the sensors Ck. It is this ultrasonic wave, measured in the absence of a fault using sensor Ck, that is then recorded in memory 36 as the reference signal Skjref associated with that sensor Ck. In this step 202, each Skjref signal is measured in the same way as during the operating phase of the fault detection system 8.

[0122] Next, a phase 210 of the fault detection system 8 operation is executed. During this phase 210, M measurement periods Pm are repeated, for example, at regular intervals. M is the total number of Pm periods executed. Generally, M is greater than ten, fifty, or one hundred. Here, the index i is a sequence number that identifies the Pm period among the set of M measurement periods. The index i ranges from 1 to M. In general, the durations of the Pm periods are all identical. Hereafter, the duration of a Pm period is denoted Dp.

[0123] Each period Pm; begins with a step 212 during which the computer 30 commands the transmitter 12 to emit a predefined ultrasonic wave into the structure 6.

[0124] In parallel, during a step 214, each of the sensors Ck measures the ultrasonic wave emitted by the emitter 12 after it has propagated through the structure 6 to the location of that sensor Ck. During step 214, each sensor Ck generates, in response, an analog measured signal Skji(t).

[0125] In parallel with step 214, during step 216, the computer 30 acquires, over the entire duration Dp of the period Pm, the signals measured in parallel by each of the sensors Ck. During this step 216, the computer 30 converts each measured analog signal Skji(t) into a measured digital signal Skji. Here, the sampling frequency fe of each measured analog signal is constant and predetermined. Thus, each Skji signal is a digital signal that takes the form of a temporal succession of P samples Skjij, where:

[0126] - P is equal to Dp*fe, and

[0127] - the index j is the order number of the sample Skjij in the temporal succession of P samples.

[0128] Each sample is a numerical value obtained during the digitization of the analog signal measured by the sensor Ck.

[0129] At the end of the M periods Pm;, during a step 220, the computer 30 constructs Sckji signals corrected for environmental variations common to sensors Ci and C 2-

[0130] Subsequently, the various steps and operations are described in the specific case of the processing applied to the S14 signals. However, all the processing applied to the signals is also applied to the S2ji signals because the C2 sensor, like the Ci sensor, is used to detect faults. More precisely, a description of the processing of the S2>i signals is obtained by reversing the indices 1 and 2 in the following description.

[0131] Step 220 begins with an operation 222 of identification, by a statistical analysis method, in each signal S1, of the variations of this signal that are correlated with variations of the signal S21. Here, the statistical analysis method used is the multiset canonical correlation analysis method, better known by the acronym mCCA (multiset Canonical Correlation Analysis). In this particular embodiment where only two sensors C1 and C2 are used, the mCCA method is also simply known by the acronym CCA (Canonical Correlation Analysis).

[0132] A description of the general principles of this mCCA method can be found in the following articles:

[0133] - KETTENRING, Jon R.: “Canonical analysis of several sets of variables”. Biometrika, 1971, vol. 58, no. 3, p. 433-451,

[0134] - TENENHAUS, Arthur and TENENHAUS, Michel: “Regularized generalized "Canonical correlation analysis." Psychometrika, 2011, vol. 76, p. 257-284.

[0135] - Jocob A. Wegelin: “A survey of Partial Least Squares (PLS) Methods, with Emphasis on the Two-Block Case”, University of Washington, Seattle, Washington, 98195 USA, Department of Statistics, Technical Report No. 371, March 2000.

[0136] Further descriptions of the principles of the mCCA method can be obtained from the following links: https: / / scikit-learn.Org / stable / modules / cross_decomposition.html#cross-decomposition and https: / / online.stat.psu.edu / stat505 / lesson / 13 / 13.1.

[0137] In the remainder of this description, the article by Kettering Jon R is referred to by the reference Kettering 1971.

[0138] The general principles of the mCCA method are not explained hereafter. Only the application of the mCCA method to Ck sensor measurements is described. Furthermore, the terminology used hereafter corresponds to that of the mCCA method.

[0139] During operation 222, the signals from sensor Ci are grouped into a measurement matrix Si. Similarly, the signals S2ji from sensor C2 are grouped into a measurement matrix S2. The matrix Sk is the measurement matrix of a sensor C k. The matrix Sk is a matrix of M rows and P columns containing, in each cell Sk[i,j], the sample Skjij. Thus, the i-th row of the matrix Sk contains all the samples that make up the signal Skji.

[0140] Subsequently, the following notations are used and their correspondences with the The mCCA method is explained:

[0141] Sk[j] denotes a variable whose different measured values ​​are the values ​​contained in the j-th column of the matrix Sk. The variables Sk[j] correspond to the variables, in the space of measurements, of the mCCA method.

[0142] wk>m is the m-th canonical vector of the mCCA method, where m is an order number that uniquely identifies this canonical vector. The index m is an integer ranging from 1 to Q, where Q is a predetermined integer less than P. The vector wk>m has dimension P. The vector wk>m has P coefficients wk>m>i to wk>m>P. The vector wk>m is therefore defined by the following relation: wk>m = [wk.mi, ..., wk>mj, ..., wkjm>P]T. The coefficients wk>m>i to wk>m>P are commonly called "weights".

[0143] Wk is the matrix that contains, in order of increasing index m, the Q canonical vectors wk>m. The matrix Wk therefore has P rows and Q columns. Each cell Wk[j,m] of the matrix Wk contains the coefficient wk>m>j.

[0144] ek>m is the m-th canonical variable. The canonical variable ek>m is a vector of dimension M defined by the following relation: ek>m = Sk.wk>m. The canonical variable ek>m is therefore representative of the following linear combination of the variables Sk[j]: ek>m = w 1] + ...+ Wk>mj:|'Sk[j]+ ... + W^mæ^SlcfF].

[0145] The matrix ek is the matrix corresponding to the projection of the matrix Sk into the latent space of the mCCA method. The matrix ek is therefore defined by the following relation: ek = Sk.Wk. The latent space is a space of dimensions lower than the dimensions of the measure space. Indeed, the number Q of columns of the matrix ek is less than the number P of columns of the matrix Sk.

[0146] pkjk jm is the canonical correlation between the canonical variables ek>m and ek >m, where k' equals two when k equals one and k' equals one when k equals two. The value of the canonical correlation pkjk >m is greater the more strongly the canonical variables ek>m and ek >m are correlated. The canonical correlation pkjk >m is normalized and ranges from zero to one.

[0147] In the mCCA method, the weights of the first-order canonical vector, i.e., the weights of the vectors wk.i, are determined to maximize the canonical correlation pk>k-4 for the pair (k, k') of sensors Ck. Then, the weights of the higher-order canonical vectors wk>m are determined, in increasing order of index m, each time to satisfy the following two constraints:

[0148] - Constraint 1): Maximize the canonical correlation pkjk >m for the pair (k, k') of CK sensors, and

[0149] - Constraint 2): for each canonical variable ek>m, the correlation between this canonical variable ek>m and any one of the canonical variables ek>m_i is zero or minimal.

[0150] Thus, the canonical correlation pk,k',m decreases as the index m increases. In other words, the canonical variables ek>i and ek >i are the two most strongly correlated canonical variables. Next, the canonical variables ek>2 and ek >2 are less strongly correlated, and so on. It follows that, in the matrix ek, the first column, which contains the canonical variable Ek>1, is the one with the strongest correlation pk>k-4 with the canonical variable Ek-4.

[0151] Here, during a sub-operation 224, the computer 30 automatically determines the matrices Wi and Ei, as well as the values ​​of the canonical correlations pi>2>m, from the measure matrices Si and S2, implementing the mCCA method. Thus, during this sub-operation 224, the computer 30 projects the matrix Si into a latent space by multiplying it by the matrix Wi to obtain the matrix Eb

[0152] In this embodiment, sensors Ci and C2 are subjected to the same variations in environmental conditions. Therefore, a variation in these environmental conditions similarly modifies the Su and S2ji signals. Conversely, the appearance of a defect in the structure does not cause the Su and S2ji signals to vary in the same way. Indeed, because sensors Ci and C2 are fixed to the structure 6 at different locations, the presence of a defect that modifies the Su signal leads to different modifications of the S2ji signal, or simply to modifications of the S2ji signal that are not perceptible. Under these conditions, the variations in the Su signals that are strongly correlated with variations in the S2ji signals are caused by variations in environmental conditions and not by the appearance of a defect.Thus, in the matrix Eb, these variations in environmental conditions correspond to the highest-order canonical variable(s) Ei>m. Conversely, in the matrix Eb, the variations in signals Su that are not caused by these variations in environmental conditions correspond to the higher-order canonical variables Ei>m.

[0153] Therefore, here, during a sub-operation 226, the computer 30 automatically selects, from the matrix Eb, a limited number of canonical variables Ei m for which the canonical correlations pi>2>m are the highest. The canonical variables Ei>m thus selected identify the variations of the signals that are correlated with variations of the signal S2ji. To this end, in this embodiment, the computer 30 uses the following selection criterion 1): All canonical variables Ei>m for which the following condition is satisfied are selected: pi>2>m > Tb where Ti is a predefined threshold. Typically, the threshold Ti is greater than or equal to 0.7 or 0.8 and, generally, less than 0.95.

[0154] Once the variations in the Su signals that are correlated with variations in the S2ji signals have been identified, in step 230, the computer 30 eliminates, in each Sij signal, the variations in that Su signal identified as being correlated with variations in the S2ji signal. For this purpose, in this first embodiment, the computer proceed as follows.

[0155] The computer 30 replaces, in the matrix Eb, each canonical variable Ekm that was not selected during operation 222 with a column of zeros to obtain a modified matrix Eeb. Next, the computer 30 projects the modified matrix Eeb into the measurement space to obtain a matrix Seb. To do this, the computer 30 applies the inverse projection of the projection used to project the matrix Si into the latent space. In other words, the matrix Sei is calculated using the following relation: Sei = EebWk4, where Wk1 is the inverse of the matrix Wk. Each row i of the matrix Sei contains the variations of the signal Su that are correlated with variations of the signal S2ji. In other words, each row i of the matrix Sei contains the variations of the signal Su that are attributed to variations in the environmental conditions common to sensors Ci and C2.

[0156] Finally, during step 230, the calculator 30 subtracts the matrix Sei from the matrix Si to obtain a matrix Sci containing in each row i of this matrix the P samples of the corrected signal Scu.

[0157] Then, in this example embodiment, during a step 240, the computer 30 constructs, from the corrected signals Scu, signals Sccu corrected for periodic variations in environmental conditions. The period T of these periodic variations in environmental conditions is known and pre-recorded in memory 36. For example, for a structure 6 exposed to the external environment, the period T is equal to 24 hours, that is, the duration of a day / night cycle. Indeed, it is known that at least some environmental conditions, such as temperature, vary at the same frequency as the day / night cycles. However, what is described here can be used to correct any known periodic variation in environmental conditions and not only those that vary according to the day / night cycles.

[0158] Here, step 240 begins with an operation 242 in which the computer 30 determines, from the measurement matrix Sci and by implementing a dimensionality reduction method, a matrix Vj of R principal components Vi.q, where:

[0159] - the Sci matrix is ​​the matrix of M rows and P columns containing in each cell Sci[i,j] the sample Scuj of the signal Scu,

[0160] - R is a predefined positive integer less than P, and

[0161] - the index q is an integer that identifies the column of the matrix Vi and which is therefore between 1 and R.

[0162] Here, the dimensionality reduction method implemented is the principal component analysis method, better known by the acronym PCA (Principal Component Analysis).

[0163] The following notations are subsequently used and their correspondences with the The PCA method is explained:

[0164] Sck[j] denotes a variable whose different measured values ​​are the values ​​contained in the j-th column of the Sck matrix. The Sck[j] variables correspond to the variables, in the measurement space, of the PCA method.

[0165] Vk>q is the q-th principal component of the PCA method, where q is an order number that uniquely identifies this principal component. The principal component Vk>q is a vector of dimension P. The principal component Vk>q has P coefficients vk>q>i to vk q p. The vector Vk>q is therefore defined by the following relation: Vk>q = [vk>q>i, ..., vk>qj, ..., vkjq>P]T.

[0166] Vk is the matrix that contains, in order of increasing index q, the Q principal components Vk>q. The matrix Vk therefore has P rows and R columns. Each cell Vk[j,q] of the matrix Vk contains the coefficient vk>qj.

[0167] The matrix yk is the matrix that corresponds to the projection of the matrix Sck into the latent space of the PCA method. The matrix yk is therefore defined by the following relation: yk = Sck.Vk.

[0168] yk>q is the variable whose different measured values ​​are found in the q-th column of the matrix yk. The variable yk>q is a vector of dimension M defined by the following relation: yk>q = Sck.Vk>q. The variable yk>q is therefore representative of the following linear combination of the variables Sck[j]: yk>q = vk>q>i*Sck[l] + ...+ vk>qj*Sck[j]+ ... + vkqp*Sck[P].

[0169] Var(yk>q) is the variance of the variable yk>q.

[0170] In the PCA method, the principal component coefficients Vkji are determined to maximize the variance Var(yk>i). Then, the principal component coefficients Vk>q, with q > 1, are determined, in ascending order of index q, each time to satisfy the following two constraints:

[0171] - Constraint 3): Maximize the variance Var(yk>q), and

[0172] - Constraint 4): the covariance between this variable yk>q and any one of the variables yk qi to yk>i, is zero or minimal.

[0173] Thus, the variance Var(yk>q) decreases when the index q increases.

[0174] Here, during a sub-operation 242, the computer 30 automatically determines the matrices Vj and yi from the matrix Sci and by implementing the PCA method. Thus, during this sub-operation 242, the computer 30 projects the matrix Sci into a latent space by multiplying it by the matrix Vj in order to obtain the matrix y^

[0175] Then, during a sub-operation 244, the calculator 30 filters the columns of the matrix yi to obtain a filtered matrix yfp. Here, this filtering of each column eliminates from that column the periodic variations of frequency equal to 1 / T as well as the periodic variations of frequencies greater than 1 / T.

[0176] To do this, for example, for each column of this matrix yb the calculator Calculator 30 constructs the power spectrum of the values ​​contained in this column. To do this, calculator 30 applies a Fourier transform to this column.

[0177] Next, the calculator 30 replaces this column with a column of zeros when the cumulative power of the frequencies which, in this constructed power spectrum, are above a threshold ST, is xb times greater than the cumulative power of the frequencies which, in this same constructed power spectrum, are below this threshold ST, where:

[0178] - the ST threshold is between 0.3*T and T, and

[0179] - xb is a factor greater than one.

[0180] For example, here the threshold ST is equal to 0.5*T and the factor xb is equal to two.

[0181] After making these changes to each column of the matrix yk, the computer obtains the filtered matrix yfi in the latent space.

[0182] During a sub-operation 246, the computer 30 projects the filtered matrix yfi into the measurement space, applying the inverse projection of the projection used to project the Sci matrix into the latent space, thus obtaining a measurement matrix Scci containing, in each row i of this Scci matrix, the P samples of the corrected signal Sccij. In other words, the Scci matrix is ​​calculated using the following relation: Scci = yfi.Vf1, where Vf1 is the inverse of the matrix Vj.

[0183] Finally, during a step 250, the computer 30 detects faults from the corrected and filtered signals Sccij and Scc2j. For example, in this embodiment, the computer 30 detects a fault from the differences between the signals Scc1 and the reference signal Si>ref and, in parallel, from the differences between the signals Scc2>i and the reference signal S2>ref.

[0184] By way of illustration only, and to verify the effectiveness of the detection method described herein, sensors Ci and C2 and emitter 12 were fixed to a tube placed outdoors and thus subjected to uncontrolled variations in environmental conditions such as variations in temperature, humidity, etc. In step 250, the presence of a defect is detected using a very simplified DIC indicator, which is equal to the amplitude of the difference between the signal Si>ref and the normalized signal Scc^ / Si ref. A defect is considered detected if the value of this DIC indicator is greater than 0.1. During the tests, a significant defect was introduced at time tdef. In this example, the defect consists of sticking a 4 cm diameter Teflon disc to the tube between emitter 12 and sensor Ci.

[0185] Figure 3 shows the evolution over time of the DIC indicator. Figure 4 shows the evolution over time of a DIb indicator identical to the DIC indicator except that it is constructed using the signals instead of the Sccij signals. In Figures 3 and 4, the horizontal dashed line represents the threshold of 0.1. As can be seen in Figure 4, in the absence of corrections for variations in environmental conditions, the DIb indicator leads to a very large number of false detections of a default. Conversely, the DIC indicator, constructed by correcting for variations in environmental conditions, presents far fewer false detections.

[0186] Figure 5 represents another embodiment of the process of Figure 2. This embodiment is identical to the process of Figure 2 except that step 220 is replaced by step 260. Step 260 is identical to step 220 except that operation 230 is replaced by an operation 262 of elimination, in each signal Su, of the variations of this signal Sij identified as being correlated to variations of the signal S2j.

[0187] During operation 262, the computer 30 replaces, in the matrix Eb, each canonical variable Ei>m that was selected during operation 222, with a column of zeros to obtain a modified matrix ECp. Then, the computer 30 performs the projection of the modified matrix eci into the measurement space to directly obtain the matrix Sep. For this, the computer 30 multiplies the matrix eci by the matrix Wf1 as described during operation 230. The matrix Sci thus obtained contains, in each row i, the P samples of the corrected signal Scij.

[0188] Chapter III: Variants:

[0189] Variants of the structure:

[0190] The detection system described herein applies to thin structures other than an aircraft fuselage panel. For example, the thin structure may also be a plate, a rail, a tube, a bar, or any other part whose thickness is small compared to its length or width. In particular, for example in the case of a bar, the thin structure does not necessarily have both an external and an internal face.

[0191] The structure is not necessarily a thin structure. For example, the structure may be a civil engineering work such as a bridge or a roadway on which a vehicle travels. In this case, the emitted signal is adapted to propagate, without being attenuated too much, within the structure and, preferably, parallel to a face of that structure. For example, for this purpose, a Lambda wave is replaced by a Rayleigh wave that propagates parallel to a face of the structure.

[0192] The structure can be made of materials other than a laminated composite material. For example, the structure can be made of a non-laminated or non-composite material. In this case, for example, the structure is a turbine or propeller blade. Thus, the detection system described here can also be used with structures made of metal or concrete.

[0193] Variants of the detection system:

[0194] Other sensor technologies can be used to realize each Ck sensor. For example, the Ck sensor can be made using:

[0195] - of an electromagnetic-magnetic-acoustic sensor, better known by the acronym EMAT (“Electro Magneto-Acoustic Transducer),

[0196] - of a film made of PVDF (Polyvinylidene Fluoride) materials, or

[0197] - of an optical fiber in which a Bragg grating is made.

[0198] The use of an EMAT sensor in the context of defect detection in a thin metallic structure is described, for example, in application FR3105554. An EMAT sensor is advantageous because the vibration signal is measured without direct contact between the sensor and the external or internal face of the thin structure. In this case, the EMAT sensor is fixed to the thin structure so as to have no degrees of freedom in a direction parallel to the face of the thin structure to which it is fixed. However, it may have a small degree of freedom in a direction perpendicular to this face.

[0199] The use of a Bragg grating as a sensor of the vibration signal is described in detail in application FR3014200.

[0200] The number of sensors used in the system 8 to measure the vibration signal can be greater than two. For example, the number of sensors Ck can be greater than or equal to three, ten, or thirty-two. In this case, according to a first variant, the sensors Ck are grouped into pairs, with the two sensors in each pair being simultaneously subjected to the same variations in environmental conditions. One of the detection methods described here is then applied to each of these pairs of sensors to obtain signals corrected for variations in environmental conditions. According to a second variant, all the measurement matrices Sk of all the sensors Ck are processed simultaneously. In this second variant, the weights of the vectors wk>m are determined, not to maximize the canonical correlation between each particular pair of sensors Ck, but a common constraint that involves all the sensors.For example, this common constraint, known as "SUMCOR," is described in the Ketteringl971 article. The common "SUMCOR" constraint aims to maximize the sum of correlations between the canonical variables of all sensors Ck. However, as described in the Ketteringl971 article, it is also possible to use other constraints instead of the "SUMCOR" constraint, such as the constraints known as "MAXVAR," "SSQCOR," "MINVAR," and "GENVAR," also described in the Ketteringl971 article. In cases where the number of sensors exceeds two, the localization of the detected defect within the structure is often even more feasible.

[0201] Other embodiments are possible for the transmitter 12. For example, a piezoelectric sensor, an EMAT sensor, or a PVDF film can also be used to generate the ultrasonic wave. In this case, the same sensor can also be used to generate the ultrasonic wave and, alternately, to measure a Skji signal.

[0202] Alternatively, the emitter 12 generates, as an ultrasonic wave, not a Lamb wave but another type of ultrasonic elastic wave such as a volume wave, a surface wave, Rayleigh wave or other. In these latter cases, it is not necessary for structure 6 to be a thin structure.

[0203] The system 8 may also include several copies of the transmitter 12 fixed at different locations on the structure 6.

[0204] Alternatively, the human-machine interface 32 is removable. In this case, the interface 32 is connected to the computer 30 only during a preventive maintenance operation to display the presence or absence of a fault.

[0205] Variants of the construction of the Sci_ signals:

[0206] The time average of the Su signals is generally found in the Sep matrix. Thus, alternatively, to find signals closer to the measured physical signals in the Scb matrix, this time average can be added to the Sc signals.

[0207] Other selection criteria can be used to select the canonical variables ek>m that correspond to the highest canonical correlations. For example, instead of selection criterion 1), the following criterion 2) can be used: All canonical variables ek>m whose index m is less than or equal to the smallest index m for which the following condition is satisfied are selected: pkjk >m - pk,k,m+i < T2, where T2 is a predefined threshold. Typically, the threshold T2 is greater than or equal to 0.3 or 0.5. In practice, the value of the canonical correlation pkjk >m drops sharply beyond a certain value of the index m. Criterion 2) therefore leads to the selection of canonical variables that have an index m less than or equal to the index m that immediately precedes this sharp drop in the value of the canonical correlation pkjk >m.

[0208] In a simplified embodiment, criterion 1) is replaced by the following criterion 3): All canonical variables ek>m whose index m is less than or equal to a predefined threshold T3 are selected. For example, the threshold T3 is chosen to be equal to one.

[0209] Alternatively, the calculator 30 uses, in combination, several selection criteria, such as a combination of several of the criteria 1) to 3). In this case, the variables ek>m that are selected are those that satisfy each of the criteria of this combination of several criteria.

[0210] Other statistical analysis methods besides the mCCA method may be used. For example, Mode A of the Partial Least Squares (PLS) statistical analysis method described in the following article may be used: Jacob A. Wegelin: “A survey of Partial Least Squares (PLS) Methods, with Emphasis on the Two-Block Case,” University of Washington, Seattle, Washington, 98195 USA, Department of Statistics, Technical Report No. 371, March 2000. It is also possible to use any of the variants of Mode A of the PLS method described in that article.

[0211] Variants of the construction of Scc^ signals:

[0212] Other dimensionality reduction methods than the PCA method may can be used. For example, the following dimensionality reduction methods can be used instead of the PCA method:

[0213] - the singular value decomposition method better known by the acronym SVD (Singular Value Decomposition), and

[0214] - the independent component analysis method better known by the acronym ICA (Independent Component Analysis).

[0215] The filtering of the columns of the matrix yk can also be carried out differently to eliminate periodic variations of period T. For example, a digital bandpass filter centered on the frequency 1 / T or a digital highpass filter whose cutoff frequency at -3 dB is equal to or less than 1 / T is used to filter each of the columns of the matrix yk. The matrix yck then contains the columns thus filtered.

[0216] In a simplified variant, the construction of the Scck>i signals is omitted. In this case, for example, the fault detection step 250 is carried out using the Sckji signals directly instead of the Scckji signals.

[0217] Variants of the detection method:

[0218] A defect can be detected from the Sckji signals by performing processing other than comparing the ultrasonic wave measured by the sensor Ck to the reference signal Sk>ref. For example, alternatively, the processing performed to detect a defect from the measurements of the sensor Ck uses vibro-acoustic modulation. Such processing is described in detail in EP4155724. In this case, it is not necessary to use a pre-recorded reference signal.

[0219] Alternatively, the emitter 12 is omitted. In this case, the Ck sensors are used to measure the acoustic wave generated in response to ambient noise. The ambient noise is typically in this case the noise generated during normal use of the structure 6

[0220] When the system 8 has at least three sensors Ck distributed over the structure 6, it is possible not only to detect the presence of a defect, but also to locate its position. For this purpose, for example, the detection methods described are combined with a known method for locating a defect from the Sccki signals measured by each of the sensors Ck. For example, the distance between a sensor and the detected defect is estimated from the propagation time of the ultrasonic wave to that sensor, and then, by triangulation, the location of the detected defect is estimated. A method based on this principle is, for example, described in application FR3014200. The defect imaging method described in the following article is advantageously implemented using the Sckji or Scckji signals instead of the Skji signals: HALL, James S. and MICHAELS, Jennifer E. Computational efficiency of ultrasonic guided wave imaging algorithms.IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 2011, vol. 58, no. 1, pp. 244–248. The imaging method described in... Request FR3075373 can also be adapted to use Sck4 or Scck4 signals to locate detected faults.

[0221] Other variants:

[0222] Alternatively, the Sek4 signals are also used to measure variations in environmental conditions. For example, if only the temperature varies, then an estimate of the variation in this temperature can be obtained from the Sek>i signals as soon as the relationship linking a temperature variation to a variation in the Sek4 signal is known.

[0223] System 8 can also be adapted to measure mechanical waves with frequencies between 1 kHz and 20 kHz. In particular, acoustic waves with frequencies between 1 kHz and 20 kHz can also be used to detect defects in structures.

[0224] The microprocessor 34 can be a generic processor, a specific processor, an application-specific integrated circuit (also known as an ASIC for "Application-Specific Integrated Circuit") or an in-situ programmable gate array (also known as an FPGA for "Field-Programmable Gate Array").

[0225] Step 240 of constructing the Scck4 signals can also be implemented independently of steps 220 or 260 of constructing the Sckji signals. In this case, for example, the Scck4 signals are constructed directly using the Skji signals instead of the Sck4 signals. Step 240 can also be implemented in a fault detection method that compensates for variations in environmental conditions by a method other than that described in steps 220 and 260. In particular, this other method does not necessarily exploit a correlation between the variations of the Su signals measured by sensor Ci and the variations of the S2,i signals measured by sensor C2. For example, step 240 can be implemented in a detection system that includes a single sensor.

[0226] Several of the variants described above can be combined in the same embodiment.

[0227] Chapter IV: Advantages of the embodiments described:

[0228] Eliminating from the signal Su the variations that are most correlated with the va By adjusting the S2>i signal and eliminating, from the S2ji signal, the variations most correlated with variations in the S14 signal, we obtain corrected signals Scu and Sc2>i that are practically independent of variations in environmental conditions, which affect sensors Ci and C2 in the same way. Using these Sc14 and Sc24 signals to detect the appearance of a defect in the structure therefore limits the number of false detections and thus increases the reliability of the detection process. Here, this increase in detection reliability is achieved without using A large number of sensors are required, as this process can be implemented using only two sensors, Ci and C2. Furthermore, all sensors, Ck, are also used to detect a defect. This simplifies the implementation of the process.

[0229] Using the mCCA method to identify correlated variations in the signals and S2ji allows for an even more reliable detection process than when other statistical methods are used to do this, such as, for example, mode A of the PLS method or one of its variants.

[0230] Obtaining the corrected signals Scu and Sc2>i directly from the inverse projection of the eck matrix simplifies the detection process.

[0231] Correcting the measured signals for variations in environmental conditions before comparing them to reference signals limits the number of false detections and substantially increases the reliability of the detection process.

[0232] Detecting defects from Scck>i signals, in addition corrected for periodic variations of period T, makes it possible to further improve the reliability of the detection process.

Claims

1. Demands A method for detecting a defect in a structure using high-frequency mechanical waves, this method comprising the following steps: - the instrumentation (200) of the structure by fixing on this structure a first and a second high-frequency mechanical wave sensors in locations where: - Both the first and second sensors are sensitive to defects that may appear in the structure, - the variations in the signals measured by the first and second sensors, which are caused by a defect in the structure, are not correlated with each other, and - The first and second sensors are subject to the same variations in environmental conditions that can cause variations in the signals measured by these first and second sensors. - during the operation (210) of the structure, the acquisition (216), by an electronic computer, during several successive measurement periods, of first signals measured by the first sensor and of second signals S2ji measured, in parallel, by the second sensor, each measured signal Skji being made up of a time sequence of P samples Skjij, where: - the index i is a sequence number that identifies the measurement period during which the Skji signal was measured, - the index k is an identifier of the sensor that measured this time sequence of P samples, this index k being equal to one to identify the first sensor and to two to identify the second sensor, and - the index j is an order number that identifies the position of the sample within a time sequence of P samples, then - the construction (220; 260) of signals Scij and Sc2ji corrected for environmental variations common to the first and second sensors, each signal Sck>i being made up of a time sequence of P samples Sck>ijj, - the detection (250) of faults from the corrected signals Scu and Sc2>i, characterized in that the construction (220; 260) of the signals Scu and Sc2>i comprises: - the identification (222), by a statistical analysis method, in each signal S u, of the variations of this Ski signal which are correlated with variations of the signal S2ji and, in each signal S2j, variations of this signal S2ji which are correlated with variations of the signal Spuis - the elimination (230; 262), in each signal S^, of the variations of this signal Si identified as being correlated with variations of the signal S2ji and, in each signal S2ji, of the variations of this signal S2ji identified as being correlated with variations of the signal S to obtain the signals Scu and Sc2>i.

2. A method according to claim 1, wherein the statistical analysis method implemented is the mCCA (multiset Canonical Correlation Analysis) method.

3. A method according to claim 2, wherein the identification (222), in each Skji signal, of the variations of this Skji signal that are correlated with variations of the Skji signal, comprises, for k equal to one and k' equal to two and for k equal to two and k' equal to one: - the projection of a measurement matrix Sk from a measurement space to a latent space to obtain a matrix ek containing M rows and Q columns, where: - the matrix Sk is a matrix of M rows and P columns containing in each cell Sk[i,j] the sample Skjij, - M is the number of measurement periods, - Q is a predefined positive integer less than P, - the matrix ek is the matrix that contains, in columns, the Q canonical variables ek>m of the mCCA method, where m is an integer index that varies from 1 to Q, the correlation between each canonical variable ek m and the canonical variable ek >m being given by the canonical correlation pkjk >m of the method mCCA,- the selection (226) in the matrix ek of a limited number of canonical variables ek>m for which the canonical correlations pkjk >m are the highest, the canonical variables ek>m thus selected identifying the variations of the Skji signals which are correlated with variations of the Sk >i signal.

4. A method according to claim 3, wherein the elimination (230), in each Skji signal, of variations of that Skji signal identified as being correlated with variations of the Skji signal, comprises, for k equal to one and for k equal to two: - the replacement in the matrix ek of each canonical variable ek>m that has not been selected by a column of zeros to obtain a modified matrix eek, then - the projection of the modified matrix eek into the measure space by applying the inverse projection of the projection used to project the matrix Sk into the latent space, thus obtaining a measure matrix Sek, then - the subtraction of the matrix Sek from the matrix Sk to obtain a matrix Sck containing in each row i of this matrix Sck the P samples of the corrected signal Sck>i.

5. A method according to claim 3, wherein the elimination (262), in each Skji signal, of variations of this Skji signal identified as being correlated with variations of the Skji signal, comprises for k equal to one and for k equal to two: - the replacement, in the matrix ek, of each canonical variable ek>m that has been selected, by a column of zeros to obtain a modified matrix eck, then - the projection of the modified matrix eck into the measurement space by applying the inverse projection of the projection used to project the matrix Sk into the latent space and thus obtain a measurement matrix Sck containing in each row i of this matrix the P samples of the corrected signal Sckji.

6. A method according to any one of the preceding claims, wherein the method also comprises: - for each sensor Ck, the construction (240) of Scck>i signals corrected for periodic environmental variations of predefined period T from the corrected Sck>i signals, the construction of these Scck>i signals comprising: - the determination (242), from a Sck matrix of measures and by implementing a dimensionality reduction method, of a Vk matrix of R principal components, where: - The Sck matrix is ​​a matrix of M rows and P columns containing, in each cell Sck[i,j], the sample Sck>ijj of the signal Sck>i, - M is the number of measurement periods, - R is a predefined positive integer less than P, then - the projection (242), using the matrix Vk, of the Sck matrix of measures from a measure space to a latent space to obtain a matrix yk containing M rows and R columns, then - filtering (244) the columns of the matrix yk to obtain a filtered matrix yfk, filtering each column eliminating from that column periodic variations of frequencies equal to 1 / T, - the projection (246) of the filtered matrix yfk into the measurement space by applying the inverse projection of the projection used to project the matrix Sck into the latent space and thus obtain a measurement matrix Scck containing in each row i of this matrix the P samples of the corrected signal Scck>i, then - the detection of faults from the corrected signals Scu and Sc2>i, includes the detection of faults from the corrected signals Scck>i.

7. A method according to claim 6, wherein the filtering (244) of the columns of the matrix yk comprises, for each column of this matrix yk: - the construction of the power spectrum of the values ​​contained in this column, then - the replacement of this column by a column of zeros when the cumulative power of the frequencies which, in this constructed power spectrum, are greater than a threshold ST, is xb times greater than the cumulative power of the frequencies which, in this constructed power spectrum, are less than this threshold ST, where: - the threshold ST is between 0.3*T and T, and - xb is greater than one.

8. A method according to claim 7 or 8, wherein the dimensionality reduction method implemented is the principal component analysis method.

9. A method according to any one of the preceding claims, wherein the method comprises: - the measurement by the first and second sensors, respectively, of a first reference signal Si>ref and a second reference signal S2>ref in the absence of a defect in the structure, each reference signal being made up of a time sequence of P samples Sk>refj, then - the recording of the signals Si>ref and S2>ref, then - during the defect detection step, a defect is detected from the deviations between a signal obtained from the corrected signal Scu and the recorded reference signal S1>ref and, in parallel, from the deviations between a signal obtained from the corrected signal Sc2>i and the pre-recorded reference signal S2>ref.

10. Information storage medium (36), readable by a microprocessor (34), comprising instructions executable by that microprocessor, characterized in that this medium comprises non-transient instructions for the execution of a process conforming to one any of the preceding claims, when these instructions are executed by the microprocessor.

11. A system for detecting a defect in a structure using high-frequency mechanical waves, this system comprising: - a first and a second high-frequency mechanical wave sensor fixed to the structure at locations where: - both the first and second sensors are sensitive to defects likely to appear in the structure, - the variations in the signals measured by the first and second sensors that are caused by a defect in the structure are not correlated with each other, and - the first and second sensors are subjected to the same variations in environmental conditions likely to cause variations in the signals measured by these first and second sensors, - an electronic computer (30) programmed to perform the following steps: - during the operation of the structure, to acquire, during several successive measurement periods, first signals Su measured by the first sensor and second signals S2.measured, in parallel, by the second sensor, each measured signal Skji being composed of a time sequence of P samples Skjij, where: - the index i is an order number that identifies the measurement period during which the Skji signal was measured, - the index k is an identifier of the sensor that measured this time sequence of P samples, this index k being equal to one to identify the first sensor and to two to identify the second sensor, and - the index j is an order number that identifies the position of the sample within a time sequence of P samples, then - construct signals Scu and Sc2>i corrected for environmental variations common to the first and second sensors, each signal Sck>i being composed of a time sequence of P samples Sck>i>j, - detect faults from the corrected signals Scu and Sc2>i, characterized in that the electronic computer (30) is also configured to,during the construction of the Scij and Sc2ji signals: - identify, using a statistical analysis method, in each signal S14, variations of this signal Su that are correlated with variations of the signal S2ji and, in each signal S2ji, variations of this signal S2ji that are correlated with variations of the signal Sij, then, - eliminate, in each signal S14, the variations of this signal Su identified as being correlated with variations of the signal S2ji and, in each signal S2j, the variations of this signal S2ji identified as being correlated with variations of the signal , to obtain the signals Scu and Sc2>i.

Images