Method and system for detecting anomalies of mechanical component, in particular aircraft component, by classifying spectrogram of acoustic signal

Abstract

A method implemented by a computer (16) in order to detect anomalies of an unknown component, comprising: determining (314) a plurality of areas (9) of the unknown component and performing at least once the following steps: generating (500, 502, 504) a spectrogram relating to an acoustic signal generated by striking a portion (6) of one area (9) of the unknown component; selecting (507), among a plurality of binary classifiers (70) respectively associated with a corresponding region (9) of the unknown component, a binary classifier (70) associated with the hit region, each of the plurality of binary classifiers (70) classifying, in two classes, a spectrogram associated with an acoustic signal generated by hitting the corresponding region, respectively, said two classes respectively indicate a spectrogram relating to an acoustic signal generated by striking an undamaged or damaged version of said corresponding area (9); performing (508), by means of the selected binary classifier, a classification of the spectrogram among one of the corresponding two classes; and detecting (510) the presence of an anomaly in the hit area of the unknown component based on the classification performed by the selected binary classifier.

Description

[0001] Cross - reference to related applications

[0002] This patent application claims the priority of European Patent Application No. 22209811.3, filed on November 28, 2022, the entire disclosure of which is incorporated herein by reference. Technical field

[0003] The present invention relates to a method and a system for detecting anomalies of mechanical components, in particular aircraft components, by classifying spectrograms of acoustic signals. Background art

[0004] It is well known that in the aviation field, there is a particular need to detect the presence of anomalies (understood as damage or defects) of aircraft mechanical components in order to ensure flight safety. To this end, for example, so - called non - destructive controls are known, which allow the general condition of the components of an aircraft to be evaluated in a relatively short time.

[0005] For example, some non - destructive controls provide for the study of possible anomalies of the components of an aircraft, which are carried out by highly specialized personnel performing visual and / or acoustic inspections of the components.

[0006] In particular, in the case of acoustic inspection, the component to be inspected is repeatedly struck with a mechanical striking tool (e.g., a hammer made of aluminum) in order to generate an acoustic response to the strike. Based on this acoustic response perceived by the ear, the person in charge of the inspection can detect, based on his / her experience, the possible presence of anomalies in the component (e.g., the presence of unbonded areas or delamination), such as in a part of the fuselage or a blade of a helicopter.

[0007] Therefore, acoustic inspection, also known as the tap test, requires the presence of trained personnel, providing the corresponding technical preparation and significant practical experience. In addition, such a process cannot be automated and is inevitably affected by uncertainties and human errors related to the capabilities of the personnel performing it. To this end, for example, human factors (fatigue, distraction, etc.) or environmental conditions (e.g., the presence of background noise) may have a negative impact on the ability of the person in charge of the inspection to detect anomalies.

[0008] The document "Defect detection with estimation of material condition using ensemble learning for hammering test" by H. Fujii et al., 2016 IEEE International Conference on Robotics and Automation (IRCA), Stockholm, May 16 - 21, 2016, pages 3847 - 3854 discloses a method for detecting material defects, which includes implementing a plurality of detectors of the weak learner type, each detector processing a corresponding frequency sub - band and analyzing the hammering sound.

[0009] US2008 / 0144927A1 relates to a non - destructive inspection device, which includes a sensor unit for detecting vibrations transmitted through a test object and a signal input unit for extracting a target signal from the electrical signals output by the sensor unit; in addition, the device includes a single neural network configured to classify a set of characteristics including a plurality of frequency components extracted from the target signal. Summary of the Invention

[0010] Therefore, an object of the present invention is to provide a solution that at least partially overcomes the disadvantages of the prior art.

[0011] According to the present invention, there is provided a method and a system for detecting anomalies as defined in the appended claims. Brief Description of the Drawings

[0012] To better understand the present invention, embodiments thereof will now be described by way of non - limiting examples only with reference to the accompanying drawings, in which:

[0013] - Figure 1 shows a front view of a part of a component of an aircraft;

[0014] - Figure 2 shows a block diagram of a system for detecting anomalies;

[0015] - Figure 3 shows a block diagram of a convolutional neural network;

[0016] - Figure 4 , Figure 9 and Figure 11 shows a flowchart of a training operation according to the present method;

[0017] - Figure 5 showsFigure 1 Enlarged front view of a part of the component shown;

[0018] - Figure 6 Shows a representation of a spectrogram;

[0019] - Figure 7 Shows a flowchart of operations for generating a spectrogram;

[0020] - Figure 8 Shows a schematic of the structure of a spectrogram;

[0021] - Figure 10 Shows a schematic of a confusion matrix; and

[0022] - Figure 12 Shows a flowchart of operations according to the present method. Detailed Description of the Invention

[0023] By way of example, the present method for detecting anomalies in components of an aircraft is described with reference to the component 1 shown in Figure 1 and the detection system 10 shown in Figure 2 , the detection system 10 being shown as operating, for example, on the component 1; as will be explained in detail below, the detection system 10 includes a striking device 12, a microphone 14, and a computer 16.

[0024] As described above, the present method for detecting anomalies provides a multi-class classifier 50 (shown in Figure 3 ) and a plurality of classifiers 70 (one shown in Figure 3 ), the multi-class classifier 50 being hereinafter referred to as the zone classifier 50, and the plurality of classifiers 70 being hereinafter referred to as the binary classifiers 70 because they are configured to classify two classes, as will be explained below.

[0025] The zone classifier 50 and the binary classifiers 70 can be trained in the manner described with reference to Figure 4 .

[0026] Specifically, the training of the zone classifier 50 provides a plurality of components identical to the component 1 without degradation, and hereinafter referred to as training components. Further, as shown in Figure 4 , the training provides for determining (block 100, Figure 4 ) a plurality of parts 2 of the component 1 (visible in Figure 1 ), which are hereinafter referred to as sub-zones 2.

[0027] In fact, component 1 is divided into sub-regions 2, each sub-region having its own outer surface S2, which is hereinafter referred to as the sub-region surface S2. Without loss of generality, the sub-regions 2 are adjacent to each other and do not overlap, such that with reference to the outer surface S1 which indicates the entire outer surface of component 1, each point of the outer surface S1 belongs to the corresponding sub-region surface S2.

[0028] More specifically, for example, component 1 can be divided into respective sub-regions 2 based on the internal structure of component 1, such that, for any sub-region 2 considered and with reference to its cross-section to indicate the cross-section of sub-region 2 taken along a plane perpendicular to the same reference direction, such cross-sections are identical to each other. Still only as an example, in the case where component 1 is formed by different lattice structures (not shown) covered by a metal surface, one or more sub-regions 2 can be defined such that each sub-region covers the corresponding lattice structure. Still as an example, one or more sub-regions 2 can be defined according to the shape of component 1, for example such that the edges of sub-regions 2 coincide with the regions where the thickness change of component 1 or the curvature change of the outer surface S1 of component 1 occurs. However, generally, the criteria used to determine the boundaries of sub-regions 2 of component 1 can be different from those described and are independent of the implementation of this method. In addition, since the training component is the same as component 1, the division into sub-regions 2 also applies to each training component.

[0029] The sub-regions 2 are stored in computer 16 and, for example, have a number equal to NUM_SUBREG.

[0030] Subsequently, the outer surface S1 is divided into a set of respective sub-parts 6, which are hereinafter referred to as basic partitions 6. In particular, the mesh 8 formed by the basic partitions 6 is determined (a part of the mesh is qualitatively shown in Figure 5 )(block 102, Figure 4 ); only as an example, each basic partition 6 can have an extension equal to approximately 1 cm 2 .

[0031] More specifically, even if in Figure 5 the mesh 8 of the basic partitions 6 only extends over a part of the outer surface S1 of component 1, the mesh 8 of the basic partitions 6 completely covers the outer surface S1. In addition, even if in Figure 5 the basic partitions 6 are shown as having an approximately square shape and being arranged to form a matrix, they can have different shapes in addition to being different from each other; the arrangement of the basic partitions 6 can also be different. In a first approximation and for the purposes of this method, the basic partitions 6 are equivalent to point-like partitions and can be struck individually.

[0032] Since the training component is the same as component 1, the mesh 8 of the basic partitions 6 also applies to the outer surface of each training component.

[0033] As described above, for each of the above-described training components, for each in sub-region 2, a plurality of spectrograms corresponding to (box 104, Figure 4 ) are generated, as described below. Figure 6 An example is shown in

[0034] Specifically, for each training component, and for each sub-region 2 of the training component, for each basic partition 6 belonging to sub-region 2, the operation shown in Figure 7 is performed.

[0035] In particular, the basic partition 6 is struck by the striking device 12 (box 200, Figure 7 ) so as to generate a corresponding acoustic signal, which is acquired by the microphone 14 (box 202, Figure 7 ).

[0036] More specifically, the acoustic signal extends over a corresponding time interval having a duration T (e.g., equal to two seconds), which is the same for all acoustic signals; furthermore, the basic partition 6 is struck periodically, for example, at a frequency equal to 3 Hz. Optionally, the acoustic signals can be acquired in a synchronized manner by striking the corresponding basic partition 6 such that during each time interval of the acoustic signal, the same number of strikes of the basic partition 6 occur; optionally, the time arrangement of the strikes can be the same for all acoustic signals. In other words, the acoustic signals are acquired during the periodic strikes of the basic partition 6.

[0037] For example, the acquisition of each acoustic signal provides for sampling the acoustic signal at a sampling frequency, for example, equal to 44 kHz (e.g., with a precision of 16 bits per sample). Thus, in a well-known manner, the acquisition of each acoustic signal requires the conversion of the acoustic signal into a corresponding electrical signal and the sampling of the electrical signal, and thus it requires the generation of a sampled electrical signal, the samples of which represent the corresponding samples of the acoustic signal.

[0038] Subsequently, in a well-known manner, the computer 16 calculates (box 204, Figure 7 ) the corresponding spectrogram for each acquired acoustic signal based on the corresponding sampled electrical signal.

[0039] As Figure 8 qualitatively shown in, each spectrogram is formed by a matrix of values; each row of the spectrogram refers to a corresponding spectral interval (two spectral intervals respectively indicated by Δf1 and Δf n ), while each column refers to a corresponding time sub-interval of the time interval over which the acoustic signal extends. The time sub-intervals of the time interval over which the acoustic signal extends can have the same duration Δt, for example, equal to 40 ms; the spectral intervals can be non-uniform, and thus they can be generated based on, for example, a logarithmic curve rather than a linear curve.

[0040] Given a column of a spectrogram, each value of the column indicates the energy content of the portion of the acoustic signal associated with the corresponding time sub-interval falling within the corresponding spectral interval. For example, the values of each column of the spectrogram are equal to the modulus of the samples of the discrete Fourier transform of the samples of the sampled electrical signal falling within the corresponding time sub-interval; in each column of the spectrogram, each element of the column may also be obtained by digital filtering of several (e.g., three) adjacent samples of the above discrete Fourier transform.

[0041] By way of example, the spectrogram may be a so-called MEL spectrogram.

[0042] Still referring to Figure 4 , the computer 16 stores (block 106, Figure 4 ) the spectrogram and the association existing between each spectrogram and the corresponding sub-region 2 (i.e., the sub-region 2 to which the base partition 6 that has been struck during the acquisition of the acoustic signal referred to by the spectrogram belongs). In fact, the computer 16 stores a corresponding label for each spectrogram, which label represents the corresponding category indicating the sub-region 2 involved by the spectrogram.

[0043] Then, the computer 16 trains (block 108, Figure 4 ) the region classifier 50 in a supervised manner based on the spectrogram and the relative label. As will be explained more specifically below, the region classifier 50 is a multi-class classifier; for example, the region classifier 50 is a convolutional neural network, as Figure 3 shown.

[0044] Specifically, the region classifier 50 includes a feature extraction stage 52, which includes a sequence of one or more hidden layers; in particular, by way of example only, Figure 3 shows a first hidden layer and a second hidden layer indicated by 54 and 54' respectively. In addition, both the first hidden layer 54 and the second hidden layer 54' include respective convolutional stages (indicated by 56 and 56' respectively) and subsequent respective pooling stages (indicated by 58 and 58' respectively).

[0045] The convolutional stages 56, 56' are configured to perform convolution, activation, and (optionally) normalization operations in a well-known manner based on respective filters starting from the data present on the respective inputs. In particular, the convolutional stage 56 of the first hidden layer 54 receives a single spectrogram at the input, while the convolutional stage 56' of the second hidden layer 54' receives the output of the pooling stage 58 of the first hidden layer 54 at the input. To this end, the pooling stages 58, 58' are configured to perform a pooling operation on the output of the corresponding convolutional stages 56, 56'.

[0046] The feature extraction stage 52 further includes a flattening layer 60, which is configured to perform a flattening operation on the output of the pooling stage 58' of the second hidden layer 54'.

[0047] The region classifier 50 also includes a fully connected layer 61 shown in a simplified and qualitative manner, which receives the output of the flattening layer 60 and classifies it into a plurality of classes equal to the number NUM_SUBREG of sub-regions 2; each class is thus associated with a corresponding sub-region 2. For the sake of simplicity of display, in Figure 3 it is assumed that the number NUM_SUBREG of sub-regions 2 is equal to 4; the four sub-regions 2 are indicated by sub-region A, sub-region B, sub-region C, and sub-region D.

[0048] Specifically, the training of the region classifier 50 can occur as Figure 9 shown.

[0049] In particular, starting from the spectrograms related to the training components stored in the computer 16, the computer 16 selects (block 300, Figure 9 ) a first subset, a second subset, and a third subset, which can be separated from each other, that is, they can not share any spectrograms. Hereinafter, the first spectrogram subset, the second spectrogram subset, and the third spectrogram subset are respectively referred to as the training set, the validation set, and the test set.

[0050] Subsequently, based on the training set and the relative labels, the computer 16 performs (block 302, Figure 9 ) the training of the region classifier 50. This training occurs in a well-known manner and is of the supervised type, as described above. For example, the training can provide the following sequence of iterative operations:

[0051] i) Based on at least a part of the spectrograms of the training set, the relative labels of the region classifier 50, and the so-called hyperparameters (such as the so-called learning rate or the type of activation function), update the parameter values of the region classifier 50 (understood as the weights and biases of the filters of the convolutional stages 56, 56' and the fully connected layer 61);

[0052] ii) Classify the spectrograms of the validation set based on the updated values of the parameters of the region classifier 50;

[0053] iii) Check the relationship of a predetermined stopping condition of a known type through the classification of the spectrograms of the validation set; and

[0054] iv) In the absence of the stopping condition, change the value of at least one hyperparameter and the iteration of the previous operations i-iii).

[0055] Therefore, when the classification of the validation set meets the stopping condition, the iteration of the above operation sequence ends. For example, when the error function indicating the difference between the classification of the spectrograms of the validation set and the actual classes is lower than a pre-established threshold, the stopping condition can occur.

[0056] Then, the computer 16 applies the region classifier 50 obtained after the operations mentioned in block 302 (and thus, where the values of the corresponding parameters are available at the end of the operations mentioned in block 302) to the spectrograms of the test set in order to classify them (block 304, Figure 9 ).

[0057] In addition, the computer 16 calculates (block 306, Figure 9 ) the confusion matrix of the classifications obtained through the operations mentioned in block 304.

[0058] The confusion matrix has dimensions NUM_SUBREG x NUM_SUBREG. By way of example only, Figure 10 shows an example of a confusion matrix related to the case where NUM_SUBREG = 4, where the classes are indicated by 1, 2, 3, 4 respectively; the elements are indexed as CM ij , where 'i' indicates the row and 'j' indicates the column; in addition, the rows of the confusion matrix represent the so-called ground truth, i.e., the actual class of the spectrogram, understood as the sub-region 2 that the spectrogram actually refers to, while the columns represent the classifications obtained through the region classifier 50. In other words, the element CM ij represents the number of spectrograms classified as related to the i-th sub-region 2 and related to the j-th sub-region 2, and thus it indicates the probability that the region classifier 50 confuses the i-th sub-region 2 with the j-th sub-region 2.

[0059] Based on the confusion matrix, the computer 16 detects (block 308, Figure 9 ) the possible presence of one or more N-tuples (with N an integer greater than or equal to 2) of classes such that, for each N-tuple, the number of spectrograms associated with the classes of the N-tuple that have been classified in a confused manner with respect to each other (i.e., the elements CM ij with values CM ij ) comply with an aggregation criterion.

[0060] For example, the computer 16 can analyze the confusion matrix row by row, initially assuming that the classes do not form any N-tuples. Having said that, considering the i-th class (where 'i' successively assumes the values 1, 2, 3, and 4), the computer 16 detects whether the i-th class already belongs to an N-tuple, in which case it increments the value of 'i' in order to analyze the next row and thus the next class, otherwise, before incrementing the value of 'i', the computer 16 checks whether there is one or more m-th classes (where'm' is different from 'i') such that CM im > TH (where TH indicates a threshold), in which case the computer 16 alternatively:

[0061] - If none of these m-th classes belong to a previously detected N-tuple, then associate the i-th class with these m-th classes such that the i-th class and these m-th classes together form a new N-tuple class; or

[0062] - If one or more of such m-th classes belong to a previously detected N-tuple, then associate the i-th class with one of such previously detected N-tuples and increase the dimension of such N-tuple by one; in particular, in the case where the number of such previously detected N-tuples is greater than one, the computer 16 may choose which of such detected N-tuples to associate with the i-th class (for example, it may choose the N-tuple with more classes in order to maximize the dimension of the N-tuple, or choose the N-tuple that includes the m-th class such that CM im takes the maximum value).

[0063] In any case, the criteria for determining the N-tuples and the dimension of the classes forming them may vary with respect to what has been described. For example, the confusion matrix may be analyzed by the computer 16 in a way different from the described manner. Additionally, it is possible to determine variants of the N-tuples of classes under the assumption that the confusion matrix is approximately symmetric in the first place, in which case the computer 16 may analyze only a subset of the confusion matrix. Additionally, it is possible to have variants where the number N is predefined; for example, if N = 2, then considering the u-th class and the p-th class, if the element CM up and / or the element CM pu exceeds a threshold, then the computer 16 may detect a pair of classes.

[0064] Subsequently, for each N-tuple of classes detected during the operation mentioned in block 308, the computer 16 aggregates (block 310, Figure 9 ) the classes of the N-tuple; in other words, the computer 16 aggregates the sub-regions 2 (the number equal to N) associated with the classes of the N-tuple in order to form a single region (understood as the aggregation of sub-regions), and this single region is associated by the computer 16 with the corresponding label (i.e., the corresponding class) (block 312, Figure 9 ). For example, Figure 3 qualitatively shows the aggregation of the classes associated with sub-region A and sub-region B.

[0065] Then, based on the sub-regions 2 and the possible aggregations performed during the operation mentioned in block 310, the computer identifies (block 314, Figure 9 ) the multiple regions 9 of the component 1 stored by the computer 16 (as Figure 4 shown).

[0066] In particular, each sub-region 2 that has not been aggregated during the operations mentioned in block 310 forms a corresponding region 9 associated with the label of the sub-region 2; in addition, each set of sub-regions 2 that are aggregated with each other forms a corresponding region 9 associated with the aggregated label of the sub-regions 2.

[0067] By way of example only, Figure 5 the first region (indicated by 9') and the second region (indicated by 9") are highlighted, the first region corresponding to the corresponding sub-region 2 that has not undergone any aggregation, and the second region corresponding to the aggregation of a corresponding pair of sub-regions 2. Still by way of example, Figure 3 it is shown how the aggregation of the classes associated with sub-region A and sub-region B results in the definition of the class associated with sub-region A, while the classes of sub-region C and sub-region D respectively correspond to the classes of region C and region D.

[0068] The aggregation of the classes mentioned in block 310 enables the fully connected layer 61 of the region classifier 50 to classify a set of classes equal to the number of regions 9, which number is hereinafter referred to as the number NUM_Z. In fact, the region classifier 50 is initially configured to classify a plurality of classes (also referred to as sub-region classes) equal to the number NUM_SUBREG of sub-regions 2; after the aggregation of the classes mentioned in block 310, the region classifier 50 is configured to classify a plurality of classes (also referred to as region classes) equal to the number NUM_Z.

[0069] For practical purposes, the region 9 of the component 1 is a zone of the component 1 that, when mechanically struck in the respective basic partitions 6, generates an acoustic signal the spectrogram of which can be classified by the region classifier 50 as being related to the acoustic signal generated by striking a part of that zone.

[0070] Thus, the training of the region classifier 50 ends.

[0071] Referring again to Figure 4 , once the operations mentioned in block 108 are ended, the computer 16 trains (block 110) the above-mentioned binary classifiers 70, the number of which is equal to NUM_Z; thus, each binary classifier 70 is associated with a corresponding region 9 of the component 1.

[0072] As will be explained more specifically below, each binary classifier 70 is trained to classify the spectrogram generated by striking the corresponding region 9 such that the classification alternatively indicates whether the region 9 is damaged.

[0073] Specifically, considering the generic binary classifier 70 associated with the k-th region 9 of the component 1, the computer 16 performs Figure 11 the operations shown.

[0074] The computer 16 selects (block 400, Figure 11)The spectrogram of the above-mentioned training set related to the k-th region 9, which forms a set of first training observations, and also selects the spectrogram of the training set related to a region 9 different from the k-th region 9, which forms a set of second training observations. As will be specifically explained below, the acoustic signals from regions 9 different from the k-th region 9 are considered to be generated by a damaged version of the k-th region 9 in order to eliminate the difficulty of finding the true damaged version of the k-th region 9. Variants are possible in any case, where the set of second training observations also includes or exclusively includes spectrograms obtained starting from acoustic signals generated by one or more damaged versions of the k-th region 9 (i.e., by components in which the k-th region is damaged), and / or spectrograms obtained starting from acoustic signals generated by one or more damaged versions of the w-th region 9 (where w is different from k) (i.e., by components in which the w-th region is damaged).

[0075] In addition, the computer 15 selects (block 402, Figure 11 )the spectrogram of the above-mentioned validation set related to the k-th region 9, which forms a set of first validation observations, and also selects the spectrogram of the validation set related to a region 9 different from the k-th region 9, which forms a set of second validation observations. Variants are possible in any case, where the set of second validation observations also includes or exclusively includes spectrograms obtained starting from acoustic signals generated by one or more damaged versions of the k-th region 9 and / or by one or more damaged versions of the w-th region 9 (where w is different from k) (i.e., by components in which the w-th region is damaged).

[0076] Then, the computer 16 initializes (block 404, Figure 11 )the binary classifier 70, as Figure 3 shown, which is formed by a convolutional neural network and includes a corresponding feature extraction stage 72 and at least one fully connected layer 81, which includes two output nodes respectively associated with the "undamaged region" class and the "damaged region" class.

[0077] In particular, the binary classifier 70 is initialized such that the corresponding feature extraction stage 72 is the same as the feature extraction stage 52 of the region classifier 50. In other words, the feature extraction stage 72 of the binary classifier 70 has the same structure as the feature extraction stage 52 of the region classifier 50; in addition, the initial values of the parameters (i.e., the weights and biases of the filters) of the feature extraction stage 72 of the binary classifier 70 are equal to the values of the corresponding parameters of the feature extraction stage 52 of the region classifier 50.

[0078] The fully connected layer 81 of the binary classifier 70 can be initialized in a well-known manner, regardless of the fully connected layer 61 of the region classifier 50.

[0079] In fact, the binary classifiers 70 are initialized in the same way, regardless of the regions 9 of the component 1 to which they refer.

[0080] Referring again to Figure 11 , after initializing the binary classifiers 70, the computer 16 trains (block 406, Figure 11 ) the binary classifiers 70 to associate the first training and validation observations with a first class indicating the fact that the spectrogram is related to an unimpaired version of the region to which the binary classifier 70 pertains, and to associate the second training and validation observations with a second class indicating the fact that the spectrogram is related to an impaired version of the region to which the binary classifier 70 pertains.

[0081] The training enables the feature extraction stage 72 of the binary classifier 70 to separate the value of the corresponding parameter from the parameter value region of the feature extraction stage 52 of the region classifier 50.

[0082] In fact, the applicant has observed that by initializing the binary classifiers 70 as described above, the relative performance can be improved in terms of the actual ability of the regions to distinguish between unimpaired and impaired regions. Regardless of the region classifier 50, it is possible in any case to initialize variants of the binary classifier 70 in a well-known manner.

[0083] Once the region classifier 50 and the binary classifiers 70 have been trained, the detection system 10 can be used to detect possible anomalies in unknown components of the same type of component 1, but whose impairment status is not known a priori. To this end, the operations shown in Figure 12 are performed.

[0084] Specifically, the striking device 12 is actuated to strike (block 500, Figure 12 ) the base partition 6 of the unknown component, which is hereinafter referred to as the unknown partition, in order to generate a corresponding acoustic signal that is acquired (block 502, Figure 12 ) by the computer 16 via the microphone 14; based on the acquired acoustic signal, the computer 16 calculates (block 504, Figure 12 ) a corresponding spectrogram, which is hereinafter referred to as the unknown spectrogram, since the status (unimpaired / impaired) of the unknown partition is not known a priori; furthermore, hereinafter, the unknown region is referred to in order to denote the region 9 of the unknown component to which the unknown partition belongs.

[0085] Subsequently, the computer 16 applies (block 506, Figure 12 ) the region classifier 50 to the unknown spectrogram in order to classify the unknown region. In fact, the region classifier 50 allows the region 9 corresponding to the unknown region to be identified among the regions 9 of the component 1.

[0086] Then, the computer 16 selects (block 507, Figure 12)The binary classifier 70 related to the identified region.

[0087] Then, the computer 16 applies the selected binary classifier 70 to the unknown spectrogram (block 508, Figure 12 ), and the selected binary classifier 70 alternatively classifies the unknown spectrogram as i) belonging to the corresponding first class, which indicates that the unknown spectrogram is related to an acoustic signal generated by striking an undamaged version of the identified region 9, or classifies the unknown spectrogram as belonging to the corresponding second class, which indicates that the unknown spectrogram is related to an acoustic signal generated by striking a damaged version of the identified region 9.

[0088] In the case where the spectrogram has been classified as belonging to the second class, the computer 16 detects (block 510, Figure 12 ) the presence of an anomaly (i.e., damage / deterioration) in the unknown region of the unknown component; in this case, in a well-known manner, the computer 16 can generate corresponding signaling.

[0089] By iterating the Figure 12 operations shown in different base partitions 6 and different regions 9 of the unknown component, anomalies that may exist in each region 9 of the unknown component can be detected.

[0090] In particular, in the case where all spectrograms related to the region 9 of the unknown component have been classified as belonging to the corresponding first class, the region 9 is undamaged; alternatively, if one or more of the spectrograms related to the region 9 of the unknown component have been classified as belonging to the corresponding second class, the region is damaged.

[0091] In the case of a given region 9, the accuracy of detection may be reduced if only spectrograms of a subset of the base partition 6 of the region 9 are classified.

[0092] The present solution allows for the advantages that are clearly obtained from the previous description.

[0093] In particular, the present method allows for the automated detection of anomalies in aircraft mechanical components and the localization of possible anomalies at the level of individual regions of the mechanical components. Nevertheless, the present method allows for the exclusion of the presence of a trained operator.

[0094] Finally, it is clear that the methods and systems for detecting anomalies as described and shown herein can be modified and varied without departing from the scope of protection of the present invention as defined in the appended claims.

[0095] For example, the region classifier and / or the binary classifier can be formed by different types of classifiers as described.

[0096] Furthermore, although in the foregoing description, for simplicity, it is assumed that the grid 8 of the base partitions 6 and therefore the definition of the shape and arrangement of the partitions struck by the striking device 12 are the same for the component 1, the training component and the unknown component, the grid of the base partitions of one or more of the training components and the unknown component may be different from the grid 8 of the base partition 6 of the component 1. In other words, for the purposes of the present method, given an area 9 of a component 1, it is not necessary to strike the corresponding area of the unknown component and / or the corresponding area of one or more of the training components at the same point, although this may be necessary to improve performance.

[0097] Furthermore, in the case where the detection system 10 is configured such that the operations mentioned in box 500 are performed on an unknown partition, the operations mentioned in box 506 can be omitted, the region to which the unknown partition belongs being known a priori. In this case, the unknown spectrogram is not classified by the region classifier 50, but is classified only by the binary classifier 70 associated with the region to which the unknown partition belongs, which is selected by the computer 16 according to the region to which it belongs.

[0098] With respect to the binary classifiers 70 , as previously described, they may be trained without being pre-initialized based on the region classifiers 50 .

[0099] In addition, the detection and aggregation operations of N-tuple classes mentioned in boxes 308 and 310 are optional. In other words, each region 9 can be consistent with the corresponding sub-region 2, in which case the region classifier 50 is configured to classify a number of classes equal to NUM_SUBREG, and in addition, the number of binary classifiers 70 is equal to NUM_SUBREG. This requires increasing the number of binary classifiers 70, and thus increases the computational burden required to train them.

[0100] Furthermore, before calculating the spectrogram, the computer 16 may perform a so-called denoising operation, ie a noise filtering operation, on the sampled electrical signal derived from the conversion of the acoustic signal, in which case the spectrogram is calculated based on the sampled electrical signal available after filtering of the noise.

[0101] Similarly, the computer 16 can perform a normalization operation of the spectrogram, and the aforementioned operation can be performed starting from the normalized spectrogram. To this end, the computer 16 can calculate the mean and standard deviation of the elements of the spectrogram associated with the training component, and can then subtract the mean from each of such spectrograms except the unknown spectrogram; in addition, the computer 16 can divide the elements of the unknown spectrogram and the spectrogram associated with the training component for the standard deviation. Other types of normalization or normalization are possible anyway.

[0102] Finally, the present method and system for detecting anomalies can also be applied to mechanical components different from aircraft mechanical components; for example, they can be applied to structurally monitor wind turbine blades or civil infrastructure and, more generally, to monitor the health status of any mechanical component.

Claims

1. A method implemented by a computer (16) for detecting anomalies in an unknown component, the method comprising determining (314) a plurality of regions (9) of the unknown component and performing the following steps at least once: - generating (500, 502, 504) spectrograms related to acoustic signals generated by striking a part (6) of a region (9) of the unknown component; - selecting (507) a binary classifier (70) related to the struck region among a plurality of binary classifiers (70) respectively related to corresponding regions (9) of the unknown component, each binary classifier (70) among the plurality of binary classifiers (70) being configured to classify spectrograms related to acoustic signals generated by striking the corresponding region on two classes respectively indicating spectrograms related to acoustic signals generated by striking an undamaged version or a damaged version of the corresponding region (9), - performing (508) classification of the spectrogram on each of the two classes by the selected binary classifier; and - detecting (510) the presence of an anomaly in the struck region of the unknown component based on the classification performed by the selected binary classifier.

2. The method according to claim 1, wherein Each binary classifier (70) has been trained in a supervised manner based on: - respective first training spectrograms related to acoustic signals generated by striking parts (6) of the corresponding regions (9) of a training component identical to the unknown component without any damage, the first training spectrograms being associated with corresponding first classes; and - respective second training spectrograms related to acoustic signals generated by striking parts (6) of corresponding regions (9) of a training component identical to the unknown component with damage in the corresponding regions, and / or related to acoustic signals generated by striking parts (6) of regions (9) different from the corresponding regions (9) of a training component identical to the unknown component without any damage, and / or related to acoustic signals generated by striking parts (6) of regions (9) different from the corresponding regions (9) of a training component identical to the unknown component with damage in regions (9) different from the corresponding regions (9), the second training spectrograms being associated with corresponding second classes.

3. The method according to any one of the preceding claims, wherein selecting (507) the binary classifier (70) related to the struck region comprises: - classifying (506) the spectrogram into a corresponding class among a plurality of region classes equal in number to the number of regions (9) of the unknown component by a multi-class classifier (50), each region class among the plurality of region classes indicating a spectrogram related to an acoustic signal generated by striking the corresponding region; and - selecting (507) the binary classifier (70) based on the classification performed by the multi-class classifier (50).

4. The method according to claim 3, wherein, The multi-class classifier (50) is trained by performing the following steps: - determining (100) a plurality of sub-regions (2) of the unknown component; and subsequently - Training (108) the multi-class classifier (50) based on a set of training spectrograms related to acoustic signals generated by striking each part (6) of a sub-region (2) of a training component identical to the unknown component without any damage, each training spectrogram in the set being associated with a corresponding sub-region class (2) indicating the sub-region (2) to which the training spectrogram pertains, such that the multi-class classifier (50) is configured to perform classification for a plurality of sub-region classes equal to the number of sub-regions (2), each sub-region class in the plurality of sub-region classes indicating a spectrogram related to an acoustic signal generated by striking the corresponding sub-region (2); and subsequently - Defining (308, 310, 312, 314) the region classes such that each region class is the same as or indicates a set of the corresponding sub-region classes, and subsequently configuring the region classifier (50) such that the region classifier (50) performs classification for the plurality of region classes.

5. The method according to claim 4, wherein the step of defining (308, 310, 312, 314) the region classes comprises: - Classifying (304) a plurality of test spectrograms related to acoustic signals generated by striking parts (6) of the sub-region (2) of a training component identical to the unknown component without any damage by the region classifier (50), such that each test spectrogram is classified into a corresponding sub-region class; - Calculating (306) a confusion matrix of the classification of the test spectrograms; And - Based on the confusion matrix, detecting (308) the existence of a set of two or more sub-regions (2) such that the test spectrograms related to the two or more sub-regions (2) have been classified in a confused manner with respect to each other in a manner that meets a threshold condition; And - For each detected set of sub-regions (2), aggregating (310) the corresponding sub-region classes to form a corresponding region class; And - For each sub-region (2) that does not belong to any detected set of sub-regions (2), setting (314) the corresponding region class to be equal to the sub-region class.

6. The method according to claim 4 or 5, wherein In each training, each binary classifier (70) has been initialized based on the region classifier (50).

7. The method according to claim 6, wherein the region classifier (50) and the binary classifier (70) are convolutional neural networks, each of the convolutional neural networks including respective feature extraction stages (52, 72); and wherein the binary classifier (70) has been initialized such that the respective feature extraction stage (72) is the same as the feature extraction stage (52) of the region classifier (50).

8. A method for detecting anomalies, comprising the steps of: - Performing the method implemented by a computer (16) according to any one of the preceding claims; - Performing (500) the striking of a part (6) of a region (9) of an unknown component; And wherein generating (502, 504) a spectrogram from an acoustic signal comprises: - Obtaining (502) the acoustic signal; and - Calculating (504) a spectrogram based on the obtained acoustic signal.

9. The method for detecting anomalies according to claim 8, wherein the striking is performed periodically.

10. A processing system comprising means (16) configured to perform the method according to any one of claims 1 to 7.

11. A system comprising: - the processing system (16) according to claim 10; - a striking device (12) configured to mechanically strike a single part (6) of an area (9) of an unknown component so as to generate a corresponding acoustic signal; and - a microphone (14) coupled to the processing system (16) and configured to acquire the acoustic signal.

12. A computer program comprising instructions which, when executed by a computer (16), cause the computer (16) to perform the method according to any one of claims 1 to 7.

13. A computer-readable computer medium having stored thereon the computer program according to claim 12.

Images