FR3105527A1 - PROCESS FOR MONITORING A STRUCTURE BY IMAGE PROCESSING - Google Patents

Abstract

Method (P) for monitoring a structure by image processing comprising: (A) a learning phase comprising steps of: (A2) determination for learning images of a surface of the structure of a learning displacement field by an image correlation technique, (A3) construction of a base of modal displacement fields by a model reduction technique, (A4) determination of a residual threshold and a range of expected values of a modal amplitude, (S) a monitoring phase comprising steps of: (S1) acquisition of a current image of the portion, (S2) determination of a current displacement field, of amplitudes current modal amplitudes and a current residual, and (S3) generating an alert when one of the current modal amplitudes takes a value outside the range of expected values of the modal amplitude or when a norm of the residual current exceeds the residual threshold. Figure for abstract: Fig. 3

Description

METHOD FOR MONITORING A STRUCTURE BY IMAGE PROCESSING

FIELD OF THE INVENTION

The present invention relates to the general field of monitoring a structure, and in particular the detection of the appearance of a fault in a structure. The structure can, for example, be located within or include a wind turbine mast, a nuclear power plant aerogenerator in thin structures or even slender structures subjected to repeated and/or random stresses over time.

The present invention also relates to the general field of non-destructive testing and monitoring of the health of structures, in particular applied to the wind industry on land and at sea.

STATE OF THE ART

Following their deployment in past years, some of the wind farms are beginning to suffer the ravages of time. The market has also, over time, become competitive. The pressure is therefore growing to reduce operating costs. One lever for reducing wind turbine operating costs lies in optimizing turbine maintenance.

The lifetime of a wind turbine is influenced by uncertain parameters related to the environment such as wind or storms, or uncertain parameters related to materials such as a weld defect or the failure of a bolt. This makes it very difficult to predict with certainty the service life of a mast. It is therefore essential to revisit the forecasting tools based on site data relating to the damage to the wind turbine mast and the structures supporting them.

Classically, for checking the state of health of structures, two techniques are possible:

- non-destructive testing, or in English Non Destructive Testing (NDT), which consists of checking the structure as a whole at regular intervals to detect the appearance of defects.

- structural health monitoring, or in English Structural Health Monitoring (SHM), consists of continuously recording data in order to detect any change in the behavior of the structure, a symptom of damage. SHM techniques are, in fact, a subset of NDT techniques.

NDT techniques include visual inspection, ultrasonic inspection and eddy current inspection (for metal structures). However, these techniques have the disadvantage that their use may involve:

- immobilization or stoppage of operation of the devices supported by the structure for a long period,

- a prohibitive cost if you want to comprehensively control a large part of the structure,

- a complexity of implementation, possibly a preparation of the surfaces, and especially

- the full-time work of specialized technicians.

Thus, NDT techniques will rather be reserved for situations where damage is suspected, localized in space and time, to carry out a more detailed diagnosis and allow the alert to be disqualified, or to remedy it. This then calls for the implementation of a continuous and automated control of the structures, SHM if possible with low-cost technologies, allowing the centralization of the control.

SHM techniques include for example damage detection by frequency tracking, an example of which for wind turbine towers has been published in “ Monitoring resonant frequencies and damping values of an offshore wind turbine in parked conditions ” Devriendt, C., Weijtjens , W., El-Kafafy, M., & De Sitter, G. (2014) IET Renewable Power Generation, 8(4), 433-441. A measured natural frequency variation of the structure is the sign of the appearance of a defect, but on the one hand it is necessary to carry out measurements at very low noise levels (delicate to obtain in a context of random stresses) and on the other hand the variations in the properties of stiffness or inertia of the structure due to the environment, such as the variation of elastic moduli with temperature or frost for example, can falsify the variations of natural frequencies measured.

It was also suggested in the publication “ Vibration analysis of 2.3 MW wind turbine operation using the discrete wavelet transform ”. Bassett, K., Carriveau, R., & Ting, DS (2010). Wind Engineering , 34 (4), 375-388 to place strain gauges at regular space intervals around the hot spots of the structure (areas of greatest strain) in order to detect localized gradients of strains which would result from the presence of cracks. This technique has the disadvantage, on the one hand that a very high density of sensors is required in order to detect any appearance of crack, on the other hand that the strain gauges undergo measurement drifts over time and may not be located in sensitive areas, thus rendering the measurement ineffective.

There is therefore a need for a technique for monitoring the state of health of structures such as wind turbine masts, robust over time, not requiring expensive measuring devices, in large quantities and with low noise level. . Moreover, the common disadvantage of all these techniques known from the prior art is to extrapolate a state of health of a structure as a whole from local information. There is a need for a technique for monitoring the state of health of wind turbine towers, the implementation of which allows a more direct determination of the nature of the mechanical damage from the results obtained.

An object of the invention is to propose a technique for monitoring the state of health of structures such as wind turbine masts, robust over time, not requiring expensive measuring devices, in large quantities and at low noise level.

Another object of the invention is to propose a technique for monitoring the state of health of wind turbine masts, the implementation of which allows a more direct determination of the nature of the mechanical damage from the results obtained.

The object is achieved in the context of the present invention thanks to a method for monitoring a structure by processing digital images, comprising:
(A) a learning phase comprising steps of:
(A2) from a plurality of training images of a portion of a surface of the structure, determination for each image of a training displacement field by an image correlation technique, the field being an estimate of a displacement of the material within the structure at each point of the portion at an instant of acquisition of the image,
(A3) from the fields, construction of a base of modal displacement fields characteristic of the plurality of images by a model reduction technique,
(A4) determination of a residual threshold characterizing a maximum admissible error of the displacement estimate and, for each modal displacement field, of a range of expected values of an associated modal amplitude characterizing a projection of the displacement fields learning on the modal displacement field,
(S) a monitoring phase comprising steps of:
(S1) acquisition of a current image of the portion,
(S2) from the current image, determination of a current displacement field, of current modal amplitudes characterizing a projection of the current displacement field on the basis of modal displacement fields and of a current residual, the field of current displacement being an estimate of a displacement of the material within the structure at each point of the portion at an instant of acquisition of the current image, the current residual characterizing an error in the estimate of the displacement, and
(S3) generation of an alert when one of the current modal amplitudes associated with one of the modal displacement fields takes a value outside the range of expected values of the modal amplitude associated with the modal displacement field or when a norm of the current residual exceeds the residual threshold.

The evaluation of the state of health being carried out by image acquisition and processing, the equipment necessary for the implementation of the process is composed of digital cameras and computers that can be chosen from medium or low quality. , that is to say at lower cost and without requiring a low level of noise, while having a satisfactory quality of the process. In addition, the process is robust over time thanks to the monitoring of a projection of the current displacement on a modal learning basis, characteristic of a normal state of health of the structure.

Such a monitoring method is advantageously completed by the following different characteristics or steps taken alone or in combination:

- step (S2) comprises a first sub-step (S2A) of determining the current displacement field by a global image correlation technique and a second sub-step (S2B) of projecting the current displacement field onto the base of modal displacement fields;

- step (S2) comprises the determination of the current displacement field by an integrated image correlation technique, corresponding to a global image correlation technique using the basis of modal displacement fields;

- step (S2) includes:

a substep (S21) of filtering the reference image and the current image to obtain a filtered reference image and a filtered current image, by applying an inhomogeneous spatial filter of the low pass type, the filter using a plurality of filter characteristic lengths for a plurality of pixels forming the image, the characteristic length for a pixel depending on a maximum observed displacement associated with the pixel in the training displacement fields,

a sub-step (S23) of determining, for each modal displacement field, a current modal amplitude of the current displacement field, from a scalar product between an extractor field associated with the modal displacement field and a difference between the filtered current image and the filtered reference image, the extractor field being an image associated with the modal displacement field constructed from the set of modal displacement fields and the filtered reference image, and

a substep (S24) of determining the current residual from the filtered current image, the filtered reference image, the current modal amplitudes and the modal displacement fields;

- the structure includes a wind turbine mast;

- the surface is inside the mast;

- the inner zone of the mast extends on either side of a weld between two collars of the mast;

- the learning phase is implemented on different portions of the surface and then the monitoring phase is implemented on these portions;

The invention also relates to a device for monitoring a structure by processing digital images, comprising an image acquisition device and an image processing unit connected to the device and adapted to store in memory images and implement a monitoring process as presented above.

The invention also relates to a wind turbine comprising a monitoring device as mentioned above, or a plurality of monitoring devices of this type for implementing the monitoring method where the learning phase is implemented on different portions of the surface and then the monitoring phase is implemented on these portions as mentioned above.

The invention further relates to a computer program comprising instructions adapted to the implementation of at least one of the steps of the method as it was presented above when said program is executed on a computer.

The invention finally relates to a database comprising the plurality of training images, the reference image, the training displacement fields, the base of modal displacement fields, the residual threshold, the ranges of expected values of the modal amplitudes, the current images, the current modal amplitudes and the current residuals determined during the implementation of a monitoring method such as could be presented above.

Such a database is advantageously supplemented by the filtered reference image, the filtered current images, the inhomogeneous spatial filter of the low-pass type and the extractor fields determined during the implementation of a monitoring method comprising a sub -step (S21) of filtering as it was presented above.

DESCRIPTION OF FIGURES

Other characteristics and advantages of the invention will emerge from the description which follows, which is purely illustrative and not limiting, and must be read in conjunction with the appended drawings in which:

FIG. 1 is a schematic representation of a device for monitoring a structure by image processing according to the invention.

FIG. 2 is an example of an image acquired during a method for monitoring a structure according to the invention.

FIG. 3 is a schematic representation of a method for monitoring a structure according to the invention.

DETAILED DESCRIPTION OF THE INVENTION

Measuring device

FIG. 1 represents a device 1 for monitoring a structure 9 by image processing in accordance with one embodiment of the invention.

The surveillance device 1 comprises a digital image acquisition device 3, a lighting device 5, a holding architecture 7 suitable for receiving and fixing the digital image acquisition device 3 and the lighting 5 on the structure 9 to be monitored. The digital image acquisition device 3 can in particular be a digital still camera or a digital camera. It should be noted that the digital image acquisition devices used can be of average or low quality. The digital image acquisition device 3 is associated with a clock to timestamp the images taken.

The monitoring device 1 also comprises an image processing unit 8 connected to the digital image acquisition device 3. The processing unit comprises a memory for recording data and processor-type calculation means for processing images.

The lighting device 5 is placed and oriented so as to illuminate a surface 15 of the structure 9.

The digital image acquisition device 3 is placed and oriented so as to take snapshots of a surface portion 11 of the structure 9, the surface portion 11 being included in the surface 15. The target surface will be chosen as centered on sensitive areas defined according to the field of application.

In the case where the structure 9 is a wind turbine or a wind turbine mast, the surface portion 11 can be an interior zone of the wind turbine mast. In this case, the surface portion 11 of the structure 9 can be chosen as an inner surface extending on either side of a weld. Indeed, a wind turbine mast is traditionally made up of several sections fixed to each other by bolts. Each section is itself formed of several ferrules which are welded together. An area of interest to be monitored in order to follow the state of health of the wind turbine mast extends on either side of a weld between two shells, and more particularly a zone located between the weld material and the material of the ferrule, usually steel. The surface portion 11 may in particular comprise a zone extending from a weld 17 between two ferrules.

In operation, the lighting device 5 illuminates the surface 15, and the image acquisition device acquires images of the surface portion 11.

This acquired image is transmitted and recorded in the image processing unit 8.

The image processing unit 8 is programmed to implement a process for monitoring the structure by processing the acquired images.

The surface portion 11 of the structure 9 can be covered with a coating having for example a mottled pattern formed of black spots 13 on a white background, as usually practiced during the implementation of the Correlation of Digital Images (cf e.g. M.A. Sutton et al., “Image correlation for shape, motion and deformation measurements: basic concepts, theory and applications”, Springer, 2009).

The resolution of the acquisition device 3 can be adapted to the contrast of the surface portion 11.

Figure 2 shows a snapshot of a portion of the inner surface of a wind turbine mast 29. The snapshot shows a weld 17 between two ferrules. A coating presenting a speckle formed of black spots 23 on a white background has been affixed to the internal surface of the mast 29.

In another embodiment, it is possible to use a plurality of monitoring devices can be installed to image different surface portions of the structure 9.

The number of monitoring devices to be used depends on the number of areas you want to monitor over time. By limiting yourself to taking images of the most sensitive areas of the structure, it is possible to monitor the entire structure with a reduced number of digital image acquisition devices, or to limit monitoring to areas specific, such as, for example, if the monitored structure is a wind turbine mast, the welds between two ferrules.

Each image acquisition device of the device can be associated with a lighting device and maintained on a support architecture adapted to extend inside or at least close to the monitored structure. A support architecture is placed and configured to accommodate and secure at least one set of image acquisition devices and one set of lighting devices.

The structure to be monitored can be a wind turbine, one or more monitoring devices being installed inside the wind turbine, for example in the wind turbine mast. In the case of a plurality of monitoring devices, the different observation portions are imaged. They may in particular correspond to the zone surrounding a weld between two ferrules of the wind turbine mast.

FIG. 3 illustrates the steps of the monitoring method implemented by the processing unit 8 from images acquired from the monitored structure.

The method comprises a first phase (or learning phase) making it possible to determine a modal base characterizing the expected displacements within the surface portion of the monitored structure and a second phase (or monitoring phase) making it possible to evaluate whether displacements measured within the surface portion of the monitored structure correspond to expected displacements.

The method takes advantage of the fact that over time, the monitored structure evolves so that two images successively acquired with the measuring device using the same digital image acquisition device placed under the same conditions are different.

The evolution of the state of health of the monitored structure can be evaluated by following the evolution over time of the images and more particularly of the displacements of material on the monitored surface. Monitoring consists of verifying whether the movements observed are acceptable or not.

Learning phase (A)

The learning phase (A) includes the following steps:

According to a step A1, the digital image acquisition device 3 acquires a reference image and a plurality of training images. The reference image and the plurality of training images are acquired during a period for which the state of health of the monitored structure is deemed “normal” or “satisfactory”.

The reference image and the learning images acquired are stored in the memory of the processing unit. In addition, for the reference image and for each acquired learning image, an associated acquisition instant is also recorded in the memory. The instant of acquisition of the learning image is not exploited during step A2 of the learning phase, so this temporal dimension does not appear in the description of this phase. However, the time dimension is important in step A3 of the learning phase so that it reappears there.

According to a step A2, the image processing unit 8 determines for each learning image, a learning displacement field by a digital image correlation technique. The learning displacement fields constitute a statistical sampling of the displacements of the material on the monitored structure observed during the learning image acquisition phase.

Brief presentation of Digital Image Correlation (CIN).

The displacement field can be measured in the following way.

Position x corresponds to a two-dimensional vector making it possible to locate the pixel within the digital image. Throughout the text, the terms in bold designate quantities which are two-dimensional vectors. The reference image is represented by a function f( x ), which associates the associated gray level f( x ) with the position pixel x . Each learning image is represented by a function g( x ), which associates the associated gray level g( x ) with the position pixel x .

To study the displacements, the training images are compared to the reference image by CIN.

The displacement field between the training image, g( x ), and the reference image, f( x ), can be represented by a two-dimensional field u ( x ); u ( x ) is the displacement vector in the plane of the images which makes it possible to pass from position x in the reference image at which an element of the photographed surface is located, to position x+u ( x ) in l training image at which the same material element of the surface is located, ideally f( x ) = g( x + u ( x )).

In reality, this equality is never satisfied. Noise necessarily taints the images, if only the noise of the digital image acquisition device and more particularly the noise intrinsic to the light intensity sensors of the digital image acquisition device . Intrinsic noise can be characterized by a numerical value or noise level.

It is useful to consider a residual η( x ) which expresses the non-conservation of gray levels in the form η( x ) = g( x + u ( x )) - f( x ) in each pixel, and a norm | η|, for example the Euclidean norm which is the square root of the sum of the squares of η( x ) over all pixels x . Other standards may be used. Ideally the residue η( x ) merges with the noise present in the reference image and the learning image. When the norm of the residue |η| takes a value which is significantly higher than the noise level estimated beforehand, then this residual signals an error made in the estimation of the displacements or a violation of the hypothesis of conservation of the levels of gray which can happen when the luminosity varied.

The training displacement field u ( x ) is obtained by minimizing the sum of the squared differences in terms of gray levels between the reference image and the training image corrected for the training displacement field.

It is also possible to use a so-called “global” CIN technique, also referred to here as a global image correlation technique. This type of technique is known to those skilled in the art. The “global” characteristic designates the use in the CIN technique of a base of displacement fields, each of which is defined over the whole of the field of study. The field of study designates here the portion of the surface of the structure considered during the use of the CIN process. Such a technique implements a decomposition of the training displacement field u ( x ) on the basis of displacement fields or shape functions n _i ( x ) according to

The coefficients a _i are unknowns to be determined to obtain the learning displacement field u ( x ).

This global CIN method is detailed in Besnard, G., Hild, F., & Roux, S. (2006). “Finite-element” displacement fields analysis from digital images: application to Portevin–Le Châtelier bands. Experimental Mechanics , 46 (6), 789-803.

Following the application of the global CIN technique, it is determined at the end of step A2:

- a learning displacement field u(x),

- a learning residue η( x ) verifying η( x ) = g( x + u ( x )) - f( x ).

Under current conditions, the uncertainty in the determination of the displacement field is about 10 ^-2 pixel. This uncertainty of determination is sufficiently low to allow for example to detect cracks invisible to the naked eye.

According to a step A3 of the learning phase, the processing unit 8 extracts from the learning displacement fields a base of modal displacement fields characteristic of these learning displacement fields by a model reduction technique

It is possible to take advantage of the fact that there are only a finite number of types of displacements within the structure. This means that, for a certain number of learning images of the monitored structure, the determined learning displacement fields can be described by a smaller number of particular displacement fields. These particular displacement fields are called modal displacement fields which together form a modal base.

The modal displacement fields make it possible to analyze the different types of displacements contained in the training images which are characteristic of the monitored structure and its state of health during the period when the training images were acquired.

If the training images all correspond to a state of health deemed "normal" or "satisfactory" of the monitored structure, the modal displacement fields are then characteristic of the displacements expected within the surface portion of the monitored structure when the supervised structure has a "normal" or "satisfactory" state of health.

To determine the modal displacement fields, a write of the learning displacement fields u ( x, t) indexed by the acquisition time t is sought as a sum of terms, each term being the product d 'a temporal part and a spatial part:

ψ _i ( x ) is the modal displacement field and corresponds to the spatial part. b _i (t) is the corresponding amplitude of ψ _i ( x ) in the learning field u ( x, t), b _i (t) is a learning modal amplitude and corresponds to the time part.

The modes ψ _i ( x ) can be determined thanks to model reduction techniques known from the prior art such as for example the so-called SVD technique (for "Singular Value Decomposition" in English known in French as "decomposition aux Valeurs singulières ").

The modal displacement fields are determined in a decreasing order of importance which is quantified by an eigenvalue of the mode. This order of determination makes it possible to determine only the modes more important than a certain threshold of importance arbitrarily chosen. When the importance associated with the last mode determined falls below the importance threshold, the determination method can be stopped. It is possible to use as a threshold of importance the uncertainty of determination of the displacement field by the CIN technique used previously.

The modal displacement fields characterize the deformation kinematics of the monitored structure in the region targeted by the measuring device. They also characterize the banal movement of the digital image acquisition device due, for example, to a vibration of the architecture holding the measuring device. The displacement modal displacement fields associated with the movements of the digital image acquisition device are called rigid body modes. These rigid body modes do not carry any interesting information concerning the local deformations of the monitored structure. However, it is essential to limit their influence as much as possible so that they do not mask the other modes. As a pre-processing of the search for kinematic modes, it is possible to subtract the average rigid translations in the principal directions using CIN by fast Fourier transform (FFT). Moreover, if rotations appear at the level of the support architecture of the measuring device, techniques also exist to limit their influence.

At the end of step A3, the processing unit 8 has determined a base comprising a certain number of modal displacement fields ψ _i ( x ). Once this base has been fixed, the processing unit 8 updates for each learning image g( x ) the learning displacement field, so that the updated learning displacement field ua( x ,t) s' writing

with b _i (t) the learning modal amplitude, and the updated learning residual ηa( x ) verifies ηa( x ,t) = g( x + ua ( x ,t),t) - f( x ).

According to a step A4 of the learning phase, the processing unit 8 determines a range of expected values of modal amplitudes characterizing a projection of the learning displacement fields on the basis of modal displacement fields and a residual threshold characterizing a maximum admissible error.

The range of expected values of modal amplitudes is determined from the learning modal amplitudes b _i (t) obtained during step A3.

During step A3, a learning modal amplitude b _i (t) per learning image is determined for each modal displacement field ψ _i ( x ). It is possible to characterize the set of learning modal amplitudes b _i (t) of the modal displacement field ψ _i ( x ) by position parameters (mean value, median, etc.) and dispersion parameters ( standard deviation, interquartile range, etc.). It is possible in particular to construct a range of expected values of the modal amplitude b _i (t) such as for example the interquartile range representing 50% of the learning modal amplitudes b _i (t) distributed symmetrically around the median of the modal learning amplitudes. The range of expected values can also be determined as an interval representing 30%, 60% or any other percentage of the learning modal amplitudes b _i (t) distributed symmetrically around the median or the mean of the learning modal amplitudes. Other ways of determining a range of expected values can also be used.

The residual threshold characterizing a maximum admissible error can be calculated from the updated learning residuals ηa( x, t). A possible method of calculation is to use the norms of these discounted learning residuals and to extract from these norms a mean value, a standard deviation, a last quartile or another quantile. From these parameters, it is possible to calculate a residual threshold equal to a sufficiently large value of the residual norm for a majority of the learning residuals to have a norm less than or equal to the residual threshold.

It is obtained at the end of step A3:

- a residual threshold characterizing a maximum admissible error from the updated learning residuals,

- for each modal displacement field ψ _i ( x ), a range of expected values of the associated modal amplitude.

Monitoring phase

The monitoring phase includes the following steps:

According to a step S1 of the monitoring phase, the digital image acquisition device 3 acquires at an acquisition time t a current image of the surface portion 11 of the structure.

This current image is taken by the same measuring device 1 which allowed the acquisition of the training images, and under the same conditions. This makes it possible to compare in a relevant way the displacements measured in the learning phase and those measured in the monitoring phase.

The current image is represented by a function h( x, t), which associates the gray level h( x , t) with the pixel at position x .

According to a step S2 of the monitoring phase, the processing unit 8 determines, from the current image, a current displacement field on the modal base selected during the learning phase and characterized, for each field of modal displacement, by a current modal amplitude and a current residue characterizing an error.

The current displacement field is an estimate of the displacement of the material within the structure at each point of the observation zone at the instant of acquisition of the current image, and the current residual characterizes an error of the estimation of this displacement.

The processing unit 8 seeks to determine, from the current image h( x, t), a current displacement field v ( x, t) which is a displacement vector in the plane of the images which makes it possible to pass from position x in the reference image at which an element of the surface photographed is located, to position x+v ( x, t) in the current image at which the same material element of the surface is located, with an error in the determination represented by the current residual ε( x, t) i.e. f( x ) = h( x + v ( x, t) , t) + ε( x, t).

The current displacement field is an estimate of displacements within the structure between the current image and the reference image. The current residual ε( x, t) is characteristic of an error in the estimate

The current displacement field is sought in the form of a linear combination of the modal displacement fields ψ _i ( x ) determined during the learning phase, the amplitude of the mode ψ _i ( x ) in the current displacement field v ( x, t) being denoted c _i (t). The amplitude c _i (t) is a current modal amplitude so that

To carry out step S2 there are several possible methodologies giving theoretically the same result. They are presented below in increasing order of their effectiveness.

According to a first methodology, step S2 comprises a first substep S2A of determining the current displacement field by a digital image correlation technique and a second substep S2B of projecting the current displacement field on the basis of modal displacement fields.

The processing unit 8 performs the first sub-step S2A by following, for example, a “global” CIN digital image correlation technique as presented previously and using a base of functions n _i ( x ).

At the end of this first substep S2A, an intermediate displacement field vi ( x, t) is determined as well as an intermediate residue εi( x, t).

During the second sub-step S2B, the intermediate displacement field is projected on the basis of modal displacement fields so as to determine the final displacement field vf ( x, t):

From the final displacement field, the intermediate displacement field, the intermediate residue, and a gradient of the reference image, the processing unit 8 determines the final residue εf( x, t) according to:

grad representing the gradient function. The star * representing the Schur matrix product. The Schur matrix product is a binary operation which, with two initial matrices of the same dimensions, associates a final matrix of the same dimension, each coefficient of which is the term-by-term product of the two initial matrices.

According to a second methodology, step S2 comprises the determination of the current displacement field by a so-called “integrated” digital image correlation technique, also referred to here as an integrated image correlation technique, using the base of fields of modal shifts. The so-called "integrated" digital image correlation technique is a global CIN-like method, in which the basis of functions n _i ( x ) used to determine the current displacement field is the basis of modal displacement fields ψ _i ( x ) determined during the learning phase. No sub-step subsequent to the application of the global CIN method is then necessary.

The so-called “integrated” approach has been the subject of publications in the literature, for example in H. Leclerc, Jean-Noël Périé, F. Hild, S. Roux. Integrated Digital Image Correlation for the identification of mechanical properties. CFM 2009 - 19th French Mechanical Congress, Marseille (France) , 2009, Marseille, France [https://hal.archives-ouvertes.fr/hal-02194355].

According to a third methodology, step S2 includes

- a first sub-step S21 of filtering the reference image and the current image by applying an inhomogeneous low-pass type spatial filter, the filter comprising a plurality of filtering characteristic lengths for a plurality of pixels forming the image, the characteristic length for a pixel depending on the maximum displacement observed at the pixel in the training displacement fields.

A filtering of an image by a low-pass spatial filter designates here a processing of the image which modifies the levels of gray composing the image so as to erase the variations of levels of gray of higher spatial frequencies, it is i.e. the strong variations of gray levels over small distances in the image. This makes it possible to smooth the variations in gray levels over a certain characteristic filtering distance or characteristic filtering length.

A filtering of an image by an inhomogeneous low-pass spatial filter designates here an image processing where the characteristic filtering length varies from one pixel to another of the image.

The processing unit 8 constructs, from the learning displacement fields, a function |u _max |( x ) which associates a pixel of position x with the maximum displacement distance, equal to a norm of the displacement vector, at position pixel x among all training displacement fields.

The processing unit 8 extracts from the function |u _max |( x ) a distance λmax which is the maximum of the function |u _max |( x ), and a plurality of distances λi comprised between 1 and λmax. In particular, the processing unit 8 can determine N distances λi with i=1…N satisfying λ1=q, λ2=q^2, λ3=q^3, …, λN=q^N=λmax.

The processing unit 8 generates N relay images which are filtered images of the reference image f( x ). To generate one of these N relay images, the processing unit 8 applies to the reference image a homogeneous spatial filter of the low-pass type characterized by a single characteristic filtering length for all the pixels, the characteristic length being equal to a distances λi. This spatial filter can for example use a convolution with a Gaussian function.

The processing unit 8 similarly generates N relay images of the current image h( x ).

The application of the homogeneous spatial filter of characteristic length λi generates a relay image fi( x ) of the reference image f( x ) and a relay image hi( x, t) of the current image h( x, t) .

The processing unit 8 applies to the reference image the inhomogeneous spatial filter of the low-pass type pixel by pixel to produce a filtered image f!( x ). At pixel x , the processing unit searches for the integer p between 1 and N, such that λ(p-1)≤|u _max |( x )≤λp. We use λ0=1.

The gray level f!( x ) of the filtered image at pixel x is obtained from the gray levels at pixel x of the relay images f(p-1)( x ) and fp( x ). More precisely, the level of gray f!( x ) is constructed by linear interpolation of the levels of gray f(p-1)( x ) and fp( x ).

Similarly, the processing unit 8 applies to the current image the inhomogeneous spatial filter of the low-pass type pixel by pixel to produce a filtered image h!( x, t) this time using the relay images hi( x, t) .

- a second sub-step S22 for determining extractor fields ξ _i ( x ) from the modal displacement fields ψ _i (x )∙ and from the filtered reference image f!( x ).

From the spatial modes ψ _i (x ), as described above, it is possible to determine the sensitivity fields s _i ( x ) = ψ _i (x )* grad (f!( x )) .

By introducing the matrix [ M ] whose coefficients Mij are defined by

it is possible to define extractor fields ξ _i ( x ), each extractor field being associated with a modal displacement field ψ _i (x ),

Or

is the coefficient ij of the inverse matrix of [ M ].

- a third sub-step S23 for determining the current modal amplitudes c _i (t),

Extractor fields have the following property: in a current image h( x, t) acquired at an acquisition time t , the modal amplitude c _i (t) associated with the modal displacement field ψ _i (x ), is directly determined by

where ρ( x ,t) = h!( x ,t) – f!( x ) is the difference between the filtered current image h!( x ,t) and the filtered reference image f!( x) .

Thus the simple "scalar product" of the extractor field ξ _i ( x ) associated with a modal displacement field ψ _i (x ), with the filtered difference of images directly gives the measurement of the current modal amplitude c _i (t) of the modal displacement field ψ _i (x ).

- a fourth sub-step S24 for determining the current residue ε( x, t).

Once the current modal amplitudes c _i (t) have been determined, it is possible to estimate a current residual ε( x, t) according to the calculation:

This calculation is an elementary operation which can be carried out in a simple manner by carrying out operations directly on the images.

It should be noted that, in all the methodologies presented above, step S2 is robust because it is based on a base of modal displacement fields which are extracted from a plurality of training images. The greater the number of learning images, the more the base of modal displacement fields characterizes in a complete and precise manner the kinematics of the structure during the learning phase. The decomposition of a current displacement field on this basis is then all the more relevant to compare the kinematics of the structure at the moment of acquisition, with the kinematics of the structure during the learning phase. Ultimately, this comparison makes it possible to characterize a state of health of the structure at the time of acquisition of the current image.

According to a step S3 of the monitoring phase, the processing unit generates a fault alert

- when one of the current modal amplitudes c _i (t) associated with a modal displacement field ψ _i ( x ) takes a value outside the range of expected values associated with the modal displacement field or
- when a norm of the current residual exceeds the residual threshold.

The range of expected values and the threshold of the residual were determined at the end of the learning phase as discussed previously.

For the current image, each current modal amplitude associated with a modal displacement field gives the weighting of an expected displacement and the current residual quantifies the part of the information contained in the current image which does not correspond to any expected displacement.

If a current modal amplitude is outside its range of expected values or if the current residual significantly exceeds the residual threshold, then an anomaly is detected which may signify a degraded state of health of the monitored structure.

The modal displacement fields make it possible to analyze the different types of displacements that are characteristic of the monitored structure and its state of health. If the state of health of the monitored structure is stable, the current modal amplitudes are within the expected value ranges. This means that new acquired images of the structure must be able to be decomposed similarly to the training images.

Two cases of alert can therefore arise:

- if the current residue significantly exceeds the residue threshold then the calculated current residue deviates significantly from the expected residue and this may signify the appearance of a new displacement whose signature has not been integrated into the base of expected displacements or even an amplitude of the expected displacement that is too large to be treated as an expected displacement. In these two cases, this may correspond to the appearance of a fault in the monitored structure;

- if one or more current modal amplitudes take a value outside its or their expected ranges, this may signify the appearance or aggravation of damage to the structure in the observed zone. For example, if a crack was initially present on the monitored structure, its expression in the deviation decomposition based on the modal displacement fields is integrated into one or more expected displacements. Its propagation or deepening over time leads to an increase in the amplitude of the mode(s) concerned.

Each image acquired provides information on the displacements within the structure and monitoring makes it possible to detect any deviation from the displacements normally observed within the structure. The unusual displacements detected are directly related to the degradation of the monitored structure.

The fault alert can be used to trigger the launch of additional investigations to further characterize the anomaly or fault detected. It should be noted that an analysis by CIN (without restriction to the modes) of the images having motivated the alert can allow, compared to the prior art, a finer diagnosis and more directly determine the nature of the mechanical damage from the results obtained. and acquired images. Such an analysis can also rule out false alarms, for example due to excessive stress that has not resulted in damage to the structure.

The monitoring method can be adapted to use a plurality of monitoring devices. Each device monitors a particular portion of the surface of the structure 9. It is then possible to more precisely characterize a state of health of the structure based on all the information given by all the monitoring devices. In particular, if the monitored structure is a wind turbine mast, monitoring devices can be deployed to monitor the area surrounding a weld between two ferrules.

It should be noted that the monitoring method does not require low levels of noise in the acquired images. Indeed, the determination of the displacement field as a combination of the modal displacement fields gives satisfactory results even with the use of a medium or low quality digital image acquisition device. This comes from the fact that the identification of the expected displacements is carried out globally over the whole of the study region (ie a set of very many pixels) because it is a question of Global Digital Image Correlation; that is to say that the quantity of information processed to carry out the decomposition is assessed on the scale of a large part of the image or even of the entire image.

The monitoring method also does not require a large amount of expensive hardware. The digital image acquisition devices used can be of average or low quality. The number of digital image acquisition devices to be used depends on the number of areas to be monitored over time. For example, if the monitored structure is a wind turbine mast, it is not necessary that the entire interior of the wind turbine mast be monitored. It is possible to limit monitoring to specific areas, such as the welds between two shells.

Finally, a computer program is proposed comprising instructions adapted to the implementation of at least one of the steps of the monitoring method as it was presented above, when said program is executed on a computer.

In the case where the monitored structure is a wind turbine mast, an alert generated at the level of the processing unit can be relayed to the control center of a plurality of wind turbines. To this end, the processing unit is connected by a communication channel to the control center.

Claims (12)

Method (P) for monitoring a structure by processing digital images, comprising:
(A) a learning phase comprising steps of:
(A2) from a plurality of training images of a portion of a surface of the structure, determination for each image of a training displacement field by an image correlation technique, the field being an estimate of a displacement of the material within the structure at each point of the portion at an instant of acquisition of the image,
(A3) from the fields, construction of a base of modal displacement fields characteristic of the plurality of images by a model reduction technique,
(A4) determination of a residual threshold characterizing a maximum admissible error of the displacement estimate and, for each modal displacement field, of a range of expected values of an associated modal amplitude characterizing a projection of the displacement fields learning on the modal displacement field,
(S) a monitoring phase comprising steps of:
(S1) acquisition of a current image of the portion,
(S2) from the current image, determination of a current displacement field, of current modal amplitudes characterizing a projection of the current displacement field on the basis of modal displacement fields and of a current residual, the field of current displacement being an estimate of a displacement of the material within the structure at each point of the portion at an instant of acquisition of the current image, the current residual characterizing an error in the estimate of the displacement, and
(S3) generation of an alert when one of the current modal amplitudes associated with one of the modal displacement fields takes a value outside the range of expected values of the modal amplitude associated with the modal displacement field or when a norm of the current residual exceeds the residual threshold. Monitoring method according to Claim 1, in which the step (S2) comprises a first substep (S2A) of determining the current displacement field by a global image correlation technique and a second substep (S2B) projection of the current displacement field on the basis of modal displacement fields. A monitoring method according to claim 1, wherein step (S2) comprises determining the current displacement field by an integrated image correlation technique, corresponding to a global image correlation technique using the base of fields of modal shifts. A monitoring method according to claim 1 wherein step (S2) comprises:
- a sub-step (S21) of filtering the reference image and the current image to obtain a filtered reference image and a filtered current image, by applying an inhomogeneous spatial filter of the low pass type, the filter using a plurality of filter feature lengths for a plurality of pixels forming the image, the feature length for a pixel depending on a maximum observed displacement associated with the pixel in the training displacement fields,
- a sub-step (S23) of determining, for each modal displacement field, a current modal amplitude of the current displacement field, from a scalar product between an extractor field associated with the modal displacement field and a difference between the filtered current image and the filtered reference image, the extractor field being an image associated with the modal displacement field constructed from the set of modal displacement fields and the filtered reference image, and
- a substep (S24) of determining the current residual from the filtered current image, the filtered reference image, the current modal amplitudes and the modal displacement fields. Monitoring method according to one of Claims 1 to 4, in which the structure comprises a wind turbine mast. A monitoring method according to claim 5, wherein the surface is inside the mast. Surveillance method according to claim 6, in which the interior zone of the mast extends on either side of a weld between two ferrules of the mast. Monitoring method according to one of Claims 1 to 7, in which the learning phase is implemented on different portions of the surface and then the monitoring phase is implemented on these portions. Device (1) for monitoring a structure by processing digital images, comprising an image acquisition device (3) and an image processing unit (8) connected to the device (3) and adapted for storing images in memory and implementing a monitoring method according to one of Claims 1 to 8. Wind turbine comprising a monitoring device according to the preceding claim. Wind turbine comprising a plurality of monitoring devices according to claim 9 arranged to implement the monitoring method according to claim 8. Computer program comprising instructions suitable for implementing one of the steps of the method according to any one of Claims 1 to 8 when said program is executed on a computer.