Method for automatic quality check of an aeronautical component

Abstract

One aspect of the invention relates to a method for automatic quality inspection of an aeronautical component (200), comprising the steps of: - using a trained artificial neural network (301) to detect a defect (103) on an image (201) of the aeronautical component (200); - training an autoencoder (302) on a database (D2) by projecting (z I ) each image (I) of the database (D2) on a small mathematical space (Z), wherein the images (I) follow a predefined probability law; - computing a plurality of metrics (M) for each image (I) of the database (D2); - supervised training of a classifier (303) as a function of the computed metrics (M); - using the autoencoder (302) and the classifier (303) to detect a defect or anomaly (107) on an image (201) of the aeronautical component (200).

Description

Technical Field

[0001] The technical field of this invention is the technical field of aircraft components, especially the technical field of quality inspection of aircraft components, also known as the technical field of quality control of aircraft components.

[0002] This invention relates to a method for automated quality inspection of aerospace components. Background Technology

[0003] To detect faults in aircraft components, it is known to use an artificial neural network trained on a training database that includes images of faulty and non-faulty aircraft components. Training is performed in a supervised manner, meaning that for each image in the training database, the artificial neural network determines whether that image has a fault.

[0004] Because it's impossible to know all the possible faults that could affect an aircraft component, the training database is not exhaustive. Images of aircraft components with faults not present in the training database are submitted to a neural network, which may then claim that the aircraft component is compliant despite having a fault.

[0005] Similarly, anomalous images—that is, images of aircraft parts that have features different from those used during the learning phase of the neural network, such as, but not exhaustively, poor-quality images or images with different lighting than images in the database—are submitted to the neural network, which may claim that the aircraft part is non-compliant even though it is not faulty, or conversely, claim that it is not faulty when it is faulty.

[0006] Therefore, it is necessary to be able to detect all faults in aircraft components, whether or not these faults are known in advance, and not to claim that aircraft components without any faults are non-compliant.

[0007] It also needs to be able to detect images submitted to an artificial neural network that have features that are different from those used during its training, which could lead to the risk of the neural network malfunctioning. Summary of the Invention

[0008] This invention overcomes the aforementioned disadvantages.

[0009] This invention relates to a method for automated quality inspection of aerospace components, comprising the following steps:

[0010] - Train the artificial neural network on the first training database;

[0011] - Obtain at least one image of an aircraft component;

[0012] - Use trained artificial neural networks to detect faults in images of aircraft components;

[0013] - Unsupervised training of the autoencoder on a second training database containing multiple images of fault-free airborne training components includes the following sub-steps:

[0014] - Project each image in the second training database onto a mathematical space with reduced dimensions relative to the dimensions of each image in the second training database, such that all projections of the images in the second database into the mathematical space follow a predefined probability law;

[0015] - Reconstruct each image in the second training database from the images in the second training database by projecting them into mathematical space to obtain the reconstructed image;

[0016] - For each image in the second training database, multiple metrics are calculated from the image in the second training database to the projection into mathematical space or from the corresponding reconstructed image to obtain a metric value for each metric;

[0017] - Supervised training of the classifier is performed based on the obtained metrics;

[0018] - Use trained autoencoders and classifiers to detect faults or anomalies in images of aircraft components;

[0019] - If the artificial neural network detects at least one fault in the image of the aircraft component, or if the artificial neural network does not detect a fault in the image of the aircraft component but the classifier detects at least one fault or anomaly in the image of the aircraft component, then the aircraft component is considered non-compliant;

[0020] Otherwise, the aircraft component is considered compliant.

[0021] This invention provides an artificial neural network used for detecting faults in aircraft components, equipped with a monitoring module comprising an autoencoder and a classifier, enabling the determination of whether the artificial neural network has correctly performed the detection task. To this end, the monitoring module performs the detection itself, not based on image features such as those of a neural network, but on calculations of a metric projected onto a mathematical space by the autoencoder after training on fault-free images, making it more robust to anomalous images and images with unknown faults. In fact, by training on faulty images, the classifier risks learning only the types of faults present and failing to function properly when the image displays an unlearned fault or a completely anomalous image.

[0022] Therefore, if the detection performance of the neural network deteriorates due to unknown faults in the image under test or anomalies in the image under test, the monitoring module will not exhibit this behavior. It can report an anomaly when its detection results differ from those of the neural network.

[0023] Preferably, since the calculation of the metric depends in part on the projection of the image to be tested onto a mathematical space by the autoencoder, an adversarial autoencoder or a variational autoencoder is chosen as the autoencoder, which allows for better control over the distribution of the image projection onto the mathematical space (called the latent space).

[0024] In addition to the features just mentioned in the preceding paragraphs, the method according to the invention may have one or more of the following supplementary features, either individually or in all technically possible combinations.

[0025] According to one aspect of the invention, the images of the aerospace components, a plurality of images in the first training database, and a plurality of images in the second training database are visible images, X-ray images, ultrasound images, or tomographic images.

[0026] According to one aspect of the invention, the method further includes: for each metric, assigning metric values ​​of images in a second training database to a first set of values ​​and a second set of values ​​according to a predefined validity criterion, wherein a classifier is trained to regard images in the second training database with metric values ​​of the first set of values ​​as having no faults or anomalies, and to regard images in the second training database with metric values ​​of the second set of values ​​as images to be rejected.

[0027] Therefore, supervised training of the classifier is possible without using images with faults or anomalies to avoid the risk of affecting the classifier's generalization ability.

[0028] According to one aspect of the invention, the step of detecting faults on images of aircraft components using a trained autoencoder and classifier includes the following sub-steps:

[0029] - An automatic encoder projects images of aerospace components into mathematical space;

[0030] - An image of the aircraft component is reconstructed by an autoencoder from an image of the aircraft component to a projection onto mathematical space;

[0031] - Calculate metrics from images of aircraft components onto mathematical space or from reconstructed images of aircraft components to obtain metric values;

[0032] - Faults on images of aircraft components are detected by a classifier using metrics obtained from images of aircraft components.

[0033] According to one aspect of the invention, for a given image, multiple metrics include the distance between the projection of the image onto a mathematical space and all projections of multiple images onto a mathematical space in a second database, and / or the distance between the image and the reconstructed image, and / or the entropy of the gap between the image and the reconstructed image.

[0034] Therefore, images whose projections onto mathematical space differ significantly from probability laws, or whose reconstructed images differ greatly from those projections, will be considered faulty or abnormal by the classifier.

[0035] According to an alternative embodiment of the prior art, the probability law is a multivariate Gaussian law, and the distance between the projection of the image onto the mathematical space and all projections of the multiple images in the mathematical space of the second database is the Mahalanobis distance.

[0036] According to one aspect of the invention, the method of the invention includes the following steps: generating an alarm if the artificial neural network does not detect a fault in an image of the aircraft component and the classifier detects at least one fault or anomaly in the image of the aircraft component.

[0037] Therefore, if the artificial neural network does not detect any faults but the monitoring module detects a fault or anomaly, an alarm is generated to indicate that the neural network has failed in its detection task or has failed to perform the detection task correctly.

[0038] According to one aspect of the invention, the training of the artificial neural network can be of the supervised training type, and the first training database includes multiple training images of aircraft components that are faulty or not.

[0039] According to one aspect of the invention, the training of the artificial neural network can also be unsupervised training, and the first training database includes multiple training images of fault-free aircraft components.

[0040] The present invention also relates to a computer program comprising program code instructions which, when executed on a computer, are used to perform the steps of the method according to the present invention.

[0041] The present invention and its various applications will be better understood by reading the following description and interpreting the accompanying drawings. Attached Figure Description

[0042] The accompanying drawings are presented for illustrative purposes only and are in no way limiting of the invention.

[0043] - Figure 1 This is a block diagram illustrating the sequence of steps in the method of the present invention.

[0044] - Figure 2 This is a schematic diagram of the first, fourth, fifth, and sixth steps of the method according to the present invention.

[0045] - Figure 3 This is a schematic diagram of the third and seventh steps of the method according to the present invention.

[0046] - Figure 4It provides a decision matrix for executing (multiple) steps of the method of the present invention based on the detection results of the artificial neural network and the detection results of the classifier. Detailed Implementation

[0047] Unless otherwise stated, the same items appearing in different figures have unique reference numerals.

[0048] This invention relates to an automated method for inspecting the quality of aircraft components.

[0049] The method is automatic, meaning it is implemented by a calculator.

[0050] In the context of this invention, "component quality inspection" means determining whether a component is compliant; that is, if no fault is detected in a component, the component is considered compliant, and if at least one fault is detected in a component, the component is considered non-compliant.

[0051] An anomalous image, or an image with anomalies, is an image that has features different from those used during the learning of an artificial neural network, or an image that is faulty.

[0052] [ Figure 1 ] Figure 1 This is a block diagram illustrating the sequence of steps in method 100 of the present invention.

[0053] [ Figure 2 ] Figure 2 This is a schematic diagram of the first step 101, the fourth step 104, the fifth step 105, and the sixth step 106 of the method 100 according to the present invention.

[0054] [ Figure 3 ] Figure 3 This is a schematic diagram of the third step 103 and the seventh step 107 of the method 100 according to the present invention.

[0055] The first step 101 of method 100 includes: training an artificial neural network 301 on a first training database Dl.

[0056] The first training database Dl contains multiple images I of training aircraft components.

[0057] In the remainder of this application, the terms “neural network” and “artificial neural network” will be used interchangeably.

[0058] For example, the artificial neural network 301 can be of the "multilayer perceptron" (MLP) type, meaning it consists of at least two layers, each containing at least one artificial neuron. The connection between two neurons is called a synapse. Each synapse is assigned a synaptic coefficient.

[0059] For example, each neuron in each layer is connected to every neuron in the previous layer and every neuron in the next layer. However, neurons in the same layer are not connected to each other.

[0060] The artificial neural network 301 can also be ResNet (https: / / arxiv.org / abs / 1512.03385v1), UNet (https: / / arxiv.org / abs / 1505.04597), or SegNet (https: / / arxiv.org / abs / 1511.00561) or any other architecture for performing semantic segmentation, detection, or classification tasks.

[0061] The first step 101 of training the artificial neural network 301, also known as the learning step, includes determining the synaptic coefficients of the neural network 301 from images I in the first training database Dl.

[0062] According to the first embodiment, the training is supervised, i.e., each image I (referred to as an input image) in the first database D1 is associated with the same image in which a fault is identified (referred to as a true output image). Therefore, the first training step 101 includes: browsing the first training database D1, and for each input image I provided to the artificial neural network 301, using an optimization algorithm to update the synaptic coefficients to minimize the gap between the output image provided by the artificial neural network 301 and the true output image associated with the input image I.

[0063] To perform this supervised training, an artificial neural network 301 can be trained to assign either a "fault" or "non-fault" category to an image.

[0064] Alternatively, an artificial neural network 301 can be trained to estimate the bounding box around a fault present in the image.

[0065] According to the second embodiment, the training is unsupervised, that is, only the images I of the first database Dl are provided to the neural network 301.

[0066] In the second embodiment, the first training database Dl includes only images I without faults.

[0067] The second step 102 of method 100 includes: acquiring at least one image 201 of the aerospace component 200 to be inspected.

[0068] The image 201 of the aerospace component 200 to be inspected can be a visible image acquired by an industrial camera, such as an X-ray image acquired by X-ray, such as an ultrasonic image acquired by ultrasound, or a tomographic image acquired by a tomographic scanner.

[0069] Image I in the first training database D1 is of the same type as image 201 of the aircraft component 200 acquired in the second step 102. Therefore, for example, if image 201 of the aircraft component 200 is acquired by ultrasound, then image I in the first training database D1 is also acquired by ultrasound.

[0070] The order of the first step 101 and the second step 102 can be reversed, that is, the second step 102 can be performed before the first step 101.

[0071] The third step 103 of method 100 includes: detecting any faults present in the image 201 of the aerospace component 200 acquired in the second step 102 for the artificial neural network 301 trained in the first step 101.

[0072] like Figure 3 As shown, at the end of the third step 103, the neural network 301 can provide labels indicating whether a fault is detected on the image 201 of the aerospace component 200, namely, a first label for the "fault" category and a second label for the "no fault" category. Figure 3 If the neural network 301 detects a fault in the image 201 of the aviation component 200, the neural network 301 returns the label KO, and if it does not detect any fault in the image 201 of the aviation component 200, the neural network 301 returns the label OK.

[0073] The neural network 301 can also provide an image 201 of the aerospace component 200, in which it will outline the detected faults and may assign a fault type to each outlined fault.

[0074] The fourth step 104 of method 100 includes: training an autoencoder 302 on a second training database D2.

[0075] An "autoencoder" refers to an unsupervised learning algorithm based on artificial neural networks that enables the construction of new representations of datasets, typically with dimensions smaller than the dataset's dimensions. To this end, an autoencoder projects the dataset onto a mathematical space with dimensions smaller than the dataset's dimensions.

[0076] Typically, an autoencoder consists of an encoder and a decoder. The encoder constructs a dimensionality-reduced representation from the initial dataset, and the decoder reconstructs the dataset from the dimensionality-reduced representation.

[0077] An autoencoder performs a series of operations based on the architecture of an artificial neural network, specifically the type of artificial neural layers that make up the network. A convolutional layer is an example of a layer type. Based on a reasonable choice of parameters for each layer of the artificial neural network, the encoder aims to transform and spatially reduce the size of the initial dataset provided at the input of the network to obtain a vector (i.e., a set of variables) that retains only the most relevant information from the initial dataset. Therefore, the encoder's purpose is to transform the initial dataset from an "initial" space to a smaller-dimensional "mathematical" or "numerical" space, which allows the input dataset to be described in a simplified vector form.

[0078] The fourth step 104 includes updating the parameters of the autoencoder 302, namely the synaptic coefficients of the neural network constituting the autoencoder 302, to minimize the error between the reconstructed dataset and the initial dataset.

[0079] Therefore, the training of the autoencoder 302 is unsupervised training, and the second training database D2 has multiple images I of the training aero components without faults.

[0080] In the context of this invention, a dataset is an image or image set, 2D or 3D.

[0081] exist Figure 3 In the process, the encoder of the automatic encoder 302 operates through projection P. Z The images I in the second training database D2 are projected onto the mathematical space Z, and each image I in the second training database D2 has a projection z onto the mathematical space Z. I The decoder of the auto encoder 302 operates via projection P. Z The inverse operation P Z -1 And the projection z onto the mathematical space Z of image I I Reconstruct each image I in the second training database D2 to obtain the reconstructed image I'.

[0082] Therefore, the fourth step 104 includes updating the parameters of the autoencoder 302 to minimize the error between each image I and the corresponding reconstructed image I' in the second training database D2.

[0083] The autoencoder 302 is preferably a variational autoencoder or an adversarial autoencoder, which allows for better control of the projection z. I The distribution in the mathematical space Z (called the latent space).

[0084] The autoencoder 302 is trained to project multiple images I from the second database D2 into the mathematical space Z. I The set Z I It follows a predefined probability law.

[0085] Probability laws include, for example, the uniformity law or the multivariate Gaussian law.

[0086] exist Figure 3 In this context, the probability law is the normal law.

[0087] Image I in the second training database D2 is of the same type as image 201 of the aircraft component 200 acquired in the second step 102 and image I in the first training database D1. Therefore, for example, if image 201 of the aircraft component 200 is acquired by ultrasound, then image I in the second training database D2 is also acquired by ultrasound.

[0088] In the first step 101 of training the artificial neural network 301 on the first training database D1, which is a first embodiment of supervised training, the second training database D2 may be a portion of the first training database D1 containing fault-free images I.

[0089] In the second embodiment of training the artificial neural network 301 on the first training database D1, the first training database D1 and the second training database D2 can be a single training database. Faulty images are used to observe metrics and the behavior of the autoencoder during solution development.

[0090] The fifth step 105 of method 100 includes: projecting each image I from the second database D2 into mathematical space Z. I Or, calculate multiple metrics M for the corresponding reconstructed image I'.

[0091] For a given image I, multiple metrics M include, for example, the projection z of the given image I onto the mathematical space Z. I The projection z of multiple images I from the second training database D2 into the mathematical space Z I The set Z I The distance between, and / or the distance between a given image I and its corresponding reconstructed image I', such as standard types L1, L2, L∞ or other distances, and / or the entropy of the gap between a given image I and its corresponding reconstructed image I'.

[0092] If the projection z of image I in the second training database D2 I The set Z I Following the multivariate Gaussian law of mean μ and covariance Cov in mathematical space Z, the projection z of a given image I onto mathematical space Z is... I The projection z of multiple images I from the second training database D2 into the mathematical space Z I The set Z IThe distance between them is, for example, the Mahalanobis Maha distance, which is defined as:

[0093]

[0094] For example, the entropy H of the gap between a given image I and a reconstructed image I' is defined as:

[0095]

[0096] For P i (I-I'), for pixel P i The gap between image I and the corresponding reconstructed image I.

[0097] exist Figure 2 In the process, step 105 is executed by module 3023.

[0098] At the end of step 5, for each image I in the second training database D2, a metric value M is obtained for each metric M.

[0099] The sixth step 106 of method 100 includes: training a classifier 303 in a supervised manner based on the metric M obtained in the fifth step 105.

[0100] For example, for each metric M, the metric value M obtained in step 5 105 is divided into a first group of values ​​G1 and a second group of values ​​G2 according to a predefined validity criterion.

[0101] The classifier 303 is then trained to treat images I in the second training database D2 with the metric M of the first set of values ​​G1 as having no faults or anomalies, and to treat images I in the second training database D2 with the metric M of the second set of values ​​G2 as having faults or anomalies.

[0102] A predefined validity criterion is, for example, placing the N worst-performing metrics M for a given metric M in a second set of values ​​G2 and placing the other M metrics in a first set of values ​​G1. Therefore, for example, in the case of a Mahalanobis distance metric, the validity criterion includes: placing the N most important metrics M in the second set of values ​​G2 and placing the other metrics M in the first set of values ​​G1.

[0103] Steps 104, 105, and 106 can be executed before or simultaneously with step 101.

[0104] The seventh step 107 of method 100 includes detecting any faults or anomalies in the image 201 of the aerospace component 200 acquired in the second step 102 for the autoencoder 302 trained in the fourth step 104 and the classifier 303 trained in the sixth step 106.

[0105] Therefore, the encoder of the auto encoder 302 projects the image 201 of the aerospace component 200 onto the mathematical space Z, and the decoder of the auto encoder 302 projects the image 201 of the aerospace component 200 onto the mathematical space Z. 201 Reconstruct image 201 of aircraft component 200 to obtain reconstructed image 202.

[0106] The image 201 of the aerospace component 200 is projected onto the mathematical space Z by module 3023. 201 Alternatively, multiple metrics M can be calculated from the reconstructed image 202 of the aerospace component 200.

[0107] The classifier 303 then detects any faults or anomalies in the image 201 of the aircraft component 200 based on the metric M calculated for the image 201 of the aircraft component 200.

[0108] like Figure 3 As shown, at the end of the seventh step 107, the classifier 303 can provide labels indicating whether a fault or anomaly is detected in the image 201 of the aerospace component 200. Figure 2 If classifier 303 detects a fault or anomaly in image 201 of aviation component 200, classifier 303 returns label KO; otherwise, if it does not detect any fault or anomaly in image 201 of aviation component 200, classifier 303 returns label OK.

[0109] Step 7, 107, can be performed before or simultaneously with step 3, 103.

[0110] [ Figure 4 ] Figure 4 It provides a decision matrix for the (multiple) steps of method 100 executed based on the detection results of artificial neural network 301 and classifier 303.

[0111] If the artificial neural network 301 detects at least one fault on the image 201 of the aerospace component 200, then the eighth step 108 of the method 100 is executed, which includes: treating the aerospace component 200 as non-compliant.

[0112] If the artificial neural network 301 does not detect any faults in the image 201 of the aviation component 200, and the classifier 303 detects at least one fault or anomaly in the image 201 of the aviation component 200, then step 8 108 is executed.

[0113] In this case, following step 108 is, for example, step 110, which includes generating an alarm. This alarm indicates that the neural network 301 failed to detect a fault.

[0114] For example, an alarm may lead to the automatic scrapping of aircraft component 200, human expertise, or the storage of image 201 of aircraft component 200 in a first training database Dl for retraining neural network 301.

[0115] If the artificial neural network 301 does not detect a fault in the image 201 of the aviation component 200 and the classifier 303 does not detect a fault or anomaly in the image 201 of the aviation component 200, then the ninth step 109 of the method 100 is executed, which includes: treating the aviation component 201 as compliant.

Claims

1. A method (100) for automated quality inspection of aircraft components (200), comprising the following steps: - Train the artificial neural network (301)(101) on the first training database (D1); - Obtain at least one image (201)(102) of the aerospace component (200); - Use a trained artificial neural network (301) to detect faults (103) on images (201) of aircraft components (200); Its characteristic is that it further includes the following steps: - The autoencoder (302) is trained unsupervised (104) on a second training database (D2) containing multiple training images (I) of fault-free aircraft parts, including the following sub-steps: o Project each image (I) from the second training database (D2) onto (z). I Project the images (I) of the second training database (D2) onto a mathematical space (Z) of a smaller dimension than that of each image (I) in the second training database (D2), such that the projection (z) of the images (I) of the second database (D2) onto the mathematical space (Z) is such that the projection (z) of the images (I) in ... training database (D2) onto the mathematical space I The set of (Z) I It follows a predefined probability law; o Projection (z) from the image (I) in the second training database (D2) to the mathematical space (Z) I Reconstruct each image (I) in the second training database (D2) to obtain the reconstructed image (I'); - For each image (I) in the second training database (D2), the projection (z) from the image (I) in the second training database (D2) to the mathematical space (Z) I Alternatively, multiple measures (M) can be calculated from the corresponding reconstructed image (I') to obtain a measure value (M) for each measure (M) (105); - Supervised training (106) of the classifier (303) based on the obtained metric (M); - Use a trained autoencoder (302) and classifier (303) to detect faults or anomalies (107) in images (201) of aerospace components (200); -If the artificial neural network (301) detects at least one fault in the image (201) of the aircraft component (200), or if the artificial neural network (301) does not detect a fault in the image (201) of the aircraft component (200) and the classifier (303) detects at least one fault or anomaly in the image (201) of the aircraft component (200), then the aircraft component (200) is considered non-compliant (108); otherwise, the aircraft component (200) is considered compliant (109).

2. The method (100) according to claim 1, characterized in that, The images (201) of the aerospace component (200), multiple images (I) of the first training database (D1) and multiple images (I) of the second training database (D2) are visible images, X-ray images, ultrasound images or tomographic images.

3. The method (100) according to any one of the preceding claims, characterized in that... It also includes the step of assigning the measure value (M) of the image (I) in the second training database (D2) to the first set of values ​​(G1) and the second set of values ​​(G2) according to a predefined validity criterion for each measure (M). The classifier (303) is trained to regard the image (I) in the second training database (D2) with the measure value (M) of the first set of values ​​(G1) as having no fault or anomaly (OK), and to regard the image (I) in the second training database (D2) with the measure value (M) of the second set of values ​​(G2) as an image to be rejected (KO).

4. The method (100) according to any one of the preceding claims, characterized in that, The step (107) of detecting faults on an image (201) of an aerospace component (200) using a trained autoencoder (302) and classifier (303) includes the following sub-steps: - An image (201) of the aerospace component (200) is projected onto mathematical space (Z) by an autoencoder (302); -The image (201) of the aerospace component (200) is projected (z) onto mathematical space (Z) by the autoencoder (302). 201 Reconstruct an image (201) of the aircraft component (200) to obtain a reconstructed image (202) of the aircraft component (200); - Projection (z) from image (201) of aeronautical component (200) onto mathematical space (Z) 201 Alternatively, the metric (M) can be calculated from the reconstructed image (202) of the aerospace component (200) to obtain the metric value (M); - Faults on the image (201) of the aircraft component (200) are detected by the classifier (303) using the metric (M) obtained from the image (201) of the aircraft component (200).

5. The method (100) according to any one of the preceding claims, wherein the autoencoder is a variational type or an adversarial autoencoder type.

6. The method (100) according to any one of the preceding claims, characterized in that, For a given image (I, 201), multiple metrics (M) include the projection (z) of the image (I, 201) onto a mathematical space (Z). I , z 201 The projection (z) of multiple images (I) from the second database (D2) into mathematical space (Z). I The set of (Z) I The distance between the two images (I, 201) and the reconstructed image (I', 202), and / or the entropy of the gap between the two images (I, 201) and the reconstructed image (I', 202).

7. The method (100) according to claim 6, characterized in that, The probability law is a multivariate Gaussian law, and the projection (z) of the image (I, 201) onto the mathematical space (Z) I , z 201 The projection (z) of multiple images (I) from the second database (D2) onto the mathematical space (Z) I The set of (Z) I The distance between them is the Mahalanobis distance.

8. The method (100) according to any one of the preceding claims, characterized in that... The step (110) is as follows: if the artificial neural network (301) does not detect a fault in the image (201) of the aviation component (200) and the classifier (303) detects at least one fault or anomaly in the image (201) of the aviation component (200), an alarm is generated.

9. The method (100) according to any one of claims 1 to 8, characterized in that, The training of the artificial neural network (301) is supervised training, and the first training database (D1) includes multiple training images (I) of aircraft parts with or without faults.

10. The method (100) according to any one of claims 1 to 8, characterized in that, The training of the artificial neural network (301) is unsupervised, and the first training database (D1) includes multiple training images (I) of aircraft parts without faults.

11. A computer program product comprising program code instructions, wherein when the program is executed on a computer, the program code instructions are used to perform the steps of the method (100) according to any one of claims 1 to 8.

Images