Method of non-destructive evaluation using ultrasonic reverberation

Abstract

The present invention relates to a method of non-destructive evaluation using ultrasound reverberated signals to determine the structural integrity of a part and comprising: simulating an acoustic propagation of an emitted ultrasound pulse by iteratively varying an angle of incidence, a signal reception gain and an acquisition time; determining those which allow a reverberated signal to be generated and detected inside the part; obtaining a signal reverberated inside the part; applying to the obtained reverberated signal a deep neural network trained by means of a set of ultrasound signals obtained from flawless reference parts and configured to determine a synthetic reverberated signal that better matches the reverberated signal obtained; and determining a reconstruction error value, a distance metric between the obtained reverberated signal and the synthetic reverberated signal; and determining the presence of flaws in the part based on said reconstruction error value.

Description

[0001] NON-DESTRUCTIVE EVALUATION METHOD USING ULTRASONIC REVERBERATION

[0002] OBJECT OF THE INVENTION

[0003] The present invention relates to a method that allows automatic detection of defects in parts using ultrasonic reverberation.

[0004] An object of the invention is to provide a method that allows obtaining a non-destructive evaluation by using reverberated ultrasound signals to determine the structural integrity of a part.

[0005] Another object of the invention is to provide a non-destructive evaluation system that allows determining the presence of defects in a part.

[0006] BACKGROUND OF THE INVENTION

[0007] Nondestructive evaluation (NDT) techniques were initially used to inspect in-service parts and detect wear problems or defects caused by use. Over time, thanks to technological advances and the need for increasingly stringent quality controls, the industry also began to use them in production plants to inspect newly manufactured parts. Currently, in industries such as aeronautics, automotive, railways, and metallurgy, the production of sample parts is unthinkable without the aid of NDT to control not only the parts produced but also the production process itself.

[0008] One of the most widely used techniques is ultrasonic inspection, as it allows defects to be identified both on the surface and inside parts. To perform these tests, ultrasonic pulses travel through the material. When they encounter an interface or defect, some of the energy is reflected back to the transducer. The information captured by the transducer is analyzed to determine the location, size, and nature of the defect. Thus, ultrasound is especially useful for detecting cracks, inclusions, internal corrosion, and variations in the material's structure. In a basic configuration, ultrasonic NDT equipment consists of one or more probes, a system responsible for emitting the pulses and receiving the signals, and control and evaluation software.To assess the condition of the inspected part, the most common approach is to use a set of different signals to form an ultrasonic image of it. There are different representations—A-scan, B-scan, C-scan, S-scan, etc.—and different ways to obtain them. These range from the most conventional techniques of performing manual or automated scanning of the part to current techniques obtained using phased-array systems.

[0009] There are techniques that determine the condition of a structure by comparing signals taken from it with those of a defect-free reference part. In some cases, the emission parameters are adjusted to focus on specific points along different paths, or guided waves are used to try to detect small echoes produced at specific positions where a defect is expected.

[0010] In any case, the images obtained are subsequently evaluated by experts to determine whether the part is correct or not, so it is essential that the image be of good quality and easily interpretable. However, the same physical phenomena that allow ultrasound to be used as an evaluation method can make image interpretation difficult, reverberation being one of them.

[0011] Ultrasonic reverberation is an effect that occurs when an ultrasonic wave propagates inside a closed environment. These waves are reflected and / or refracted at the internal interfaces and boundary surfaces of the insulated object. This effect is repeated multiple times in different directions and amplitudes until all the energy is dissipated and the ultrasonic wave becomes undetectable. In general, reverberation is an undesirable effect in nondestructive ultrasound evaluation, as it generates a series of artifacts or ghost echoes whose interpretation is complex and unclear, ultimately impairing the obtained ultrasound image.

[0012] There are multiple ultrasonic inspection systems and methods to eliminate and prevent this effect. The most widely used method involves using detection doors or windows to form ultrasonic images with time sections of the signals where this effect does not occur or is sufficiently attenuated to be considered negligible, as it does not exceed a certain threshold. There are also studies that use reverberation to determine the resonance of the inspected materials and calculate the elastic parameters of the inspected structure.

[0013] DESCRIPTION OF THE INVENTION

[0014] The present invention is based on the use of ultrasonic reverberation to determine whether an inspected part is functionally correct or not. Thus, this invention relates to a non-destructive evaluation system that allows the complete sonication of the part to be inspected, a method that allows determining the ultrasonic inspection parameters to acquire the reverberated signal generated in the part, and applying an artificial neural network to automatically determine whether the reverberated signal originates from a functionally correct part or not.

[0015] The main advantage of this technique over conventional methods is that it allows for a shorter evaluation time for ultrasonic inspections. In general, to evaluate a part or a region thereof, multiple ultrasonic pulses must be generated, time-limited to avoid the reverberated signal, acquired, represented in an image, and evaluated using an expert or any automatic method. The method of the present invention allows the validity of a part, or a region thereof, to be determined by emitting a single pulse, acquiring the reverberated signal, and classifying this signal using a neural network. This technique allows for the rapid and automatic detection of invalid parts, based on prior training, without the need to generate or evaluate images.

[0016] The non-destructive evaluation method of the invention makes use of reverberated ultrasound signals to determine the structural integrity of a part and comprises the steps of: simulating an acoustic propagation of an emitted ultrasonic pulse by iteratively varying inspection parameters, which comprise: an angle of incidence, a signal reception gain and an acquisition time, to determine at least one set of inspection parameters that allow generating and detecting a reverberated signal within the part; obtaining a reverberated signal within the part to be evaluated, by means of a multi-element transducer (phased-array) or a single-element transducer;applying to the obtained reverberated signal a deep neural network, which may be of the autoencoder type, trained using a set of ultrasonic signals obtained from reference pieces without defects, or with one or more previously known defects, and configured to: o determine a synthetic reverberated signal belonging to the learned acoustic pattern that best corresponds to the obtained reverberated signal, and / or determine a reconstruction error value, being a distance metric between the obtained reverberated signal and the synthetic reverberated signal; and determine the presence of defects in the piece based on whether the determined reconstruction value is greater or less than one or more previously defined reconstruction error thresholds.;

[0017] Preferably, reconstruction error thresholds are defined based on a system tolerance.

[0018] In deep neural network training, signals are preferably used in which the position and coupling of the transducer are varied.

[0019] Furthermore, the deep neural network can be trained to achieve minimal reconstruction error by adjusting its hyperparameters. Preferably, techniques will be applied to avoid overfitting and / or underfitting.

[0020] Deep neural network hyperparameters may include: number of nodes in the encoder, latent vector, and decoder, number of epochs, percentage of data used for validation, batch size, maximization, and loss functions.

[0021] More preferably, the deep neural network may be trained to reach a threshold value in at least one of the output metrics selected from: the evolution of the data over time, histograms, individual scalar metrics or data distributions.

[0022] In particular, the deep neural network can be trained to minimize a loss value per epoch. The loss value per epoch indicates how far the synthetic reverberated signal generated in an epoch is from an input signal from that same epoch. In this way, the configuration of each neuron in each layer that minimizes the loss value per epoch is stored as the best learning state.

[0023] The method may also comprise an additional step of changing the weights of the neurons in each epoch before moving to the next epoch.

[0024] The method of the invention may also comprise an initial calibration step comprising obtaining two or more thresholds, in particular: a first threshold (PB|PM): which is the smallest model reconstruction error for the case trained with multiple evaluations of standard pieces; and a second threshold (PM|SP): obtained by evaluating the measurements made with the probe without contact with the part.

[0025] Thus, if in the stage of determining the presence of defects in the part the reconstruction error is lower than the first threshold it is determined that the part is good (PB), if the reconstruction error is between the first and second threshold it is determined that the part is bad (PM) and if it is above the second threshold it is determined that there is no part or signal available (SP)

[0026] Additionally, as many thresholds as necessary can be established to classify different degrees or types of alteration.

[0027] The method of the invention may further comprise a step of validating the reconstruction error thresholds using parts with artificial defects, detected with a conventional inspection, that is, without using reverberation and scanning the entire area to be inspected.

[0028] Regarding the stage of simulating the acoustic propagation of the emitted ultrasonic pulse, this can be carried out using ray tracing. Furthermore, this stage can include the steps of simulating the trajectory of the ultrasonic pulse and simulating an acoustic field generated along these trajectories, which will also allow the areas between rays to be insonified. On the other hand, in the stage of obtaining a reverberated signal, the ultrasonic parameters can be iteratively varied until it is determined that the volume that the reverberated signal has not traversed is smaller than the size of the defect to be detected.

[0029] The present invention also relates to a non-destructive evaluation system, comprising: a transducer configured to generate an ultrasonic pulse; a receiver configured to obtain a reverberated signal within the piece to be evaluated; a processing module connected to the transducer and to the receiver configured to: o simulate an acoustic propagation of an emitted ultrasonic pulse by iteratively varying inspection parameters, comprising: an angle of incidence, a signal reception gain and an acquisition time, to determine at least one set of inspection parameters that allow generating and detecting a reverberated signal within the piece and transmitting it to the transducer; o obtain a reverberated signal within the piece to be evaluated from the receiver;or applying to the obtained reverberated signal a deep neural network trained using a set of ultrasonic signals obtained from reference pieces without defects and configured to:;

[0030] ■ determine a synthetic reverberated signal belonging to the learned acoustic pattern that best corresponds to the obtained reverberated signal, and

[0031] ■ determine a reconstruction error value, being a metric of distance between the obtained reverberated signal and the synthetic reverberated signal; and / or determine the presence of defects in the part based on whether the determined reconstruction error value is greater or less than one or more previously defined reconstruction error thresholds

[0032] Finally, the invention also relates to a computer program configured to run on the processing module of the system of the invention and to carry out the steps of the defined method. The described method allows for the automatic verification of the integrity of the entire volume of the analyzed part with respect to a reference standard without any alteration.

[0033] Thus, the method of the invention provides the following advantages over the state of the art:

[0034] • It uses existing technology, so it does not require investment in new equipment.

[0035] • Reduces inspection time for the part or areas of special interest.

[0036] • It does not depend on the training of an operator to decide whether defects exist or not.

[0037] • No specific operator training is required to detect different types of defects. Currently, each type of defect can produce different image distortions, and operators must be trained to recognize them.

[0038] • Minimizes the risk of not detecting a defect that, due to poor orientation or accessibility, does not directly generate a distortion in the image that the operator typically evaluates.

[0039] • The designed artificial intelligence model only needs to be trained using signals obtained from defect-free part specimens or from the area of ​​interest, but not from every possible type of alteration that may occur. Therefore, sample specimens for each type of defect to be detected are not required.

[0040] DESCRIPTION OF THE DRAWINGS

[0041] To complement the description being made and in order to help better understand the characteristics of the invention, in accordance with a preferred example of practical implementation thereof, a set of drawings is attached as an integral part of said description, in which the following has been represented for illustrative and non-limiting purposes:

[0042] Figure 1 shows a block diagram explaining the steps of an embodiment of the method of the invention. Figures 2A and 2B show a part with a reverberated signal inside and a diagram of its corresponding echo.

[0043] Figures 3A, 3B, 3C and 3D.- They show two examples of a conventional inspection, and its corresponding diagram, in which multiple acquisitions would have to be made to detect the two defects.

[0044] Figures 4A and 4B.- Show an example of steel metal test pieces, without defects and with artificial defects.

[0045] Figure 5.- Shows an example of the path of the reverberated signal inside the defect-free part shown in Figure 4.

[0046] Figures 6A, 6B and 6C.- Show an example of the path of the reverberated signal in a piece without a defect, in a piece with one defect and in a piece with three defects.

[0047] Figure 7.- Shows a diagram of the loss per epoch during the training of the neural network in a particular embodiment of the invention.

[0048] Figures 8A, 8B, 8O, and 8D - Show evaluation examples using the method of the invention: a case obtained with poor transducer-part coupling; another case obtained from a part with no defects; another case obtained from a part with three defects on the bottom surface; and another case obtained from a part with a non-through lateral defect and a hole in the rear area.

[0049] PREFERRED EMBODIMENT OF THE INVENTION

[0050] The method of the invention consists of different procedures that can be divided into three stages, as shown in the diagram in Figure 1:

[0051] 1. Obtaining the reverberated signal.

[0052] It is necessary to check that a reverberated signal can be generated that runs through the entire area of ​​interest from different angles so that the possible internal and external defects (4) in turn generate reflections that modify the usual reverberation produced by the geometry of the piece (1).

[0053] To this end, given a position of the transducer (2) in the part (1), three main parameters will have to be determined: angle of incidence, signal reception gain, and acquisition time. In general terms, these parameters will be selected based on geometric simulation tools for acoustic propagation that make it possible to verify that the emitted ultrasonic pulse has traveled through the entire area of ​​interest several times, impacting at different angles. In the method of the invention, it is feasible that for the same area of ​​interest, different sets of parameters will have to be used to ensure the sought-after complete insonification objective. In a preferred embodiment, a multi-element system (phased-array) is used, which makes it possible to easily select the three parameters for emitting and receiving the ultrasonic pulse. Alternatively, single-element systems can be used if it is verified that it is not necessary to use many different angles of incidence.

[0054] Figures 2A and 2B show that the reverberation of the ultrasonic pulse will allow the inspection of the entire volume of the area of ​​interest if it is incident at the appropriate angle, the acquisition time is sufficiently long and the system gain allows sufficient amplification of the different echoes produced on the different surfaces of the area of ​​interest and that return to the receiver (3) following multiple trajectories.

[0055] Figures 3A to 3D show the difference with conventional inspection, where sometimes paths including multiple reflections can be used to inspect hard-to-reach areas. In this case, an attempt is made to prevent reverberated echoes from appearing by limiting the gain and acquisition time so that only the echoes of potential defects (4) appear, if any. Consequently, to inspect the same area of ​​the part (1), different sweeps of the transducer (2) must be performed in different positions, since limiting reverberation makes it impossible to inspect the entire part (1) from a single position.

[0056] 2. Training the neural network.

[0057] The method of the invention makes use of a deep neural network of the “autoencoder” type that is trained to associate a pattern of ultrasound signals, generated by multiple reflections when traveling along multiple paths inside a specific volume of parts (1) free of defects (4), with a reference state. The resulting network will be able to, given input data, generate a synthetic signal belonging to a previously learned acoustic pattern, which is the one that best corresponds to the input provided. The distance, with a chosen metric, between the input and output data of this network, called reconstruction error, will be the indicator of the degree to which the inspected part (1) belongs to the set of parts (1) free of defects (4).Thus, the output stage of the neural network will not be a classification, for example, part (1) with defects (4) or without defects (4), but a real value between 0 and 1 that will indicate the deviation with respect to the standard inspections. If the value is close to 0 it will mean that there is a great similarity with the reverberated signal of a part (1) without defects (4) while if the value is close to 1 it will mean that there is no similarity.

[0058] For training, a set of signals obtained from a collection of parts (1) without defects (4) will be used using the parameters determined in the previous stage. The different inspection parameters will be slightly varied, including the position and coupling of the transducer (2) within the margins determined by the type of inspection to be performed. Generally, these margins will be greater when the inspection is performed manually than when it is performed using automatic positioning systems, where the repeatability of the measurement is greater.

[0059] In the learning process, the neural network parameters, also known as hyperparameters, are adjusted until an optimal fit between the training data and the data reserved for validation is obtained, in this case, until a minimum reconstruction error is achieved. The tuning of these hyperparameters (number of nodes in the encoder, in the latent vector, and in the decoder, number of epochs, percentage of data used for validation, batch size, maximization and loss functions, etc.) can be done manually or semi-automatically based on various available metrics or output indicators, such as data evolution over time, histograms, individual scalar metrics, or data distributions.

[0060] It is important to achieve a minimal reconstruction error, but without overfitting or underfitting. In the first case, the model will produce a reconstruction error equal to 0 for the training data set, but will not correctly recognize new data. In the second case, the reconstruction error will be highly variable even with the training data. The latter will generally be due to insufficient data available for training or excessive simplification of the model.

[0061] 3. Defect detection (4)

[0062] Once the network has been trained, a calibration process will be carried out that will determine the threshold values ​​of the reconstruction error that will allow the reverberated signal acquired from the inspected part (1) to be classified and thus determine its integrity.

[0063] In this case, three classes are established:

[0064] Good part (1) (PB): The reverberated signal comes from a part (1) without defects (4) or with defects (4) within the acceptable tolerance for the part

[0065] (1).

[0066] Bad part (1) (PM): the reverberated signal comes from a part (1) with defects (4) that prevent its use and that will need to be discarded or a conventional inspection done to determine the defect (4) more accurately.

[0067] No part (1) (SP): the reverberated signal does not exist or does not have sufficient amplitude. In general, this will be due to a missing or incorrect coupling of the transducer.

[0068] (2) ultrasonic to the piece (1).

[0069] To carry out this classification, two thresholds will need to be determined: PB|PM and PM|SP. The first threshold will be set by performing multiple evaluations of pattern pieces (1) and will correspond to the lowest reconstruction error of the model for the trained case. The second threshold will be set by evaluating the measurements made with the probe without contact with the piece (1). Reconstruction errors lower than the first threshold will determine that the piece (1) is good (PB), errors between the first and second threshold will determine that the piece (1) is bad (PM) and above the second threshold they will be classified as no piece (1) or signal not available (SP).

[0070] From that moment on, the trained and adjusted model for the piece (1) to be evaluated will be available in the area through which the reverberated signal travels. The evaluation result (PB / PM / SP) will be associated with a numerical value that will determine the deviation of the reconstruction with respect to a standard signal, which will provide information not only on the presence of defects (4) but also on the severity of the alteration found. Furthermore, since the values ​​are continuous, as many thresholds as necessary can be established to classify different degrees or types of alteration.

[0071] To determine the validity of these thresholds, parts (1) with artificial defects (4) that can be detected with conventional inspection can be used. This calibration process can be repeated each time the inspection system or operating method is modified.

[0072] Below is a preferred embodiment of the invention, specifying the parameters necessary for its practical execution.

[0073] In this embodiment, use is made of inspection equipment consisting of a phased array of 32 elements with an emission frequency of 5 MHz and from which it is possible to select different ultrasonic parameters such as angle of incidence, gain and duration of the trace that will allow the reverberated signal to be generated within the pieces (1).

[0074] In this example, the evaluation of pieces (1) of steel whose geometric characteristics are: Length 70 mm x Width 15 mm x Height 30 mm and ultrasonic velocity of longitudinal waves: 5700 m / s, as shown in figures 4A and 4B. These pieces (1) will be subjected to different modifications in their structure, artificial defects (4), to demonstrate the validity of the method. As explained, the stages of the method are:

[0075] 1. Obtaining the reverberated signal.

[0076] To obtain the reverberated signal, an ultrasonic transmission simulation program using ray tracing has been used, which provides the most appropriate inspection parameters so that the reverberated signal travels through the entire volume of the piece (1), as shown in Figure 5.

[0077] In order to limit errors in the determination, the ultrasonic parameters of the system are varied until the volume of the areas that the reverberated signal has not crossed is smaller than the size of the defects (4) to be detected. In this case, the simulation program used only draws the trajectory of the ultrasonic rays but not the acoustic field generated along these trajectories, which will generally also allow the areas between the rays to be insonified. This feature could be included in the simulation program.

[0078] In the case presented, the ultrasonic parameters that will allow the piece (1) to be evaluated are: angle of incidence 25°, gain 25 dB and trace duration 180 mm.

[0079] Once these parameters are determined, simulations of parts (1) with defects (4) can be carried out, which allow us to verify that the reverberated signal is altered in the presence of defects (4), as shown in Figures 6A to 6C.

[0080] 2. Training the neural network.

[0081] The AI ​​model is trained using a set of ultrasonic signals, in this case A-Scans, obtained from a "defect-free" reference volume (4) or standard specimen. In this case, 20,000 A-Scans from approximately 8,000 samples were used, and the hyperparameter search was performed automatically after a basic manual tuning.

[0082] To validate the progress of model optimization, the various available metrics or output indicators must be considered, such as data evolution over time, histograms, individual scalar metrics, or data distributions. In this model, the main objective is to minimize the loss value per epoch. That is, at the end of each epoch, or a complete iteration through all the model layers, a value is obtained that indicates how far the reconstructed signal is from the input of that same epoch. When the output value of the epochs remains unchanged for a certain number of times, determined by the value in early_stoppin _patience, the training process is automatically stopped, and the configuration of each neuron in each layer is stored as a representation of the best possible learning state.

[0083] As a result of all the optimizations performed, the configuration of the main hyperparameters that provided the best performance to the model is the following: Total_training_data = 1000

[0084] Number_of_hidden_layers = 10

[0085] Number_of_nodes = 256

[0086] Kernel_initializer= 'zeros'

[0087] Activation= 'gelu'

[0088] Monitoring = 'valjoss'

[0089] Optimizer = 'Adam'

[0090] Early_stopping_patience = 5

[0091] Batch_size = 128

[0092] Learning_rate = 1e-4

[0093] Num_epochs = 20

[0094] The final result of the metrics used for optimizing the indicated hyperparameters appears in Figure 7. The main metric considered in this model is the loss value per epoch. Thus, at the end of each epoch, a value is obtained that indicates how far the reconstructed signal is from the original data at the input of that epoch. Before moving on to the next epoch, the weights (states) of the neurons are changed in each epoch, and the process is repeated. When the output value of the epochs remains unchanged for a certain number of times, determined by the value in "early_stoppin_patience" (in this case 5), the training process is automatically stopped, and the model then saves the configuration of each neuron in each layer as a representation of the best possible learning state.

[0095] In the example shown in Figure 7, that moment corresponds to when the value on the graph no longer falls below about 5 3 approximately, and a horizontal line begins to appear. Therefore, increasing the number of epochs would no longer improve the model, and the learning process ends.

[0096] 3. Defect detection (4):

[0097] The evaluation of the parts (1) is carried out by applying the proposed AI model and setting thresholds that determine the sensitivity of the system. In this case, the threshold to determine that the part (1) is correct has been set to 0.060, a bad part (1) greater than 0.060 and less than 0.500. Finally, if the reconstruction error is greater than 0.500, it is considered that there is no part (1). These thresholds could be modified at any time by the user to adapt the tolerance of the system without having to modify the model.

[0098] Figures 8A to 8D show some real examples of part evaluation (1). They show that in addition to indicating whether it is a good part, a bad part (1) or no part (1), a numerical value is associated that determines the deviation of the reconstruction with respect to the learned pattern signal, which provides information not only on the presence of defects (4) but also on the importance of the alteration found.

Claims

MODIFIED CLAIMS received by the International Bureau on July 30, 2025 (30.07.2025) 1. A non-destructive evaluation method using reverberated ultrasonic signals to determine the structural integrity of a part (1), comprising the steps of: simulating an acoustic propagation of an emitted ultrasonic pulse by iteratively varying inspection parameters, comprising: an angle of incidence, a signal reception gain and an acquisition time, to determine at least one set of inspection parameters that allow generating and detecting a reverberated signal within the part; obtaining a reverberated signal within the part (1) to be evaluated;applying to the obtained reverberated signal a deep neural network trained using a set of ultrasonic signals obtained from reference pieces (1) without defects (4) or with one or more previously known defects and configured to: o determine a synthetic reverberated signal belonging to a learned acoustic pattern that best corresponds to the obtained reverberated signal, and o determine a reconstruction error value, being a distance metric between the obtained reverberated signal and the synthetic reverberated signal; and determine the presence of defects (4) in the piece (1) based on whether the determined reconstruction error value is greater or less than one or more previously defined reconstruction error thresholds.; 2. Method according to claim 1, wherein the step of obtaining a reverberated signal within the piece (1) to be evaluated is carried out with a multi-element phased array transducer (2) or a single-element transducer (2).

3. Method according to claim 1, wherein the deep neural network used is an autoencoder.

4. Method according to claim 1, wherein the deep neural network is trained with signals in which the position and coupling of the transducer (2) are varied.

5. Method according to claim 1, wherein the deep neural network is trained to adjust its hyperparameters until obtaining a minimum reconstruction error.

6. Method according to claim 5, wherein the deep neural network is trained to avoid overfitting and underfitting.

7. Method according to claim 5, wherein the hyperparameters of the deep neural network comprise a number of nodes in the encoder, in the latent vector and in the decoder, number of epochs, percentage of data used as validation, batch size, maximization and loss functions.

8. Method according to claim 5, wherein the deep neural network is trained until reaching a threshold value in at least one of the output metrics selected from: the evolution of the data over time, histograms, individual scalar metrics or data distributions.

9. Method according to claim 7, wherein the deep neural network is trained to minimize a loss value per epoch, a value that indicates how far the synthetic reverberated signal generated in an epoch is from an input signal of that same epoch, storing as the best learning state that configuration of each neuron of each layer that minimizes the loss value per epoch.

10. Method according to claim 9, further comprising a step of changing the weights of the neurons in each epoch before moving to the next epoch.

11. Method according to claim 1, further comprising a step of obtaining two thresholds: a first threshold: is the smallest reconstruction error of the model for the case trained with multiple evaluations of pattern pieces (1), such that that if the reconstruction error is lower than the first threshold it is determined that the part (1) is good (PB); and a second threshold: obtained by evaluating the measurements made with the probe without contact with the part (1), so that, in the stage of determining the presence of defects in the part, if the reconstruction error is between the first and second threshold it is determined that the part (1) is bad (PM) and if it is above the second threshold it is determined that there is no part (1) or signal available (SP).

12. Method according to claim 1, wherein the step of determining the presence of defects (4) in the part (1) comprises establishing as many thresholds as necessary to classify different degrees or types of alteration.

13. Method according to claim 1, further comprising a step of validating the reconstruction error thresholds using parts (1) with artificial defects (4), detected with a conventional inspection.

14. Method according to claim 1, wherein the step of simulating an acoustic propagation of an emitted ultrasonic pulse is performed by ray tracing.

15. Method according to claim 1, wherein in the step of obtaining a reverberated signal, the ultrasonic parameters are varied iteratively until a volume that the reverberated signal has not passed through is less than the size of the defect (4) to be detected.

16. Method according to claim 1, wherein the step of simulating an acoustic propagation of an emitted ultrasonic pulse comprises simulating a trajectory of the ultrasonic pulse and an acoustic field generated along these trajectories, which will also allow the areas between rays to be insonified.

17. Method according to claim 1, wherein in the step of determining the presence of defects (4) in the part (1) the thresholds are determined based on a tolerance of the system.

18. Non-destructive evaluation system, comprising: 22 a transducer (2) configured to generate an ultrasonic pulse; a receiver (3) configured to obtain a reverberated signal inside the piece (1) to be evaluated, a processing module connected to the transducer (2) and to the receiver (3) configured to: o simulate an acoustic propagation of an emitted ultrasonic pulse by iteratively varying inspection parameters, comprising: an angle of incidence, a signal reception gain and an acquisition time, to determine at least one set of inspection parameters that allow generating and detecting a reverberated signal inside the piece (1) and transmitting it to the transducer (2); o obtain a reverberated signal inside the piece (1) to be evaluated from the receiver (3); o apply to the obtained reverberated signal a deep neural network trained by means of a set of ultrasonic signals obtained from reference pieces (1) without defects (4) and configured to: ■ determine a synthetic reverberated signal belonging to a learned acoustic pattern that best corresponds to the obtained reverberated signal, and ■ determine a reconstruction error value, being a metric of distance between the obtained reverberated signal and the synthetic reverberated signal; and / or determine the presence of defects (4) in the part (1) based on whether the determined reconstruction error value is greater or less than one or more previously defined reconstruction error thresholds 19. Computer program configured to be executed in the processing module of the system according to claim 18 and to carry out the steps of the method according to any of claims 1 to 17. [0001]DECLARATION ACCORDING TO ARTICLE 19.1 [0002]According to Article 19.1 of the PCT Treaty, concerning the modification of claims before the International Bureau: [0003]A new set of claims is provided that replaces those originally presented: [0004]-Claim 1 has been modified to address clarity issues regarding the use of undefined elements and to standardize the use of the terms. [0005]- Amended claim 2 limits the use of parentheses to reference signs. [0006]- Claim 3 has been modified to avoid the use of confusing expressions. [0007]- Claim 7 has been modified to address clarity issues, adapting to the examiner's interpretation. [0008]- The dependence of claim 9 has been modified since it referred to elements not described in claim 8 but in claim 7. [0009]- In claim 11, the use of brackets has been limited to reference signs, and has also been modified to solve clarity problems, adapting to the examiner's interpretation.

Images