Method and system for detecting one or more anomalies in a structure

Abstract

The invention relates to a method (60) for detecting one or more anomalies in a structure, comprising: a learning phase (62) comprising the following steps: - obtaining (64) healthy data respectively associated with N distinct sets of use conditions of said structure; - projecting (66) the healthy data into a latent space with a dimension smaller than the dimension of said healthy data; - determining (68) the outline of said set of healthy data projected into said latent space; and an operational phase (70) comprising the following steps: - obtaining (74), through a preliminary measurement (72), test data representative of the current state of said structure; - projecting (76) said test data into said latent space; - detecting (78) at least one current anomaly in said structure as soon as a test data element is outside said outline.

Description

[0001] TITLE: METHOD AND SYSTEM FOR DETECTING ANOMALY(IES) IN A STRUCTURE

[0002] The present invention relates to a method for detecting anomaly(ies) in a structure and for monitoring the SHM health of said structure, an anomaly corresponding to modifications of the physical and / or geometric properties of the structure, said structure carrying at least one sensor for measuring at least one characteristic of said structure, the method comprising at least one learning phase and at least one operational phase.

[0003] The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement such a method of detecting anomaly(ies) in a structure.

[0004] The invention also relates to an electronic device for detecting anomaly(ies) in a structure, said structure carrying at least one sensor configured to generate and receive guided waves, the device comprising a learning unit and a test unit.

[0005] The invention also relates to a system for detecting anomaly(ies) in a structure.

[0006] The present invention relates to the field of structural health monitoring or SHM (from the English Structural Health Monitoring) aimed at detecting and characterizing damage (i.e. anomalies) of structures / infrastructures. Such structural anomalies correspond to modifications of the physical and / or geometric properties of the structure considered likely to affect its performance and / or reliability.

[0007] Such structural health monitoring is implemented using integrated sensors, for example piezoelectric sensors capable of emitting and receiving ultrasonic guided elastic waves.

[0008] The permanent integration of sensors on or within structures / infrastructures to monitor their condition makes it possible in particular to guarantee the safety of load-bearing structures corresponding to thin and / or long mechanical components, such as wind turbine blades, aircraft engine fuselages or components, metal or composite pipes, tension cables, bridge suspension cables, train rails, etc.

[0009] In particular, guided elastic ultrasonic waves emitted by such sensors make it possible to detect structural defects leading to discontinuities or variations in geometry, such as cracks, delaminations in composite fuselages, corrosion leading in particular to a loss of thickness in metals, etc., at an early stage and thus to monitor them for decades. Such detection is generally well mastered in the laboratory where external effects are limited, controlled and calibrated.

[0010] The main challenge associated with such a structural diagnosis is, in real conditions, linked on the one hand to the presence of epistemic uncertainties on the structure or instrumentation such as the positions of the sensors, the properties of the sensors, the elastic or geometric properties of the structure, and on the other hand linked to the presence of unsupervised (i.e. unknown) external influence effects evolving with distinct temporal dynamics.

[0011] Note that by "epistemic uncertainties" we mean uncertainties due to a lack of knowledge that can be reduced by trying to acquire more knowledge through, for example, data collection, expert assistance, accelerated testing, etc. Note that in the SHM context, reducing these uncertainties is generally prohibitive in terms of cost, hence the use of the term uncertainties.

[0012] Such unsupervised external influence effects can be classified on the one hand into so-called "fast" effects, presenting a variation on a time scale ranging from an hour to a day, such as the variation in temperature, humidity and stress (i.e. constraints exerted on a material), and on the other hand into so-called "slow" effects, presenting a variation on a time scale of the order of months, years, etc., such as the aging of sensors, the modification of material properties, of the structure or of the coupling between sensors and structure, and due to thermomechanical cycles or other.

[0013] In practice, the signatures of early defects are much smaller than the perturbations on the measurements associated with the presence of such epistemic uncertainties and / or the presence of such unsupervised influence effects. In other words, guided waves, in particular elastic and ultrasonic, suitable for detecting structural defects, via their signature, are not only sensitive to defects, but also to external and operational effects such as temperature, humidity, uncertainties on the sensors, aging, etc.

[0014] A classic solution aims to compensate for epistemic uncertainties by means of a reference state, i.e. a measurement in the absence of a fault and then comparing the current state to the reference state, in particular by subtraction, correlation, etc. while assuming that the only difference between the reference and current states can only be attributed to the presence of a fault, but that all other operating parameters, except at best a single rapid effect parameter such as temperature, are equal between these two reference and current states.

[0015] To compensate for rapid effects, such as temperature variation, signal processing is generally used. Such a conventional solution is limiting, often prohibitive in terms of experimental design, and likely to lead to the detection of a false positive or a false negative. Indeed, such a state-of-the-art solution first implies "perfect" instrumentation of the structure before the presence of a defect under identical operating conditions or at best within a factor such as temperature.

[0016] Temperature compensation is also limited in terms of the maximum compensable range, which is currently around a maximum temperature difference of 15°.

[0017] Furthermore, effects other than temperature, such as external forces, are more difficult to reproduce in the laboratory, and the simultaneous presence of several effects has not really been explored until now, nor has the evolution of the reference state over time, a reference state which is likely to change despite the absence of a fault.

[0018] In other words, conventional solutions do not take into account the aforementioned slow effects, and are implemented on the assumption that the system behaves nominally over long periods, which is of course not satisfactory when considering the instrumentation of a structure over several decades during which the aging of the measurement sensor and its coupling to the structure is inevitable and causes a slow drift in the signals received by the measurement sensor, which could be wrongly interpreted as a fault, or could hide one.

[0019] The aim of this invention is therefore to propose a method and a device capable of providing an early and reliable diagnosis of structural defects in real conditions of use in the presence of epistemic uncertainties and unsupervised external effects with both rapid dynamics (i.e. of the order of an hour to several days) and slow dynamics (i.e. of the order of a month to several years).

[0020] To this end, the subject of the invention is a method for detecting anomaly(ies) in a structure and for monitoring the SHM health of said structure, an anomaly corresponding to modifications of the physical and / or geometric properties of the structure, said structure carrying at least one sensor configured to generate and receive guided waves, the method comprising at least one learning phase and at least one operational phase, the learning phase comprising the following steps:

[0021] - obtaining a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct two by two, of conditions of use of said structure, N being an integer, two distinct sets having at least one condition of use distinct from one set to another, said obtaining of a set of healthy data being implemented via a plurality of Q measurement sensors carried by said structure forming a network of sensors, Q being an integer greater than one, at least one of said Q sensors being configured to generate and receive guided ultrasonic elastic waves;

[0022] - projection of said healthy data set into a latent space of reduced dimension relative to the dimension of said healthy data set;

[0023] - determination of the contour of said healthy data set projected into said latent space; the operational phase comprising the following steps:

[0024] - obtaining, by prior measurement via said at least one sensor, a set of test data representative of the current state of said structure;

[0025] - projection of said test data set into said latent space;

[0026] - detection of at least one current anomaly of said structure as soon as an element of said test data set is outside said contour.

[0027] This method aims to quantify the extent to which signal data, corrupted over various time scales, can be used for the analysis of structural states. Such a method thus makes it possible to compensate for the influence of the aforementioned supervised and / or unsupervised external effects without erasing the signature of the small defects that one seeks to detect.

[0028] More specifically, the invention consists in obtaining all the data of the signals likely to be measured in real healthy conditions, said data being subsequently called healthy data, that is to say in the absence of a defect, with the aim of detecting a defect by its absence from this set. In other words, such obtaining amounts to constructing a digital twin describing all the healthy signals to then detect an anomaly as beyond the contour of the "healthy" space described by this twin. Such a "healthy" space obtained according to the present invention is an important and differentiating element compared to conventional solutions. Indeed, according to the present invention, no attempt is made to model defects, because these will be detected by their absence from this set of healthy data. Not modeling the defects is a significant advantage because it amounts to not making assumptions related to the typology, position or size of the defects.

[0029] Since there are an infinite number of fault-free signals, the present invention further proposes to use a dimensionality reduction, via a latent space, to describe the characteristics of the signals in a healthy state. According to other advantageous aspects of the invention, the method for detecting anomaly(ies) in a structure comprises one or more of the following characteristics, taken individually or in all technically possible combinations:

[0030] - said set of healthy data is obtained following a preliminary phase of calibration of said structure for the N distinct sets two by two of conditions of use of said structure;

[0031] - said set of healthy data is obtained following a preliminary phase of simulation of said structure for the N distinct sets two by two of conditions of use of said structure;

[0032] - said set of healthy data is obtained following a preliminary hybrid phase of calibration and / or simulation of said structure for the N sets, distinct two by two, of conditions of use of said structure;

[0033] - the method comprises, during said preliminary hybrid phase, a compensation step by transfer learning in the event of a difference between calibration and simulation for the same set of conditions of use of said structure;

[0034] - said reduced-dimensional latent space is obtained by supervised or unsupervised dimensional reduction;

[0035] - said unsupervised dimensional reduction is implemented by means of one of the elements belonging to the group comprising at least:

[0036] - a principal component analysis,

[0037] - an auto-encoder previously trained to compress and then decompress the signals from the healthy data set, and of which only the part dedicated to compression is used to implement said dimensional reduction.

[0038] - an autoregressive process;

[0039] - said supervised dimensional reduction is implemented by means of a neural network;

[0040] - said neural network is a neural network whose type belongs to the group comprising:

[0041] - a convolutional neural network;

[0042] - a recurrent neural network;

[0043] - a multi-layer perceptron;

[0044] - said determination of contour consists of:

[0045] - search, in said latent space, for the spherical or elliptical envelope of minimal radius(es) encompassing the points of said healthy data set projected into said latent space, or - search in the latent space for the hyperplane furthest from the origin which separates, from the origin, the points of said healthy data set projected into said latent space, or

[0046] - use other types of anomaly detection such as robust estimation of the covariance matrix, known isolation forest, or even with outlier detection with a local outlier factor.

[0047] The invention also relates to a computer program comprising software instructions which, when executed by a computer, implement a method for detecting anomaly(ies) in a structure as defined above.

[0048] The invention also relates to a system for detecting anomaly(ies) in a structure and for monitoring the SHM health of said structure, an anomaly corresponding to modifications of the physical and / or geometric properties of the structure, said system comprising a plurality of Q sensors for measuring at least one characteristic of said structure carried by said structure and forming a network of sensors, Q being an integer greater than one, at least one of said Q sensors being configured to generate and receive guided ultrasonic elastic waves, said system comprising a device in said structure and comprising a learning unit and a test unit, the learning unit comprising:

[0049] - a first obtaining module configured to obtain, via said plurality of Q measurement sensors, a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct two by two, of conditions of use of said structure, N being an integer, two distinct sets having at least one condition of use distinct from one set to another;

[0050] - a first projection module configured to project said set of healthy data into a latent space of reduced dimension relative to the dimension of said set of healthy data;

[0051] - a determination module configured to determine a contour of said healthy data set projected into said latent space; the test unit comprising:

[0052] - a second obtaining module configured to obtain, by measurement via said at least one sensor, a set of test data representative of the current state of said structure;

[0053] - a second projection module configured to project said test data set into said latent space provided by the projection module of said learning unit;

[0054] - a detection module configured to detect at least one current anomaly of said structure as soon as an element of said test data set is outside said contour provided by said determination module of said learning unit. These characteristics and advantages of the invention will appear more clearly on reading the description which follows, given solely by way of non-limiting example, and made with reference to the appended drawings, in which:

[0055] [Fig 1] Figure 1 illustrates the effect of a structural defect on the propagation of guided waves on the surface of a structure.

[0056] [Fig 2] Figure 2 is a schematic view of an electronic device for detecting anomaly(ies) in a structure according to the present invention;

[0057] [Fig 3] Figure 3 is a flowchart of a method for detecting anomaly(ies) in a structure according to the present invention;

[0058] [Fig 4] [Fig 5] Figures 4 and 5 illustrate respectively the learning phase and the operational phase of the method according to the present invention.

[0059] In Figure 1, an example of propagation of guided waves on the surface of a structure 10 is illustrated for the structure without a defect in view A and in the presence of a defect in view B.

[0060] More specifically, in view A of this example, the defect-free structure 10 corresponds to a defect-free cylindrical pipe, of length L, and carrying on its surface at least one measuring sensor in an uncontrolled environment, for example configured to generate, via the element 12 and receive via the element 14, guided waves 16, in particular elastic and ultrasonic. For example, such a sensor belongs to the group comprising at least: piezoelectric sensors, an electro-magneto-acoustic transducer EMAT (from the English Electro magneto acoustic transducer), a polyvinylidene fluoride PVDF sensor (from the English Polyvinylidene fluoride), etc. According to a hybrid instrumentation, reception via the element 14 can also be ensured by one of the fiber optic Bragg gratings.

[0061] Other types of sensors are suitable for use as long as they are capable of carrying out a measurement in an uncontrolled environment, for example sensors based on the use of ultrasound in general, including, as mentioned above, sensors suitable for generating and receiving guided waves, but also sensors for measuring eddy currents or vibration measurements, etc. In view B, the presence of a defect 18 on the surface of the pipe modifies the propagation of the guided waves by generating, in particular, reflected guided waves 20.

[0062] Figure 2 is a schematic view of an electronic device 30 for detecting anomaly(ies) in a structure according to the present invention. Such a device 30 comprises a machine learning unit 32 comprising a first obtaining module 34 configured to obtain a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct two by two, of conditions of use of said structure, N being an integer, two distinct sets having at least one condition of use distinct from one set to another.

[0063] The learning unit 32 further comprises a first projection module 36 configured to project said healthy data set (also called healthy base comprising N samples, each sample being of dimension P) into a latent space E of reduced dimension compared to the dimension of said healthy data set.

[0064] According to an optional aspect, such a latent space E is configured to describe as simply as possible the aforementioned perturbative effects, and is ideally linear, if not monotonic.

[0065] The learning unit 32 also comprises a determination module 38 configured to determine a contour C, in particular multidimensional, of said set of healthy data projected into said latent space.

[0066] Furthermore, the electronic device 30 comprises a test unit 40, capable of receiving as input the output S of the learning unit 32, and comprising a second obtaining module 42 configured to obtain, by measurement via said at least one sensor, a set of test data representative of the current state of said structure 10.

[0067] The test unit 40 further comprises a second projection module 44 configured to project said test data set into said latent space provided by the first projection module 36 of said learning unit 32.

[0068] The test unit 40 also comprises a detection module 46 configured to detect at least one current anomaly of said structure as soon as an element of said test data set is outside said contour provided by said determination module 38 of said learning unit 32.

[0069] In other words, the learning unit 32 is capable of constructing by training a digital twin describing all of the healthy data to detect an anomaly as well as beyond the healthy space described by this digital twin.

[0070] In the example of Figure 2, the electronic device 30 for detecting anomaly(ies) in a structure comprises an information processing unit 50 formed for example of a memory 52 and a processor 54 associated with the memory 52.

[0071] In the example of Figure 2, the first obtaining module 34, the first projection module 36, the determination module 38, the second obtaining module 42, the second projection module 44 and the detection module 46 are each produced in the form of software, or a software brick, executable by the processor 54. The memory 52 of the electronic device 30 for detecting anomaly(ies) in a structure is then capable of storing, for implementing a learning phase, a first obtaining software, a first projection software, and a determination software, and for implementing a test phase, a second obtaining software, a second projection software, and a detection software.The processor 54 is then able to execute each of the software programs among the first obtaining software program, the first projection software program, the determination software program for a learning phase, and for a test phase, a second obtaining software program, a second projection software program, and a detection software program.

[0072] In a variant not shown, the first obtaining module 34, the first projection module 36, the determination module 38, the second obtaining module 42, the second projection module 44 and the detection module 46 are each produced in the form of a programmable logic component, such as an FPGA (Field Programmable Gate Array), or a GPU (Graphics Processing Unit), or in the form of an integrated circuit, such as an ASIC (Application Specific Integrated Circuit).

[0073] When the electronic device 30 for detecting anomaly(ies) in a structure is produced in the form of one or more software programs, that is to say in the form of a computer program, also called a computer program product, it is also capable of being recorded on a medium, not shown, readable by a computer. The computer-readable medium is, for example, a medium capable of storing electronic instructions and of being coupled to a bus of a computer system. For example, the readable medium is an optical disk, a magneto-optical disk, a ROM memory, a RAM memory, any type of non-volatile memory (for example EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card. A computer program comprising software instructions is then stored on the readable medium.

[0074] In a manner not shown, the invention also relates to a system for detecting anomaly(ies) in a structure, said system comprising such an electronic device 30 for detecting anomaly(ies) in a structure, illustrated by FIG. 2, and further comprising a plurality of Q sensors carried by said structure, configured to generate and receive guided waves and forming a sensor network, Q being an integer greater than one. According to a particular aspect, at least one of said Q sensors is configured to generate and receive guided elastic ultrasonic guided waves.

[0075] The information is therefore, in this case, processed at the sensor network level by exploiting Q 2 -Q signals corresponding to the paths between the Q sensors, or Q 2signals if the signals emitted and received by the same sensor are used. The sensor network Q is, if necessary, supplemented by an acquisition chain (not shown) and configured to process the measurements captured via said sensor network Q. Indeed, it is not excluded that a signal acquired by a sensor element 14 as illustrated in FIG. 1, in the presence of a fault under operating conditions Ui, is identical to a signal acquired in the absence of a fault under operating conditions Us. To remove this ambiguity, the present invention then proposes to rely on the data from a sensor network. Indeed, due to the asymmetry of the faults, the contribution of a given fault is variable depending on the paths traveled by the waves, which is not the case for operational conditions. The data is therefore processed at the sensor network level, Q 2 -Q signals corresponding to the paths between the Q sensors.

[0076] The operation of the electronic device 30 for detecting anomaly(ies) in a structure will now be described with reference to FIG. 3 which schematically illustrates an example of implementation, according to the present invention, of a method 60 for detecting anomaly(ies) in a structure, such as for example the structure 10 of FIG. 1.

[0077] The method 60 according to the present invention firstly comprises an automatic learning phase 62, also illustrated in more detail in FIG. 4 described below.

[0078] According to a first step 64 of the learning phase 62, the electronic device 30 for detecting anomaly(ies), via its first obtaining module 34, obtains OBT A a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct two by two, of conditions of use of said structure, N being an integer, two distinct sets presenting at least one condition of use distinct from one set to another.

[0079] The N sets, distinct two by two, of conditions of use of said structure aim in particular to cover the main operating parameters likely to influence the measurements, namely those associated with the aforementioned epistemic effects, such as the placement of sensors, the properties and geometry of the structure, the properties of the sensors, those associated with the aforementioned rapid effects, such as temperature, humidity, pressure, the forces applied to the structure considered, those associated with slow effects such as the aging of the sensor(s), of the structure and of the coupling between the structure and the sensor(s).

[0080] As illustrated by Figure 2, such obtaining 64 follows one of the three optional preliminary phases 642, 644, 646, implemented by said electronic device 30 for detecting anomaly(ies) or by a separate device, the first optional preliminary phase 642 corresponding to a preliminary phase of calibration CAL of said structure for the N distinct sets two by two of conditions of use of said structure. In other words, according to this first option, the set of healthy data is determined beforehand by experience during a calibration phase 642, often likely to be prohibitive in terms of costs / duration, and consisting of instrumenting healthy structures and acquiring data over long periods.Such a calibration phase 642 involves both the supervision of the measured data to ensure that there are no defects, and their exhaustiveness in order to ensure that a statistically sufficient number of combinations of the parameters are explored. Depending on the application, such a calibration is available and carried out beforehand, for example during a research and development phase of the structure considered, the obtaining step 64 then consisting of recovering the set of healthy data resulting from the calibration phase 642 CAL.

[0081] According to a second preferred option, said set of healthy data is obtained at the end of a preliminary phase 644 of SIM simulation of said structure for the N distinct sets two by two of conditions of use of said structure, the obtaining step 64 then consisting of recovering the set of healthy data resulting from the phase 644 of SIM simulation.

[0082] This second option is a preferred approach, according to the present invention, the prior simulation 644 making it possible to establish, quickly and at a lower cost than the first experimental option, models (i.e. simulations) of the signals subject to the disturbing parameters corresponding to the epistemic uncertainties and the aforementioned unsupervised external effects, and thus to generate healthy data on the fly.

[0083] Indeed, depending on the application, it may prove prohibitive to acquire all possible healthy signals in experience, in particular on the one hand because the aging effects are only visible after several years, and on the other hand because it would be necessary to instrument many structures to observe all the variabilities of structures / instrumentation.

[0084] Such a second option by simulation also implies having available models of all the phenomena influencing the guided waves, such models being able to be learned / calibrated / tested / validated on samples representative of the structure and the application considered, greatly limiting the cost compared to the experimental solution associated with the first option mentioned above. Advantageously, the models (i.e. simulations) of defects are not necessary, because the present invention proposes to focus on healthy data.

[0085] According to a third option, said set of healthy data is obtained following a prior hybrid phase 646 of calibration and / or simulation of said structure for the N sets, distinct two by two, of conditions of use of said structure, the obtaining step 64 then consisting of recovering the set of healthy data resulting from the hybrid phase 646 HYB.

[0086] According to a particular optional aspect of this third option, said hybrid preliminary phase comprises a step, not shown, of compensation by transfer learning in the event of a difference between calibration and simulation for the same set of conditions of use of said structure. In other words, such an optional aspect is notably implemented in the case where the models (i.e. simulations) are not perfect, i.e. without rigorous superposition between the simulated and experimental signals with identical parameters, and proposes to compensate for the difference by transfer learning, an example of which is described in patent application FR 3 1 13 530, such “transfer learning” making it possible in a hybrid manner to recalibrate the models using an affordable number of experiments.

[0087] Regardless of the option implemented prior to the obtaining step 64, as indicated previously, it should be noted that for each option, a system for detecting anomaly(ies) in a structure is considered, said system comprising the electronic device 30 for detecting anomaly(ies) in a structure, illustrated by FIG. 2, and further comprising a plurality of Q sensors carried by said structure, configured to generate and receive guided waves and forming a network of sensors, Q being an integer greater than one, and N being such that N=Q 2 -Q. The network of Q sensors makes it possible to ensure that the data set recovered during the obtaining step 64 is healthy.

[0088] Indeed, external effects are relatively homogeneous and perceived by all sensors. This is the case, for example, of temperature, which tends to affect all measurement sensors relatively homogeneously. On the other hand, a defect will only influence certain sensors and in quite different ways, due to its limited size and natural asymmetry.

[0089] Furthermore, independently of the option implemented prior to the obtaining step 64, the learning phase 62 of the method 60 according to the present invention further comprises a step 66 of projection PA of said set of healthy data into a latent space E of reduced dimension compared to the dimension of said set of healthy data. Indeed, the set of healthy data is generally of large dimension, in particular of the order of 10 3 at 10 5 which is likely to slow down anomaly detection during the operational testing phase.

[0090] Projection step 66 therefore aims to reduce the dimensions of the healthy learning data, by describing them (i.e. projecting) into a latent space of reduced dimension, for example of the order of 10 to 10 2 for original dimensions of 10 respectively 3 at 10 5 .

[0091] Said reduced-dimensional latent space is obtained by supervised or unsupervised dimensional reduction.

[0092] It should be noted that since healthy data do not contain defects, an unsupervised 66-dimensional reduction allows for no unfavorable a priori for defect detection. In other words, an unsupervised 66-dimensional reduction is suitable for implementation to avoid creating a bias against the signature of defects, absent from the initial set of healthy data.

[0093] According to a first option, when the dimensional reduction 66 is unsupervised, it is implemented by means of one of the elements belonging to the group comprising at least:

[0094] - a principal component analysis (PCA),

[0095] - an auto-encoder previously trained to compress and then decompress the signals from the healthy data set, and of which only the part dedicated to compression is used to implement said dimensional reduction.

[0096] - an autoregressive process

[0097] In other words, the first projection module 36 is capable of implementing a principal component analysis, in particular ensuring that the loss of variance is less than the influence of a defect subsequently sought during the operational phase. In the case of the implementation of an auto-encoder, the first projection module 36 is configured to drive the first half of said auto-encoder to compress / decompress the signals, and the dimensional reduction is obtained by retaining only the central component of the auto-encoder commonly called "embedding".

[0098] According to a second option, when the dimensional reduction 66 is supervised, in particular to best adapt a set of healthy data resulting from a prior simulation, the physics of which is well understood, and to promote the detection of defects, such a supervised dimensional reduction 66 is implemented by means of a neural network, the type of which belongs for example to the group comprising:

[0099] - a convolutional neural network;

[0100] - a recurrent neural network;

[0101] - a multi-layer perceptron. Such a neural network is particularly suitable for modeling data and disturbing effects such as temperature, humidity, etc., the reduced-dimensional latent space then corresponding to all the outputs of the neural network.

[0102] In addition, as illustrated by Figure 2, the learning phase 62 of the method 60 according to the present invention further comprises a step 68 of determining the contour C, implemented automatically via said determination module 38 or manually, of said set of healthy data projected into said latent space E. In other words, during said step 68, one seeks, by training, to describe the contours of the set of healthy data, in order subsequently, during an operational phase, to detect an anomaly outside of said contour.

[0103] To do this, as described in particular by V. Chandola et al. in the publication entitled “Anomaly detection: A survey”, ACM Computing Surveys Volume 41 Issue 3 July 2009 Article No.: 15pp 1-58, said contour determination 68 is for example suitable for comprising:

[0104] - the search, in said latent space E, for the envelope, in particular spherical or elliptical, of minimal radius(es) encompassing the points of said set of healthy data projected into said latent space, according to the technique known under the English name of “support vector data description”, or

[0105] - the search in the latent space E for the hyperplane furthest from the origin which separates, from the origin, the points of said set of healthy data projected into said latent space, according to the technique known under the English name of “one class support vector machine”, or

[0106] - use other types of anomaly detection such as robust covariance estimation, isolation forest, or outlier detection with the local outlier factor (LOF).

[0107] Once the learning phase 62 has been carried out, the outputs S, comprising the latent space E and the contour C, of ​​the learning unit 32, are transmitted to the test unit 40 suitable for implementing an operational phase 70, repeated after each step 72 of measurement M of a current state of the structure considered. As indicated previously, each measurement 72 is in particular obtained using a plurality of Q sensors carried by said structure, configured to generate and receive guided waves and forming a network of sensors, Q being an integer greater than one, and N being such that N=Q 2 -Q.

[0108] It should be noted that the learning phase 62 is suitable for being repeated in particular to take into account additional parameters suitable for influencing the propagation of the guided waves depending on the application of the desired structure, for example in the event of a change in the climatic zone of operation of said structure.

[0109] Furthermore, according to an aspect not shown, the present invention makes it possible to take into account the knowledge of a physical parameter of the condition of use at the time of the measurement 72, such a physical parameter corresponding in particular to the temperature, to reduce the set of healthy data resulting from the learning 62 to a set of healthy data restricted by one dimension due to this known physical parameter.

[0110] In other words, in this case the learning 62 is repeated to obtain said restricted healthy data set, and its contour C' in a latent space E' associated with this dimension reduction linked to the perfect knowledge of a usage parameter such as the temperature during the measurement 72. Such a contour C' is necessarily more restricted than the contour C associated with the healthy data set obtained for a plurality of distinct operating temperatures of the structure. Such a more restricted contour C' makes the approach more reliable, and in the extreme case where all the influential parameters would be known, which does not seem possible in practice, due to epistemic uncertainties, the healthy data set would then contain a single data set.

[0111] Such an operational phase 70, also illustrated and described below in relation to FIG. 5, comprises a first step 74 of obtaining OBT_T, via the second obtaining module 42 of FIG. 2, a set of test data representative of the current state of said structure considered, in which the presence of a defect is unknown.

[0112] Then, during a step 76, implemented by the second projection module 44 of FIG. 2, the set of test data representative of the current state of said structure considered is projected, via a projection P_T, into the latent space E, used during the learning phase 62.

[0113] Then, during a step 78, implemented automatically by the detection module 46 of FIG. 2 or manually, at least one current anomaly of said structure is detected as soon as an element of said test data set is outside said contour C.

[0114] In a manner not shown, the method 60 is also capable of comprising an evaluation of the biases and errors due to modeling and machine learning in order to provide a measure of confidence in the fault diagnosis (i.e. the anomaly detection) proposed according to the present invention.

[0115] Figure 4 illustrates, schematically, in more detail the learning phase 62 previously described, with step 64 of obtaining a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct two by two, of conditions of use of said structure, N being an integer, two distinct sets presenting at least one condition of use distinct from one set to another.

[0116] More precisely, as illustrated in FIG. 4, during this obtaining step 64, the healthy data associated with as many signals N as there are distinct sets of usage conditions of a healthy structure without defects are obtained. In other words, each of the N signals represented is obtained with a variation of at least one usage parameter such as the position of at least one sensor, the temperature and / or the aging of said at least one sensor.

[0117] Step 66 of projecting said healthy data set into a latent space of reduced dimension M relative to the dimension N of said healthy data set is illustrated in Figure 4 by M 2 two-dimensional representations of healthy data represented by crosses.

[0118] It should be noted that such two-dimensional representations are used here only for illustrative and explanatory purposes of the multidimensional aspect of the latent space, the multidimensional representation of which is complex.

[0119] Step 68 of determining the contour of said healthy data set (also called healthy base comprising N samples, each sample being of dimension P) projected into said latent space is illustrated in Figure 4 by the contours Ci and C2, the contour Ci surrounding the crosses of healthy data projected during step 66 into a two-dimensional part with dimension 1 on the abscissa and dimension 2 on the ordinate of the latent space of M dimensions (i.e. of size M <P nécessairement du fait de la réduction de dimension), le contour C2 entourant les croix de données saines projetée lors de l’étape 66 dans une partie bidimensionnelle avec la dimension 1 en abscisse et la dimension 3 en ordonnée de l’espace latent à M dimensions.

[0120] Figure 5 illustrates, schematically, in more detail the operational phase previously described, with step 74 of obtaining, by prior measurement via said at least one sensor, a set of test data representative of the current state of said structure.

[0121] Step 76 of projecting said test data set into said latent space, resulting from the learning 62 illustrated by figure 4, is represented in figure 5 by M 2 two-dimensional representations of the T1 and T2 test data represented by a cross in each two-dimensional representation.

[0122] The anomaly detection step 78 compares, in each two-dimensional representation, that the test data cross is in the associated healthy data contours, otherwise an anomaly 82 is detected as soon as a cross is outside a two-dimensional contour, which is the case for the cross T 1 which is outside the contour Ci in the two-dimensional space of dimensions 1 and 2, whereas when dimensions 1 and 3 are considered the cross T2 is in the contour C2. In other words, as soon as a test data position outside the multidimensional contour is detected, in particular here for explanatory purposes and simplicity of representation within the M 2 two-dimensional representations, an anomaly is detected, and if necessary in a manner not shown, an alert not shown is raised and / or a maintenance request for said structure.

[0123] In other words, such detection 78 is certainly crude but effective, and consists of considering that if the representation in the latent space of the test data set is too far from the contour C of the healthy data set, a defect is present.

[0124] According to a particular aspect, very dependent on the use case considered, the distance from the test data representative of an anomaly (here the cross T1) to the contour here Ci could be suitable for translating the criticality of the detected defect.

[0125] Those skilled in the art will understand that the invention is not limited to the embodiments described, nor to the particular examples of the description, the embodiments and variants mentioned above being suitable for being combined with each other to generate new embodiments of the invention.

[0126] Thus, the present invention proposes a method and a device for detecting anomaly(ies) in a structure which make it possible to construct, by training, a digital twin describing all of the healthy data associated with a healthy state of the structure considered in order to then detect an anomaly as beyond the healthy space described by this digital twin.

[0127] In other words, the present invention makes it possible to avoid the creation of defects, whether experimentally or by simulation, by learning by modeling only healthy states.

[0128] According to one aspect of the invention implementing the obtaining of healthy data at least in part by prior simulation, it is possible to dispense with a prior calibration phase on real structures and in real usage conditions, or to greatly limit it in the case of hybrid prior processing involving at least in part simulation.

[0129] Furthermore, the present invention is generic with respect to the influential parameters, because once the healthy data set is obtained by learning, it is possible to quickly evolve it to take into account additional conditions of use of the structure and to recalculate the multidimensional contour in the latent space. Thus, the anomaly detection obtained via the present invention has intrinsic robustness to external parameters by integrating them from the start in the learning of healthy data, so that the diagnosis (i.e. the detection) does not make any a priori assumptions about the external conditions at the time of its application to the measurement of a current state of the structure.

[0130] Such anomaly detection is likely to reduce the environmental footprint of instrumented structures through preventive maintenance and extending the lifespan by detecting anomalies (i.e. faults) at an early stage.

Claims

DEMANDS 1. Method (60) for detecting anomaly(ies) in a structure and for SHM health monitoring of said structure, an anomaly corresponding to changes in the physical and / or geometric properties of the structure, said structure having at least one sensor for measuring at least one characteristic of said structure, the method comprising at least one learning phase (62) and at least one operational phase (70), the learning phase (62) comprising the following steps: - obtaining (64) a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct in pairs, of conditions of use of said structure, N being an integer, two distinct sets presenting at least one condition of use distinct from one set to another, said obtaining of a set of healthy data being implemented via a plurality of Q measurement sensors carried by said structure forming a sensor network, Q being an integer greater than one, at least one of said Q sensors being configured to generate and receive guided ultrasonic elastic waves; - projection (66) of said healthy data set into a latent space of reduced dimension compared to the dimension of said healthy data set; - determination (68) of the outline of said healthy data set projected into said latent space; the operational phase (70) comprising the following steps: - obtaining (74), by prior measurement (72) via said at least one sensor, a set of test data representative of the current state of said structure; - projection (76) of said test data set into said latent space; - detection (78) of at least one current anomaly of said structure as soon as an element of said test dataset is outside said contour.

2. Method (60) according to claim 1, wherein said healthy data set is obtained following a prior calibration phase (642) of said structure for the N distinct pairs of usage conditions of said structure.

3. A method (60) according to claim 1, wherein said healthy data set is obtained following a preliminary simulation phase (644) of said structure for the N distinct pairs of conditions of use of said structure.

4. Method (60) according to claim 1, wherein said healthy data set is obtained following a prior hybrid calibration and / or simulation phase (646) of said structure for the N pairs distinct sets of conditions of use of said structure.

5. Method (60) according to claim 4, comprising, during said hybrid preliminary phase, a transfer learning compensation step in case of discrepancy between calibration and simulation for the same set of conditions of use of said structure.

6. Method (60) according to any one of the preceding claims, wherein said reduced dimension latent space is obtained by supervised or unsupervised dimensional reduction.

7. A method (60) according to claim 6, wherein said unsupervised dimensional reduction is implemented by means of one of the elements belonging to the group comprising at least: - a principal component analysis, - an autoencoder trained beforehand to compress and then decompress the signals from the healthy dataset, and of which only the part dedicated to compression is used to implement said dimensionality reduction. - an autoregressive process.

8. Method (60) according to claim 6, wherein said supervised dimensional reduction is implemented by means of a neural network.

9. Method (60) according to claim 8, wherein said neural network is a neural network whose type belongs to the group comprising: - a convolutional neural network; - a recurrent neural network; - a multilayer perceptron.

10. A method (60) according to any one of the preceding claims, wherein said contour determination (68) consists of: - search, within said latent space, for the spherical or elliptical envelope of minimum radius(s) encompassing the points of said healthy data set projected into said latent space, or - search in the latent space of the hyperplane furthest from the origin that separates, from the origin, the points of said healthy data set projected into said latent space, or - use other types of anomaly detection such as robust estimation of the covariance matrix, known isolation forest, or outlier detection with a local aberration factor.

11. Computer program comprising software instructions which, when executed by a computer, implement, at least in part, a method for detecting anomaly(ies) in a structure according to any one of the preceding claims.

12. Anomaly detection system in a structure and SHM health monitoring of said structure, an anomaly corresponding to changes in the physical and / or geometric properties of the structure, said system comprising a plurality of Q sensors measuring at least one characteristic of said structure carried by said structure and forming a sensor array, Q being an integer greater than one, at least one of said Q sensors being configured to generate and receive guided ultrasonic elastic waves, said system comprising a device (30) in said structure and comprising a learning unit (32) and a test unit (40), the learning unit (32) comprising: - a first acquisition module (34) configured to obtain, via said plurality of Q measurement sensors, a set of healthy data representative of N healthy states of said structure respectively associated with N sets, distinct in pairs, of conditions of use of said structure, N being an integer, two distinct sets presenting at least one condition of use distinct from one set to another; - a first projection module (36) configured to project said healthy data set into a latent space of reduced dimension compared to the dimension of said healthy data set; - a determination module (38) configured to determine an outline of said healthy dataset projected into said latent space; the test unit (40) comprising: - a second acquisition module (42) configured to obtain, by measurement via said at least one sensor, a set of test data representative of the current state of said structure; - a second projection module (44) configured to project said test dataset into said latent space provided by the projection module of said learning unit; - a detection module (46) configured to detect at least one common anomaly of said structure as soon as an element of said test dataset is outside said contour provided by said determination module of said learning unit.

Images