System and method for reconstructing time-domain ultrasonic detection data from compressed data

Abstract

A method for reconstructing ultrasonic testing data includes obtaining compressed data representing a plurality of ultrasonic signals acquired from an ultrasonic test of a test object, applying a trained machine learning model to the compressed data to reconstruct a plurality of amplitudes associated with the plurality of ultrasonic signals, and generating a plurality of reconstructed amplitude signals having the plurality of reconstructed amplitudes associated therewith, and outputting the plurality of reconstructed amplitude signals.

Description

[0001] Cross-references to related applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 591,354, filed October 18, 2023, the contents of which are incorporated herein by reference. Technical Field

[0003] These improvements generally relate to the fields of nondestructive testing and acoustic inspection, and more specifically, to the reconstruction of time-domain ultrasonic inspection data from compressed data. Background Technology

[0004] Non-destructive testing (NDT) can refer to the use of one or more different techniques to inspect areas on or inside an object, such as to determine the presence of defects or flaws, or to otherwise characterize the object under inspection. One approach to NDT may include the use of acoustic inspection techniques, such as one or more electroacoustic transducers, to allow sound to penetrate areas on or inside the object under inspection and to detect and process the scattered or reflected acoustic energy. This scattered or reflected energy can be referred to as an acoustic echo signal. Typically, such acoustic inspection schemes involve the use of acoustic frequencies within the ultrasonic frequency range, for example, including pulses of energy within a specific range, which may include values ​​from, for example, hundreds of kilohertz to tens of megahertz, as an illustrative example.

[0005] To achieve miniaturization and the use of low-power devices, several methods have been proposed, including using lower amplitude resolution for encoding NDT data. For example, one-bit quantization can be used to binarize 12-bit or 16-bit resolution data to one-bit resolution. This binarization has the advantage of reducing data load, thus enabling bandwidth-limited links, high-speed transmission, or more channels. However, a drawback of this method is the loss of amplitude information discarded by the binarized signal.

[0006] Therefore, improvements are necessary. Summary of the Invention

[0007] According to one aspect, a method for reconstructing ultrasonic detection data is provided. The method includes: obtaining compressed data representing multiple ultrasonic signals acquired from ultrasonic detection of an object; applying a trained machine learning model to the compressed data to reconstruct multiple amplitudes associated with the multiple ultrasonic signals, and generating multiple reconstructed amplitude signals having the multiple reconstructed amplitudes associated therewith; and outputting the multiple reconstructed amplitude signals.

[0008] In some embodiments, multiple ultrasound signals are acquired from a full mastric capture (FMC) scan of the object being tested, which includes sequentially emitting a first ultrasound wave from each of multiple elements of a multi-element ultrasound probe toward the object being tested, and simultaneously receiving a second ultrasound wave scattered and / or reflected by the object being tested at the multiple elements in response to the emission of the first ultrasound wave.

[0009] In some embodiments, the multiple ultrasound signals include multiple amplitude time trajectories, providing a time-series representation of multiple second ultrasound waves.

[0010] In some embodiments, the multiple amplitude time trajectories include multiple basic A-scans.

[0011] In some embodiments, obtaining compressed data includes encoding multiple amplitude time trajectories using one-bit quantization.

[0012] In some embodiments, obtaining compressed data includes recording at least one of the rising phase leading edge position and the falling phase leading edge position associated with a plurality of ultrasound signals.

[0013] In some embodiments, obtaining compressed data includes receiving a plurality of ultrasound signals acquired in a compressed format.

[0014] In some embodiments, the machine learning model is a U-NET convolutional autoencoder neural network.

[0015] In some embodiments, the method further includes: performing multiple finite element simulations to generate simulated ultrasonic testing data, generating a training dataset based on the simulated ultrasonic testing data, and training a machine learning model using the training dataset.

[0016] In some embodiments, generating a training dataset includes encoding simulated ultrasonic detection data using one-bit quantization.

[0017] In some embodiments, generating the training dataset further includes applying random noise to the simulated ultrasonic detection data before encoding.

[0018] According to another aspect, a system for reconstructing ultrasonic detection data is provided. The system includes a processing unit and a non-transient memory communicatively coupled to the processing unit and including computer-readable program instructions executable by a computing unit for: obtaining compressed data representing multiple ultrasonic signals acquired from ultrasonic detection of an object; applying a trained machine learning model to the compressed data to reconstruct multiple amplitudes associated with the multiple ultrasonic signals and generating multiple reconstructed amplitude signals having the multiple reconstructed amplitudes associated therewith; and outputting the multiple reconstructed amplitude signals.

[0019] In some embodiments, multiple ultrasound signals are acquired from a full mastric capture (FMC) scan of the object being tested, which includes sequentially emitting a first ultrasound wave from each of multiple elements of a multi-element ultrasound probe toward the object being tested, and simultaneously receiving a second ultrasound wave scattered and / or reflected by the object being tested at the multiple elements in response to the emission of the first ultrasound wave.

[0020] In some embodiments, the multiple ultrasound signals include multiple amplitude time trajectories, providing a time-series representation of multiple second ultrasound waves.

[0021] In some embodiments, computer-readable program instructions may be executed by a processing unit for encoding multiple amplitude time trajectories using one-bit quantization to obtain compressed data.

[0022] In some embodiments, computer-readable program instructions may be executed by a processing unit to record at least one of the rising phase leading edge position and the falling phase leading edge position associated with a plurality of ultrasound signals to obtain compressed data.

[0023] In some embodiments, the machine learning model is a U-NET convolutional autoencoder neural network.

[0024] In some embodiments, computer-readable program instructions may be executed by a processing unit for performing multiple finite element simulations to generate simulated ultrasonic testing data, generating a training dataset based on the simulated ultrasonic testing data, and training a machine learning model using the training dataset.

[0025] In some embodiments, computer-readable program instructions may be executed by a processing unit to encode simulated ultrasonic detection data using one-bit quantization to generate a training dataset.

[0026] In some embodiments, computer-readable program instructions may be executed by a processing unit to apply random noise to the simulated ultrasonic test data prior to encoding.

[0027] Those skilled in the art will understand, upon reading this disclosure, many other features and combinations thereof related to the embodiments described herein. Attached Figure Description

[0028] In the diagram,

[0029] Figure 1A This is a block diagram of an example nondestructive testing (NDT) system according to one embodiment;

[0030] Figure 1B yes Figure 1A A block diagram of the processing system;

[0031] Figure 2 It is used as a diagram. Figure 1A A schematic diagram of the NDT system performing total focusing method (TFM) imaging;

[0032] Figure 3 This is a schematic diagram of an example U-NET architecture according to one embodiment;

[0033] Figure 4A It shows the basis Figure 1A Example full matrix capture (FMC) images obtained from finite element simulation of the NDT system;

[0034] Figure 4B It shows the basis Figure 4A Example TFM image obtained from FMC image;

[0035] Figure 5A , Figure 5B and Figure 5C The diagram illustrates the use of Figure 1A Experimental results obtained from the NDT system;

[0036] Figure 6 This is a flowchart of a method for reconstructing ultrasonic test data according to one embodiment;

[0037] Figure 7 It is used to implement Figure 1A NDT systems and / or Figure 6 A block diagram of an example computing device for the method.

[0038] It should be noted that in the accompanying drawings, the same features are identified by the same reference numerals. Detailed Implementation

[0039] Acoustic inspection, such as ultrasound-based inspection (also known as ultrasonic testing or ultrasound detection), can include focusing or beamforming techniques to help construct data maps or images representing regions of interest within the tested sample. As used herein, the term "ultrasound" or "ultrasound" refers to a technique that uses high-frequency sound waves (typically above 20 kHz) to detect defects (e.g., cracks, voids, or bulges) and measure material thickness without damaging the object being inspected. Using an array of ultrasonic transducer elements can include employing a phased array beamforming method, which may be referred to as Phased Array Ultrasound Testing (PAUT). For example, delay-and-sum beamforming techniques can be used, such as involving coherent summation of the time-domain representations of the received acoustic signals from the individual transducer elements or apertures. Specifically, a method called "phase coherence imaging" (PCI) beamforming can be used, which involves coherent summation of a normalized or quantized data representation corresponding to phase information. In another approach, Total Focusing Method (TFM) beamforming technology is used, where one or more elements in the array (or apertures defined by these elements) are used to emit ultrasonic pulses, while other elements are used to receive scattered and / or reflected ultrasonic energy, constructing a time-series (e.g., A-scan) matrix that corresponds to a series of transmit-receive cycles, where the emission occurs at different elements (or corresponding apertures) in the array. This TFM method, which obtains A-scan data for each element (or each defined aperture) in the array, can be called "full matrix capture" (FMC) technology.

[0040] During FMC acquisition, binarization (or any other suitable quantization technique) can be used to compress the amount of data associated with the time-series signal. The term "binarization" as used herein refers to a data reduction technique where an analog signal is quantized (or compressed) between two distinct levels (e.g., 0 and 1). This can also be called single-bit digitization. This technique has proven useful because FMC acquisition generates a large amount of data, and FMC acquisition signals are typically digitized with 12 to 16 signed bits (or any other suitable number of bits), posing significant challenges to data management. For dense probes (e.g., those with 64 or 128 transducer elements) or matrix probes including arrays of transducer elements (e.g., 8x8, 11x11, etc.), the amount of data generated by FMC acquisition is even more substantial. By reducing the number of bits (e.g., encoding the FMC acquisition signal using one-bit quantization), data compression can be achieved, thereby significantly reducing the storage space required for FMC-related data.

[0041] While this paper mentions a single-bit digitization or binarization scheme for compressing data acquired during the inspection of an object, it should be understood that other compression methods are also envisioned. For example, more aggressive compression methods can be applied, including but not limited to recording the rising and / or falling phase leading edge positions associated with the acquired signal, which would further eliminate redundant 0s and 1s in the alternating positive and negative plateau regions separating the acquired signal.

[0042] However, using data compression (e.g., compressing acquired A-scan data using binarization) can lead to the loss of amplitude information, even if this information may still prove useful in some applications. Specifically, while binarized data may make it easier to detect some small defects, the loss of amplitude information has a side effect: images generated using TFM can degrade the visualization of reflective structures in the sample under inspection. Therefore, using data compression can make the visual interpretation of A-scan data more complex.

[0043] To address this issue, this paper proposes a machine learning approach to reconstruct amplitude data acquired during FMC (Focused Motion Control), where the amplitude time series is regenerated from compressed (e.g., binarized) time-domain data representing the FMC acquisition signal. The regenerated time series data illustratively has a larger amplitude resolution (e.g., 12 to 16 bits) than the compressed data (1 bit). The proposed method relies on the spatial / temporal coherence between adjacent time series generated by the FMC format: large amplitude signals (with extended spatial coherence) tend to appear more frequently in adjacent channels than weak signals (with limited spatial coherence), while weak signals are quickly lost in noisy backgrounds. Once output, the regenerated amplitude time series data can then be used for other processing, such as PAUT (Parallel Approach to Waveform) or TFM (Temporal Beamforming). For example, a TFM image can be generated based on the regenerated amplitude data.

[0044] Figure 1A An exemplary non-destructive testing (NDT) system 100 according to one embodiment is illustrated. The NDT system 100 can be used to inspect or otherwise test various components or objects of inspection, such as component 102. Component 102 can be any suitable processed or manufactured component and can be made of any suitable material. By way of the first set of examples, component 102 can be a part of an engine, vehicle, personal electronic device, etc. By way of another set of examples, component 102 can be a composite component, polymer composition, or the like. Other types of components are also contemplated, and the above examples should be understood as non-limiting examples.

[0045] The NDT system 100 consists of an ultrasonic testing (UT) tool 104 and a processing system 106 (see below for reference). Figure 1B (Further details omitted). The UT tool 104 is operable to perform ultrasonic testing on the pair of components 102 and can be configured to perform any suitable type of ultrasonic testing. For example, the UT tool 104 can be configured to generate and emit one or more plane waves, or another suitable type of ultrasonic (or ultrasound) and / or sound waves, which are directed at and impact the component 102. In the illustrated embodiment, the UT tool 104 generates ultrasonic waves, which can be generated by the UT tool 104 in any suitable manner, such as using a multi-element ultrasonic phased array (PA) probe (see reference). Figure 2 (Ref. 206). In some embodiments, ultrasound may use an array of piezoelectric transducer elements (see reference 206). Figure 2The piezoelectric transducer element array (207) can be used as part of or separate from a multi-element ultrasonic phased array probe. As needed, the transducer elements can be commanded to generate a single plane wave or multiple plane waves of the FMC as one or more focused beams. The UT tool 104 (e.g., using one or more transducer elements) then receives and records the returned ultrasonic waves (i.e., waves scattered and / or reflected by component 102) as a function of time.

[0046] In one embodiment, according to the FMC technique, each transducer element of the PA probe is sequentially excited, and at least some transducer elements simultaneously receive the returned waves, thereby generating a basic A-scan matrix for the PA probe. In some embodiments, all transducer elements receive the returned waves. In other embodiments, all transducer elements except the transmitting transducer elements receive the returned waves. However, it should be understood that sparse FMC may also be applicable, where some transducer elements are skipped (e.g., every two (2) elements, every four (4) elements, every eight (8) elements, etc.), such that only a predetermined number of transducer elements are excited. Furthermore, it should be understood that while the use of the FMC technique and the A-scan data obtained from the transducer elements of the PA probe are mentioned herein, other embodiments may also be applicable. In one embodiment, B-scan data can be obtained using the UT tool 104. The B-scan format can be used to plot thickness as a function of time or position while scanning the transducer (e.g., the PA probe) along assembly 102 to provide its depth profile. By correlating the ultrasonic data with the actual transducer location, a scaled view can be drawn, and the data can be correlated and traced to specific areas of assembly 102. Any suitable device can be used for position tracking. For example, electromechanical devices such as encoders can be used to record the position relative to a user-defined scan pattern and index resolution for each data acquisition. Therefore, it should be understood that the systems and methods described herein do not require obtaining all possible transmit-receive combinations of transducer elements in all embodiments. Spatial proximity between the transducer elements (or groups of transducer elements) of the probe may be sufficient. In fact, signal redundancy in multiple scans obtained from adjacent transducer elements (or groups of transducer elements) allows the processing system 106 to distinguish noise from data associated with a coherent structure having well-defined amplitudes.

[0047] The returned ultrasound waves are then recorded and can be used (using any suitable method) to evaluate various properties of component 102, including detecting the presence of defect 108. The UT tool 104 can perform partial or full evaluation of component 102 itself, or it can provide processing system 106 with information about the recorded ultrasound waves (referred to as “ultrasound information”). In some embodiments, as will be further described below, processing system 106 includes one or more artificial intelligence (AI) based tools such as 110. Processing system 106 can be configured to analyze the ultrasound information in any suitable manner and can produce various results or conclusions about the properties of component 102. This may include information about the dimensions of component 102, the material composition of component 102, etc. Furthermore, analysis of the ultrasound information may reveal the presence of defect 108 in component 102 when defect 108 is present. Figure 1A In the example shown, defect 108 is a "flat-bottomed hole." However, it should be understood that the techniques described herein can be applied to other types of defects, including but not limited to electrical discharge machining (EDM) cracks and fatigue cracks, side-drilled holes (SDH), metallic inclusions or delamination in polymers, voids, etc. In some cases, ultrasonic information can be processed to generate an image of component 102, which can be presented (e.g., via a screen or other display device) for evaluation, for example, to a trained operator. In other cases, processing system 106 can be configured to detect defect 108 based on analysis of ultrasonic information. Parties responsible for the production or use of component 102 may choose to discard or repair component 102 upon detection of defect 108, or may choose to permit the use of component 102 upon detection that it is free of defects.

[0048] like Figure 1B As shown, in one embodiment, the processing system 106 includes an input module 112, an ultrasonic data acquisition module 114, an optional data compression module 116, a data reconstruction module 118, a training module 120, a component evaluation module 122, and an output module 124. At least the data reconstruction module 118 may include one or more AI-based tools (see reference). Figure 1A (110 in the middle), these tools can be trained using training module 120.

[0049] Apart from Figure 1A and Figure 1B In addition, now also referencing Figure 2 . Figure 2 The illustration shows the use of Figure 1AAn exemplary diagram 200 illustrates the principle of TFM imaging performed by an NDT system. In one embodiment, ultrasound reflected and / or scattered by component 102 can be acquired using FMC technology, which involves acquiring a comprehensive dataset by capturing all possible transmit-receive combinations from the transducer element array, thereby providing a dense and highly detailed ultrasound data matrix consisting of amplitude time trajectories (e.g., A-scans). Those skilled in the art will understand that FMC can be analogized to a three (3)-dimensional matrix, where the first dimension corresponds to the time axis, the second dimension corresponds to the transmitting transducer element, and the third dimension corresponds to the receiving transducer element. The first dimension is controlled by the number of acquired time increments and the sampling frequency, while the second and third dimensions depend on the number of transducer elements N, with a total size of [missing information]. Then, TFM is used to process the rich dataset, focusing the wave onto a predefined pixel grid, for example... Figure 2 The pixel grid 202 is used. Using pixel grid 202, the TFM at each pixel location (x,z) can be calculated using the following equation:

[0050] (1)

[0051] in It is a given time The analyzed signal at a given time This corresponds to the flight time between transmitter k and pixel, and back to receiver l.

[0052] When using a wedge 204 placed on the component under test 102, multiple beam paths can be considered. Since most PA probes (such as 206) are longitudinally polarized, the wave paths in the wedge 204 are typically omitted in the nomenclature. T represents the transverse mode, and L represents the longitudinal mode. The symbols “TT,” “TT-T,” “T-TT,” and “TT-TT” used herein refer to the wave paths used when generating TFM images. A TT-T path indicates that the transmitted shear beam travels downwards to the rear wall 210a of component 102, reflects back to the defect 108, and then reflects back to the probe 206. A TT-TT path indicates that the transmitted shear beam travels downwards to the rear wall 210a, upwards to the defect 108, where it is reflected downwards back to the rear wall 210a, and then upwards back to the probe 206. For example, Figure 2 The TT-T path 208 indicates that there is a jump (with a reflection from the rear wall 210a of component 102 using a lateral mode) before defect 108, and there is a T direct path from defect to front wall 210b of component 102.

[0053] Reference Figure 1A and Figure 1BData captured during FMC acquisition is received illustratively at input module 112 and recorded at ultrasonic data acquisition module 114. The amount of recorded data can be reduced by compression, i.e., by reducing (or compressing) the amplitude of the FMC acquired signal to a compressed (e.g., binary) value. This compression can be performed by data compression module 116. In one approach, multiple analog signals can be acquired from transducer elements and received at input module 112. The received analog signals can then be processed at ultrasonic data acquisition module 114 (e.g., using a low-noise amplifier, time gain compensation, low-pass filter, and analog-to-digital converter) to generate sampled data. Data compression module 116 can then apply a compression algorithm to the sampled data to obtain compressed data. In another approach, a selective single-bit digitization process can be considered, in which multiple analog signals are acquired from transducer elements and received at input module 112. The analog signal is then processed using a low-noise amplifier and a low-pass filter at the ultrasonic data acquisition module 114, and then compressed (e.g., binarized) at the data compression module 116 using a comparator, retaining either the full compressed version of the signal (e.g., binary version) or only the rising and / or falling edges of the signal. However, it should be understood that any method suitable for ultrasonic data reduction can be applied and performed by the data compression module 116. It should also be understood that while the use of data compression or quantization (e.g., binarization) to compress (e.g., binarize) the FMC acquisition signal and generate compressed (e.g., binary) data is mentioned, compressed data can also be obtained directly using a suitable (e.g., binary) acquisition system (not shown) configured to directly acquire the FMC acquisition signal in compressed (e.g., binary) format. For example, the ultrasonic data acquisition module 114 can be part of a binary acquisition system and can be configured to output binarized FMC data, thereby reducing the need for the data compression module 116 to perform data compression on the analog FMC acquisition signal.

[0054] Amplitude data can then be reconstructed from the compressed (e.g., binarized) data obtained from the ultrasonic data acquisition module 114 or the data compression module 116 using the data reconstruction module 118 (by regenerating the amplitude time series). For this purpose, one or more AI-based tools 110 included in the data reconstruction module 118 can be used. These AI-based tools 110 include at least one machine learning (ML) model configured to reconstruct amplitude information from compressed data (e.g., binarized FMC). In one embodiment, the ML model is a U-NET autoencoder neural network. However, it should be understood that other types of ML models (e.g., other types of neural networks) and / or other types of AI-based tools are also considered. However, using an ML model with a multi-layered architecture may be ideal. Using the training module 120, the ML model (e.g., neural network) described herein can be trained (to reconstruct amplitude data from compressed data) using any suitable dataset, in any suitable manner, and with any suitable number of epochs, depending on the implementation. The amount of training performed by the training module 120 can be selected based on the required training time and / or available computing power. A method for training U-NET will be described in more detail below. The output of the trained ML model can be provided to the component evaluation module 122, which is configured to analyze the data in any suitable manner and produce various results or conclusions about the characteristics of component 102. The results or conclusions are then provided to the output module 124 so that they can be presented in any suitable manner (e.g., via a screen, display, or other suitable output device).

[0055] Now refer to Figure 3The diagram illustrates an exemplary U-NET autoencoder neural network architecture 300 according to one embodiment. As used herein, the term "autoencoder" refers to an artificial neural network that uses an unlabeled database. An autoencoder architecture consists of an encoder and a decoder. The encoder reduces the dimensionality of the input data provided to it in order to extract the most important features from the data into a latent space. The decoder uses the latent space to reconstruct the input data. Once trained, these autoencoder networks can be used to generate data similar to the original dataset. Autoencoders can also be used for image or signal denoising by using low-noise or noise-free data for loss calculation. As used herein, the term "U-NET" refers to a special case of a convolutional autoencoder with a symmetrical U-shaped structure. The encoder typically consists of repeated convolution and max-pooling operations, allowing the extraction of important features from the latent space. These features are then upsampled by the corresponding decoder. The encoder features are preserved and then concatenated to the corresponding decoder layers via connection paths. This allows the neural network to better convey the context of the input data along different structural layers, thereby improving the accuracy of the reconstructed data.

[0056] like Figure 3 As shown, the U-NET 300 includes an encoder 302 and a decoder 304, whose parameters are defined after multiple tests to achieve the desired efficiency level (e.g., p-level). In the illustrated example, the input to the U-NET 300 (also referred to herein as the "input image") consists of multiple binarized FMCs of size 4096 (number of samples) × 1024 (32 element probes). The encoder 302 uses a combination of convolutional and pooling layers to extract and capture features from the input image while reducing its spatial dimensionality. The encoder 302 consists of two consecutively applied 3x3 convolutions, each followed by a ReLU activation function (indicated by "Conv. ReLU" arrow 306) to increase the number of channels in order to capture higher-level features from the input image. Downsampling is performed using a 2x2 max pooling operation with a stride of 2 (indicated by "Max. pool" arrow 308) to halve the spatial dimensionality of the feature map.

[0057] Decoder 304 uses a combination of convolution and upconvolution operations to combine learned features and upsample the input feature map until a segmentation map is generated. Each step in decoder 304 begins with upsampling using a 2x2 transposed convolution (indicated by arrow 310, "Convolution") to upsample the spatial dimensions by a factor of two. Corresponding feature maps from encoder 302 are concatenated (indicated by arrow 312, "Concatenate"). Multiple double 3x3 convolutions (each followed by ReLU activation, "Arrow 306") are then applied to reduce the number of channels. For the final layer, a 1x1 convolutional layer (indicated by arrow 314, "Transpose Conv") is applied. Padding 316 is further added in each convolutional layer to avoid losing any boundary pixels and to eliminate the cropping required in the original version. In the illustrated embodiment, the output of U-NET 300 consists of an amplitude FMC of size 4096 x 1024. It should be understood that although the number of transducer elements is limited to 32 due to the requirements of the graphics memory during network training, any suitable number of elements other than 32 can be used per probe. It should also be understood that... Figure 3 The U-NET architecture shown is for illustrative purposes only, and other embodiments may also be applicable. For example, the number of layers can be modified (e.g., by adding a layer), or the size of the convolutional layers can be changed.

[0058] For example, U-NET 300 illustratively uses a classification method, thus being trained to evaluate the appropriate category for each pixel in a one-bit quantized input image, which forms the input data provided to U-NET 300. However, it should be understood that while this document describes U-NET 300 as being trained using classification, U-NET 300 can alternatively use a regression method. The input image is generated by compressing (e.g., binarizing) a set of fixed dynamic range amplitude signals (e.g., FMC acquisition signals) and evaluating the appropriate category over a fixed dynamic range (e.g., 16-bit dynamic range) associated with these signals. In one embodiment, U-NET 300 is trained using a training dataset comprising simulated ultrasound detection data generated in any suitable manner. The simulated data includes a dataset of simulated ultrasound detections performed using a simulation model representing one or more ultrasound detection scenarios in a simulated version of an experimental NDT setup involving an object (with one or more detectors) being detected. Simulated data can be generated using finite element (FE) simulations (e.g., on a graphics processing unit (GPU) accelerated solver or any suitable simulation suite) to replicate the FMC acquisition process described above. Specifically, FE simulations can be used to generate a set of amplitude signals with a fixed dynamic range (e.g., 16 bits or less) that can be used as truth values. In one embodiment, random noise is applied (i.e., added) to the amplitude signals obtained from the FE simulation, and the results are compressed (e.g., binarized by applying one-bit quantization or compressed in any other suitable manner) to generate simulated data ultimately used for training the U-NET 300. Adding random noise to the signal transforms a coherent but low-valued signal into an incoherent structure. This, in turn, allows the simulation to yield experimentally obtained results. In some cases, the entire training dataset can be generated via FE simulation. In other cases, a relatively large portion of the training dataset can be generated via FE simulation, and the training dataset can be augmented with a relatively small portion based on experimental data acquired using the NDT system 100. Other implementations may also be applicable.

[0059] Once U-NET 300 training is complete, the trained U-NET 300 can be validated using any suitable procedure. In one embodiment, experimental results not used during training (e.g., time-domain data binarized from experiments) can be presented to U-NET 300. The ability of U-NET 300 to correctly reconstruct amplitude information from the experimental compressed data is then evaluated. The output of the trained U-NET 300 (i.e., the predicted amplitude data) can then be compared with the expected output (i.e., the amplitude associated with the signal used as the ground truth). Any suitable loss function can be used for the comparison, including but not limited to the Mean Square Error (MSE) loss function (used to calculate the average distance), to determine the error rate (or loss) between the U-NET's output and the ground truth. Backpropagation is then performed to update the parameters (e.g., weights) of U-NET 300 based on the error rate.

[0060] In one embodiment, the FE model used for FE simulation consists of a 25 mm homogeneous isotropic steel plate with a density of ρ = 8000 kg / m³. 3 The Young's modulus is E = 210 GPa, and the Poisson's ratio is ν = 0.30, modeled in two dimensions (2D). The transmitting system consists of a PA probe with a 5 MHz linear array of 32 elements, associated with an ultrasonic PA wedge. Only the first 32 elements are modeled to constrain the size of the FMC. A 5 mm diameter flat-bottom hole (FBH) is presented on the back wall of the plate, its position varying from 25 mm to 75 mm in front of the probe centerline (5 increments), and its depth varying from one-quarter to one-half of the plate thickness (4 increments), corresponding to a total of 20 models. Absorbing boundaries are added at both ends (x-axis) of the plate to reduce the model size by minimizing edge echoes. The excitation signal used is a 5-cycle, 5 MHz-centered Hanning-windowed toneburst. The length of the excitation signal is similar to the time response of the PA probe elements when excited by a half-cycle pulse (burst) generated by most industrial inspection equipment. The element size is defined to ensure that there are at least 25 elements for each shortest wavelength, and to ensure convergence of results, the time step is chosen so that the mode does not skip elements within the time increment.

[0061] The obtained FMC was then resampled at a sampling frequency of 50 MHz and reorganized into a 4096×1024 dimension (32-element FMC). Initial impulses were canceled out as they were unavailable parts of the matrix. Filtered white noise was added within the probe's bandwidth to approximate the signal available in the experiment. The FMC was then binarized by associating positive values ​​with high states and zero or negative values ​​with low states. The binary FMC was used as input to the neural network, while the standard (non-binary) FMC was used for loss calculation. Specifically, the output of the U-NET 300 was compared to the corresponding standard FMC using the MSE loss function. Several learning rates were considered, and one was selected. This value. The training epochs are controlled by an early stop procedure. When the validation loss stops improving within a predetermined number of epochs (e.g., 100 epochs), the early stop procedure halts the training process. When this occurs, the weights of U-NET 300 are set to the configuration that minimizes the validation loss.

[0062] In one embodiment, using FE simulations to generate training data can facilitate the generation and control of parameters such as sample thickness and the nature, size, and magnitude of defects. However, it should be understood that although this document mentions training the U-NET 300 using simulation data generated using FE simulations, the U-NET 300 can also be trained using simulation data generated in any other suitable manner. It should also be understood that in some embodiments, the U-NET 300 can be trained using experimental data instead of simulation data.

[0063] To evaluate the ability of the system and method described in this paper to reconstruct FMC amplitude information from compressed data (e.g., binary data), the TFM algorithm was used with a basic (or standard) FMC and an FMC reconstructed using the system and method described in this paper. The two images were then compared using the Structural Similarity Index Measure (SSIM), an index used to compare one image with another reference image. SSIM is typically calculated over windows f and g in both images using the following equation:

[0064]

[0065] in, and These are the means of windows f and g, respectively. and It is the variance in windows f and g. Let f be the covariance of f and g, and c1 and c2 be two constants used to stabilize the function when the denominator is close to zero. In general, and Where k1 = 0.01, k2 = 0.03, and L is the dynamic range of the pixel values. The values ​​of the TFM image are normalized to between 0 and 1, so L = 1. Once all local SSIMs are calculated, the global SSIM value can be obtained by averaging the local SSIM values. The more similar the two images being compared, the closer the SSIM value will be to 1. Conversely, for two very different images, the value will tend to 0.

[0066] Figure 4A The paper presents an example of reconstruction using a training dataset generated based on the above FE simulation, showing Figure 400 of the FMC obtained through simulation, its binarized version Figure 410, and the version reconstructed by trained U-NET 300, Figure 420. Figure 4B The diagram illustrates the basis Figure 4A The TFM images were calculated from the FMC images. Figure 430 corresponds to the TFM images obtained using the TT wavepath and the simulated FMC (Figure 400), while Figure 440 uses the reconstructed FMC (Figure 420). Figure 450 shows a comparison between the two TFM images obtained using the SSIM index. As can be seen from Figure 450, the main errors (low SSIM values) appear at the edges of the indicators, where defects or artifacts are present. As for the larger indicators, the centers are reconstructed with high fidelity (SSIM values ​​close to 1). Despite the observed errors, the two TFM images remain very similar, and it is difficult to determine with the naked eye which image was created using simulated data and which image was generated by the U-NET 300.

[0067] Figure 5A , Figure 5B and Figure 5C The experimental results obtained using the system and methods described herein are illustrated. Various configurations of the samples were evaluated, including different block thicknesses and various defect characteristics (e.g., size, orientation, nature, and number). The first plate examined was a 19 mm steel plate with four (4) EDM opening notches of gradually increasing inclination, each 0.3 mm wide and extending vertically for 5 mm, relative to the back wall of the plate (i.e., relative to the horizontal plane or Figure 2 The tilt angles of the x-axis (in the figure) are 90°, 80°, 70°, and 60°. The results are as follows: Figure 5A As shown in Figures 502a, 502b, 502c, and 502d, TFM images obtained through experimental measurements are displayed. These were obtained using U-NET (reference). Figure 3The TFM images obtained from the reconstructed data (300°) are shown in Figures 504a, 504b, 504c, and 504d. The corresponding SSIMs are shown in Figures 506a, 506b, 506c, and 506d. Figures 502a, 504a, and 506a were obtained using measurements of a 90° vertical crack. Figures 502b, 504b, and 506b were obtained using measurements of an 80° vertical crack. Figures 502c, 504c, and 506c were obtained using measurements of a 70° vertical crack. Figures 502d, 504d, and 506d were obtained using measurements of a 60° vertical crack. To better visualize the crack inclination angle under different conditions, the wavepaths used for reconstruction differ according to the following configuration: TT-T paths were used for 90° and 80° cracks, and TT-TT paths were used for 70° and 60° cracks. Figure 5A It can be observed that the errors in the TFM image are mainly present at the edges of the markers rather than their positions, so their impact on the size and location of the defects is negligible. It can also be seen that the number or intensity of artifacts appears to be lower when using the reconstructed data compared to the experimental data.

[0068] The second plate used comprises four (4) groups, each group consisting of two SDHs manufactured by EDM, each 0.3 mm in diameter, vertically spaced 5 mm apart, and located at the center of the thickness. The axes connecting the SDHs are inclined at 90°, 80°, 70°, and 60° relative to the bottom of the workpiece. Figure 5B Various TFM and SSIM images are shown: Figures 512a, 512b, 512c, and 512d show TFM images obtained through experimental measurements; Figures 514a, 514b, 514c, and 514d show TFM images obtained using data reconstructed from U-NET 300; and Figures 516a, 516b, 516c, and 516d show the corresponding SSIMs. The organization of the subgraphs is similar to... Figure 5A Similarities: Figures 512a, 514a, and 516a are obtained for a dip angle of 90°; Figures 512b, 514b, and 516b are obtained for a dip angle of 80°; Figures 512c, 514c, and 516c are obtained for a dip angle of 70°; and Figures 512d, 514d, and 516d are obtained for a dip angle of 60°. Figure 5A The difference in the results is that all images use the TT-TT path. According to Figure 5B It can be observed that the original TFM and the TFM obtained using the reconstructed FMC show the marker in the same location in both cases. Due to SSIM, the error of the marker edge is visible, but it cannot be distinguished directly on the TFM image.

[0069] Now for reference Figure 5CFurther experimental measurements were performed on a 24.54 mm thick steel bar containing a fatigue crack that occupied 35% of the plate thickness (8.6 mm in length). This allows for the detection of real defects with more irregular profiles compared to EDM machined notches or SDH. Figure 522a shows a TFM image obtained for the experimental FMC, Figure 522b shows a TFM image obtained using the FMC reconstructed with U-NET 300, and Figure 522c shows the corresponding SSIM. In this case, it can be noted that the fatigue crack is clearly visible in both TFM images (see Figures 522a and 522b). These reconstructions are similar to those obtained with a perfectly machined notch, but less similar to those with a fatigue crack. Furthermore, it is also possible to... Figure 5A and Figure 5B The same observation was made, namely that there was an error in the amplitude of the marker edge, but no positioning error. In addition, the reconstructed image (Fig. 522b) appears to have fewer artifacts than the original image (Fig. 522a).

[0070] refer to Figure 6 The illustration shows a method 600 for reconstructing ultrasonic detection data. Method 600 employs a machine learning model, such as... Figure 3 The U-NET 300 in Figure 1. Step 602 includes obtaining compressed (e.g., binary) data representing multiple ultrasound signals acquired from the object being examined (e.g., component 102 in Figure 1). In one embodiment, the ultrasound signals are acquired from a full mastric capture (FMC) scan of the object being examined. An FMC scan can be performed using the NDT system 100 in Figure 1. As described above, such an FMC scan can include each element from multiple elements of a multi-element ultrasound probe (e.g., Figure 2 The PA probe 206 shown has transducer elements 207 that sequentially emit first ultrasonic waves toward the object being detected, and simultaneously receive second ultrasonic waves scattered and / or reflected by the object being detected at multiple elements in response to the emission of the first ultrasonic waves. In one embodiment, the multiple ultrasonic signals include multiple amplitude time trajectories, providing a time-series representation of multiple second ultrasonic waves. For example, the multiple amplitude time trajectories may include multiple basic A-scans as described above. In step 602, compressed data can be obtained by encoding the amplitude time trajectories using one-bit quantization. In other embodiments, obtaining compressed data includes receiving ultrasonic signals acquired in a compressed format (e.g., acquired in binary format using a binary acquisition system as described above). As previously mentioned, B-scans may also be obtained in other embodiments.

[0071] Step 604 includes applying a trained ML model to the compressed data to reconstruct multiple amplitudes associated with multiple ultrasound signals and generating multiple reconstructed amplitude signals with the multiple reconstructed amplitudes associated with them. The ML model can be a neural network, more specifically a convolutional neural network, such as U-NET (e.g., Figure 3 (U-NET 300 in the example). U-NET can be trained using a training dataset generated from simulated ultrasonic testing data, which is produced through multiple finite element simulations as described above. In one embodiment, the simulated ultrasonic testing data is encoded using one-bit quantization to generate the training dataset. In another embodiment, random noise is applied to the simulated ultrasonic testing data prior to encoding.

[0072] Step 606 includes outputting multiple reconstructed amplitude signals. These signals can be output in any suitable format and via any suitable output device (e.g., screen, display, etc.) for subsequent evaluation of one or more characteristics of the detected object.

[0073] In one embodiment, the systems and methods described herein can facilitate the use of advanced ultrasound NDT techniques (e.g., TFM) and simplify the interpretation of A-scans by allowing the reconstruction of the amplitude of signals that have been compressed (e.g., binarized) or previously stored in compressed (e.g., binary) formats. The systems and methods described herein can indeed help users visualize A-scans, as it is difficult for users to distinguish any features in a binarized A-scan. Furthermore, using the systems and methods described herein allows for a simplified front-end of the ultrasound inspection system by compressing (e.g., binarizing) the signal. Moreover, compressed (e.g., binarized) signals are much smaller than 12-bit or 16-bit quantized data, while retaining sufficient information for proper reconstruction. This saves digital space in system memory on the one hand; on the other hand, it optimizes data availability by enabling faster transmission on one or more of the system's transmission channels (i.e., buses). Therefore, the overall footprint of the inspection system's electronics can be reduced due to the smaller digitization of the acquired data. Furthermore, using compressed (e.g., binarized) data as the basic data unit reduces the electrical energy required to generate any usable signal from the component or part under test. In fact, higher excitation levels (i.e., high voltage) are typically required to generate a sufficient return signal for proper digitization of the acquired signal. By compression (e.g., binarization), the focus can be placed solely on detecting phase changes, thereby reducing the excitation level requirement. The system and method described herein can then retrieve amplitude data without the high voltage requirements of conventional systems.

[0074] refer to Figure 7The illustration shows a schematic diagram of an example computing device 700. As shown, the computing device 700 includes at least one processing unit (also referred to herein as a processor) 702, a memory 704 storing instructions 706, and at least one I / O interface (illustrated as "input" and "output"). One or more elements (including part or all of the UT tool 104 and / or part or all of the processing system 106) and / or methods (refer to...) of the NDT system (refer to 100 in Figure 1) described herein... Figure 6 All of the 600s in the diagram can be implemented in the form of computing devices 700. For simplicity, only one computing device 700 is shown in the diagram, but the system may include more computing devices 700, which users can operate to access remote network resources and exchange data. Computing devices 700 can be of the same type or different types. The components of computing devices 700 can be connected in various ways (including direct coupling, indirect coupling via a network), and can be distributed over a wide geographical area and connected via a network (this can be called "cloud computing").

[0075] Each processor 702 may be, for example, any type of general-purpose microprocessor or microcontroller, digital signal processing (DSP) processor, integrated circuit, field-programmable gate array (FPGA), reconfigurable processor, programmable read-only memory (PROM), or any combination thereof.

[0076] Memory 704 may include any suitable combination of computer memories, whether internal or external, such as random-access memory (RAM), read-only memory (ROM), compact disc read-only memory (CD-ROM), electro-optical memory, magneto-optical memory, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferroelectric RAM (FRAM), etc.

[0077] The I / O interface enables the computing device 700 to interconnect with one or more input devices (such as a keyboard, mouse, camera, touchscreen, and microphone) or with one or more output devices (such as a display and speakers).

[0078] In some embodiments, computing device 700 includes one or more network interfaces to enable computing device 700 to communicate with other components, exchange data with other components, access and connect to network resources, service applications, and perform other computing applications by connecting to networks (or networks) capable of carrying data. These networks include the Internet, Ethernet, plain old telephone service (POTS) lines, public switch telephone network (PSTN), integrated services digital network (ISDN), digital subscriber line (DSL), coaxial cable, fiber optic, satellite, mobile, wireless (e.g., Wi-Fi, WiMAX), SS7 signaling network, fixed line, local area network, wide area network, and other networks including any combination of these.

[0079] Throughout this disclosure, references to servers, services, interfaces, portals, platforms, or other systems formed by computing devices may be made multiple times. It should be understood that the use of such terms is considered to refer to one or more computing devices having at least one processor configured to execute software instructions stored on a computer-readable tangible, non-transient medium. For example, a server may include one or more computers operating as a web server, database server, or other type of computer server in a manner that fulfills the described roles, responsibilities, or functions.

[0080] Various embodiments may take the form of a software product. The software product may be stored on a non-volatile or non-transient storage medium, such as a compact disk read-only memory (CD-ROM), a USB flash drive, or a removable hard drive. The software product includes instructions that enable a computer device (personal computer, server, or network device) to perform the methods provided in the embodiments.

[0081] The embodiments described herein are implemented using physical computer hardware, including computing devices, servers, receivers, transmitters, processors, memory, displays, and networks. The embodiments described herein provide useful physical machines and specially configured computer hardware devices. The embodiments described herein relate to electronic machines and methods implemented by electronic machines suitable for processing and converting electromagnetic signals representing various types of information. The embodiments described herein are generally and holistically related to machines and their uses; and the embodiments described herein have no practical significance or applicability other than for use with computer hardware, machines, and various hardware components. Replacing the physical hardware specially configured to perform various actions with non-physical hardware (e.g., using thought processes) may significantly affect how these embodiments operate. Such limitations of computer hardware are clearly essential elements of the embodiments described herein, and they cannot be omitted or replaced with thought processes, which would otherwise have a substantial impact on the operation and structure of the embodiments described herein. Computer hardware is essential for implementing the various embodiments described herein, and it is not merely for performing steps in a fast and efficient manner.

[0082] This disclosure provides numerous exemplary embodiments. While each embodiment represents a single combination of inventive elements, other examples may include all possible combinations of the disclosed elements. Thus, if one embodiment includes elements A, B, and C, and a second embodiment includes elements B and D, other remaining combinations of A, B, C, or D may also be used. The term “connection” or “coupling” may include direct coupling (where two mutually coupled elements are in contact with each other) and indirect coupling (where at least one additional element is located between the two elements).

[0083] While these embodiments have been described in detail, it should be understood that various changes, substitutions, and modifications can be made to this document without departing from the scope defined by the appended claims. It should be understood that the above description and examples illustrated herein are for illustrative purposes only.

[0084] Furthermore, the scope of this application is not intended to be limited to the specific embodiments of the processes, machines, manufactures, material compositions, means, methods, and steps described in the specification. Those skilled in the art will readily understand from the disclosure of this invention that any existing or future processes, machines, manufactures, material compositions, means, methods, or steps that can perform substantially the same function or achieve substantially the same result as the corresponding embodiments described herein can be utilized. Therefore, the appended claims are intended to include such processes, machines, manufactures, material compositions, means, methods, or steps within their scope.

Claims

1. A method for reconstructing ultrasonic detection data, the method comprising: Obtain compressed data representing multiple ultrasonic signals acquired from ultrasonic testing of the object under test; The trained machine learning model is applied to the compressed data to reconstruct multiple amplitudes associated with the multiple ultrasound signals and generate multiple reconstructed amplitude signals with the multiple reconstructed amplitudes associated with them. as well as Output the multiple reconstructed amplitude signals.

2. The method according to claim 1, wherein, The plurality of ultrasound signals are acquired from a full mastric capture (FMC) scan of the object being tested, the full mastric capture scan of the object being tested comprising sequentially emitting a first ultrasound wave from each of a plurality of elements of a multi-element ultrasound probe toward the object being tested, and simultaneously receiving a second ultrasound wave scattered and / or reflected by the object being tested at the plurality of elements in response to the emission of the first ultrasound wave.

3. The method according to claim 2, wherein, The plurality of ultrasonic signals include a plurality of amplitude time trajectories, providing a plurality of time sequence representations of the second ultrasonic waves.

4. The method according to claim 3, wherein, The multiple amplitude time trajectories include multiple basic A-scans.

5. The method according to claim 3, wherein, Obtaining the compressed data includes encoding the plurality of amplitude time trajectories using one-bit quantization.

6. The method according to claim 1, wherein, Obtaining the compressed data includes recording at least one of the rising phase leading edge position and the falling phase leading edge position associated with the plurality of ultrasound signals.

7. The method according to any one of claims 1 to 4, wherein, Obtaining the compressed data includes receiving the plurality of ultrasound signals acquired in a compressed format.

8. The method according to any one of claims 1 to 7, wherein, The machine learning model is the U-NET convolutional autoencoder neural network.

9. The method according to any one of claims 1 to 8, further comprising: Multiple finite element simulations are performed to generate simulated ultrasonic testing data, a training dataset is generated based on the simulated ultrasonic testing data, and the machine learning model is trained using the training dataset.

10. The method according to claim 9, wherein, Generating the training dataset includes encoding the simulated ultrasonic detection data using one-bit quantization.

11. The method according to claim 10, wherein, Generating the training dataset also includes applying random noise to the simulated ultrasonic detection data before the encoding.

12. A system for reconstructing ultrasonic detection data, the system comprising: Processing unit; as well as A non-transient memory, communicatively coupled to the processing unit and including computer-readable program instructions executable by the processing unit for: Obtain compressed data representing multiple ultrasonic signals acquired from ultrasonic testing of the object under test; The trained machine learning model is applied to the compressed data to reconstruct multiple amplitudes associated with the multiple ultrasound signals and generate multiple reconstructed amplitude signals with the multiple reconstructed amplitudes associated with them. as well as Output the multiple reconstructed amplitude signals.

13. The system according to claim 12, wherein, The plurality of ultrasound signals are acquired from a full mastric capture (FMC) scan of the object being tested, the full mastric capture scan of the object being tested comprising sequentially emitting a first ultrasound wave from each of a plurality of elements of a multi-element ultrasound probe toward the object being tested, and simultaneously receiving a second ultrasound wave scattered and / or reflected by the object being tested at the plurality of elements in response to the emission of the first ultrasound wave.

14. The system according to claim 13, wherein, The plurality of ultrasonic signals include a plurality of amplitude time trajectories, providing a plurality of time sequence representations of the second ultrasonic waves.

15. The system according to claim 14, wherein, The computer-readable program instructions, which can be executed by the processing unit, are used to encode the plurality of amplitude time trajectories using one-bit quantization to obtain the compressed data.

16. The system according to claim 12, wherein, The computer-readable program instructions, which can be executed by the processing unit, are used to record at least one of the rising phase leading edge position and the falling phase leading edge position associated with the plurality of ultrasound signals to obtain the compressed data.

17. The system according to any one of claims 12 to 16, wherein, The machine learning model is the U-NET convolutional autoencoder neural network.

18. The system according to any one of claims 12 to 17, wherein, The computer-readable program instructions can be executed by the processing unit to perform multiple finite element simulations to generate simulated ultrasonic testing data, generate a training dataset based on the simulated ultrasonic testing data, and train the machine learning model using the training dataset.

19. The system according to claim 18, wherein, The computer-readable program instructions, which can be executed by the processing unit, are used to encode the simulated ultrasonic detection data using one-bit quantization to generate the training dataset.

20. The method according to claim 19, wherein, The computer-readable program instructions can be executed by the processing unit to apply random noise to the simulated ultrasonic detection data prior to the encoding.

Images