Methods, apparatus, equipment and storage media for testing ceramic layers in thermal barrier coatings

By acquiring terahertz time-domain echoes at multiple physical scanning locations and performing feature mapping and position encoding fusion, a trained model is used to detect ceramic layer thickness, solving the problem of insufficient detection accuracy in existing technologies and achieving more stable and accurate ceramic layer thickness detection.

CN123112005APending Publication Date: 2026-11-17UNIV OF SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611636882.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for detecting thermal barrier coatings struggle to fully utilize relevant information to improve the accuracy of ceramic layer detection when using terahertz time-domain echoes.

Method used

By acquiring multiple terahertz time-domain echoes of the thermal barrier coating under test at multiple different physical scanning positions, feature mapping and position encoding are performed, the echo feature sequences are fused, and the ceramic layer thickness is detected using a trained model.

Benefits of technology

It improves the stability and accuracy of ceramic layer detection results, reduces confusion caused by lack of location characterization, and enhances the integrity of the detection area.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123112005A_ABST
    Figure CN123112005A_ABST
Patent Text Reader

Abstract

This application relates to a method, apparatus, device, and storage medium for detecting ceramic layers in thermal barrier coatings. The method includes: acquiring multiple terahertz time-domain echoes from the thermal barrier coating under test; each terahertz time-domain echo being the time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions; performing feature mapping on each terahertz time-domain echo to obtain echo embedding vectors corresponding to each physical scanning position; obtaining position encoding vectors corresponding to each physical scanning position based on their spatial physical location; fusing the echo embedding vectors and position encoding vectors corresponding to each physical scanning position to obtain a fused echo vector corresponding to each physical scanning position, thereby obtaining an echo feature sequence of the thermal barrier coating under test; and determining the detection result of the ceramic layer in the thermal barrier coating under test based on the echo feature sequence. This method can improve the accuracy of the ceramic layer detection results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of terahertz nondestructive testing technology, and in particular to a method, apparatus, equipment and storage medium for testing ceramic layers in thermal barrier coatings. Background Technology

[0002] With the development of high-temperature equipment technologies such as aerospace and gas turbines, thermal barrier coatings (TBCs), as important thermal protection materials, are widely used in aero-engine turbine blades, combustion chambers, and other high-temperature service components. A thermal barrier coating typically consists of a ceramic top coat (TC), a bond coat (BC), and a metal substrate. The thickness of the ceramic top coat and the quality of the interfacial bonding directly determine the thermal insulation performance and structural stability of the thermal barrier coating. Therefore, accurate detection of the ceramic top coat thickness and / or interfacial defects is of significant engineering importance.

[0003] In related technologies, thermal barrier coating inspection methods mainly include ultrasonic testing, infrared testing, eddy current testing, and X-ray testing. Among them, ultrasonic testing usually requires a coupling agent, and its detection effect on complex structures is easily affected by surface condition and structural morphology; infrared testing has a wide range of applications, but its sensitivity to deep interface information is insufficient; eddy current testing is mainly suitable for conductive materials, and its detection capability for ceramic layers is limited; X-ray testing can intuitively display the internal structure, but it poses a risk of ionizing radiation and is not suitable for frequent use in the inspection of aerospace components.

[0004] Terahertz time-domain spectroscopy (THz-TDS) is a non-contact non-destructive testing technique with advantages such as high time resolution, low energy, and strong anti-interference capabilities. It can penetrate ceramic layers and acquire interface echo information, thus showing promising application prospects in the inspection of thermal barrier coatings. However, when using terahertz time-domain echoes for ceramic layer inspection, the accuracy of the results remains a challenge due to the dependence of the terahertz time-domain echo on factors such as the inspection location, the condition of the area under test, and the inspection conditions. How to more fully utilize the information related to ceramic layer inspection in the terahertz time-domain echoes to improve the accuracy of the inspection results still requires further improvement. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, equipment, and storage medium for detecting ceramic layers in thermal barrier coatings that can improve detection accuracy, in response to the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for detecting the ceramic layer in a thermal barrier coating, comprising: Multiple terahertz time-domain echoes are obtained from the thermal barrier coating under test; each terahertz time-domain echo is a time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. Feature mapping is performed on each of the terahertz time-domain echoes to obtain the echo embedding vectors corresponding to each of the physical scan positions; Based on the spatial physical location of each physical scanning location, obtain the location encoding vector corresponding to each physical scanning location; For each physical scanning position, the echo embedding vector and the position encoding vector corresponding to the physical scanning position are fused to obtain the fused echo vector corresponding to the physical scanning position, so as to obtain the echo feature sequence of the thermal barrier coating under test. The detection result of the ceramic layer in the thermal barrier coating to be tested is determined based on the echo feature sequence.

[0007] In one embodiment, prior to acquiring multiple terahertz time-domain echoes of the thermal barrier coating under test, the method further includes: Multiple training samples and their corresponding real labels are obtained; the training samples include multiple terahertz time-domain echoes of the thermal barrier coating under a simulated working condition; each terahertz time-domain echo is a time-domain echo response obtained by performing terahertz simulation scanning on the corresponding thermal barrier coating at different simulation scanning positions; the real labels include at least the real thickness information of the ceramic layer in the corresponding thermal barrier coating. The training samples include multiple terahertz time-domain echoes, which are input into the embedding representation module in the detection model to obtain the sample echo embedding vectors at each of the simulated scanning positions; the detection model also includes a position encoding module and an inversion network. The spatial physical locations of the multiple simulated scanning positions are input into the position encoding module to obtain the sample position encoding vector corresponding to each simulated scanning position. For each simulated scanning position, the sample echo embedding vector and the sample position encoding vector corresponding to the simulated scanning position are fused to obtain the sample echo feature sequence corresponding to the training sample; The sample echo feature sequence is input into the inversion network to obtain the sample prediction result corresponding to the training sample; the sample prediction result includes at least the predicted thickness information of the ceramic layer in the thermal barrier coating of the corresponding sample. Based on the sample prediction results and the true labels, the detection model is updated by backpropagation to obtain the trained detection model.

[0008] In one embodiment, obtaining multiple training samples and the real labels corresponding to each of the training samples includes: Parametric geometric modeling of thermal barrier coatings is performed based on Bézier curves and / or Bézier surfaces to obtain multiple original thermal barrier coating geometric models. The control point positions of the Bézier curves and / or Bézier surfaces corresponding to the multiple original thermal barrier coating geometric models are adjusted to obtain multiple sample thermal barrier coating geometric models with different interface morphologies. Material parameters of multiple sample thermal barrier coating geometric models were set to obtain sample thermal barrier coatings under various simulation conditions. For each of the thermal barrier coatings of the sample, terahertz simulation scanning is performed at multiple simulation scanning positions to obtain the sample terahertz time-domain echo corresponding to each of the simulation scanning positions. Multiple terahertz time-domain echoes obtained under the same simulation condition are used as a training sample, and the actual thickness information of the ceramic layer in the thermal barrier coating of the corresponding sample is used as the real label of the training sample.

[0009] In one embodiment, the true thickness information includes the true local thickness corresponding to multiple control point locations; the predicted thickness information includes the predicted local thickness corresponding to multiple control point locations; and the backpropagation update of the detection model based on the sample prediction results and the true labels includes: Based on the thickness difference between the predicted local thickness and the actual local thickness corresponding to the multiple control point locations, the relative thickness loss is obtained; From the multiple control point locations, determine the control point locations where the actual local thickness is less than the preset thin layer thickness threshold, and use these as the thin layer control point locations. Based on the thickness difference between the predicted local thickness and the actual local thickness corresponding to the thin layer control point location, the thin layer penalty loss is obtained. The detection model is updated by backpropagation by combining the relative thickness loss and the thin layer penalty loss.

[0010] In one embodiment, the step of performing feature mapping on each of the terahertz time-domain echoes to obtain the echo embedding vector corresponding to each of the physical scan positions includes: For each terahertz time-domain echo, the terahertz time-domain echo is windowed to obtain echo data corresponding to multiple time sub-windows; Based on the echo energy corresponding to the echo data within each time sub-window, the embedding weight corresponding to each time sub-window is determined; wherein, the embedding weight is positively correlated with the echo energy; Vector mapping is performed on the echo data within each time sub-window to obtain the sub-window embedding vector corresponding to each time sub-window. Based on the embedding weights corresponding to each of the time sub-windows, the embedding vectors of each sub-window are weighted and fused to obtain the echo embedding vector corresponding to the terahertz time-domain echo.

[0011] In one embodiment, obtaining the location encoding vector corresponding to each of the physical scan locations based on their spatial physical locations includes: Obtain the scan step size and total scan length corresponding to multiple physical scan positions; the scan step size is used to characterize the actual physical distance between adjacent physical scan positions, and the total scan length is used to characterize the total distance covered by multiple physical scan positions along a preset scan direction. Based on the scanning step size and the total scanning length, a spatial frequency range is determined; the spatial frequency range is the range of values ​​for the spatial frequency, and the spatial frequency is used to characterize the scale of positional change along the preset scanning direction; Based on the spatial frequency range, the spatial physical position of each physical scanning position is encoded to obtain the physical guidance position code corresponding to each physical scanning position. Based on the physical guidance position code and position correction vector corresponding to each of the physical scanning positions, the position code vector corresponding to each of the physical scanning positions is obtained; the position correction vector is used to correct the position representation deviation of the corresponding physical guidance position code.

[0012] In one embodiment, determining the detection result of the ceramic layer in the thermal barrier coating under test based on the echo feature sequence includes: The echo feature sequence is subjected to self-attention processing to obtain a fine branch feature sequence; wherein, the fine branch feature sequence includes fine branch features corresponding to each of the physical scan positions, and the fine branch features are used to characterize the local echo detail features corresponding to the physical scan positions; The echo feature sequence is sequentially subjected to downsampling, self-attention processing, and sequence length recovery processing to obtain a coarse branch feature sequence; wherein, the coarse branch feature sequence includes coarse branch features corresponding to each of the physical scan positions, and the coarse branch features are used to characterize the spatial correlation features between the corresponding physical scan position and other physical scan positions; By combining the fine branch feature sequence and the coarse branch feature sequence, a first fusion weight and a second fusion weight corresponding to each of the physical scan positions are determined; wherein, the first fusion weight is used to characterize the contribution of the local echo detail features corresponding to the corresponding physical scan position in the fusion process, and the second fusion weight is used to characterize the contribution of the spatial correlation features corresponding to the corresponding physical scan position in the fusion process. Based on the first fusion weight and the second fusion weight corresponding to each of the physical scanning positions, the fine branch features and coarse branch features corresponding to each of the physical scanning positions are fused position by position to obtain a fused feature sequence. The fused feature sequence is sequentially subjected to feedforward processing, global pooling processing, and regression processing to obtain the detection results of the ceramic layer in the thermal barrier coating to be tested.

[0013] Secondly, this application also provides a device for detecting ceramic layers in thermal barrier coatings, comprising: The acquisition module is used to acquire multiple terahertz time-domain echoes of the thermal barrier coating under test; each terahertz time-domain echo is a time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. An embedding module is used to perform feature mapping on each of the terahertz time-domain echoes to obtain the echo embedding vectors corresponding to each of the physical scanning positions. The location module is used to obtain the location encoding vector corresponding to each of the physical scanning locations based on the spatial physical location of each of the physical scanning locations; The fusion module is used to fuse the echo embedding vector and the position encoding vector corresponding to each physical scanning position to obtain the echo feature sequence of the thermal barrier coating under test. The detection module is used to determine the detection result of the ceramic layer in the thermal barrier coating to be tested based on the echo feature sequence.

[0014] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for detecting ceramic layers in thermal barrier coatings provided in the first aspect of this application.

[0015] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for detecting ceramic layers in thermal barrier coatings provided in the first aspect of this application.

[0016] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for detecting ceramic layers in thermal barrier coatings provided in the first aspect of this application.

[0017] The aforementioned method, apparatus, equipment, and storage medium for detecting ceramic layers in thermal barrier coatings acquire multiple terahertz time-domain echoes of the thermal barrier coating under test at multiple different physical scanning positions. This allows the ceramic layer detection results to be determined not only based on echo information from a single position, but also through comprehensive analysis of echo information corresponding to multiple physical scanning positions. This improves the amount of information used for ceramic layer detection and the completeness of the detection area characterization, thereby enhancing the stability of the ceramic layer detection results. Furthermore, because the echo feature sequence integrates echo feature information from multiple positions and spatial physical location information, the features corresponding to each terahertz time-domain echo not only characterize the echo response at the corresponding physical scanning position but also retain the correspondence between the echo response and the actual scanning position. This reduces the possibility of echo features corresponding to different physical scanning positions being confused due to a lack of location characterization, and allows the detection process to combine the distribution of echo features at different physical scanning positions for judgment, thus improving the accuracy of the ceramic layer detection results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a diagram illustrating the application environment of a method for detecting ceramic layers in thermal barrier coatings in one embodiment. Figure 2 This is a flowchart illustrating a method for detecting the ceramic layer in a thermal barrier coating in one embodiment; Figure 3 This is a schematic diagram of the propagation of terahertz waves in a thermal barrier coating in one embodiment. Figure 4 This is a schematic diagram of the overall architecture of the detection model in one embodiment; Figure 5 This is a flowchart illustrating the training process of a detection model in one embodiment; Figure 6 This is a schematic diagram of the geometric modeling of the thermal barrier coating in one embodiment; Figure 7 This is a schematic diagram of an embedded representation module in one embodiment; Figure 8 This is a schematic diagram of the MSHA module in one embodiment; Figure 9 This is a structural block diagram of a ceramic layer detection device in a thermal barrier coating in one embodiment; Figure 10This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0022] Please see Figure 1 The ceramic layer detection method in thermal barrier coatings provided in this application embodiment can be applied to a detection platform including a terahertz transceiver 102, a scanning control mechanism 104, and a computer device 106. The scanning control mechanism 104 controls the terahertz transceiver 102 to emit terahertz waves to the thermal barrier coating under test at different physical scanning positions and receive the echoes. The computer device 106 processes the received multiple terahertz time-domain echoes to obtain the detection results of the ceramic layer.

[0023] The computer device 106 can be a terminal or a server. A terminal can be, but is not limited to, various personal computers, laptops, smartphones, and tablets. A server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0024] The thermal barrier coating under test can be applied to the surface of aero-engine turbine blades, combustion chamber components, or other high-temperature service components. The thermal barrier coating may include a ceramic layer, an adhesive layer, and a metal substrate. The ceramic layer may be located on the outer side of the thermal barrier coating to reduce the temperature experienced by the metal substrate. The adhesive layer may be located between the ceramic layer and the metal substrate to improve the bonding performance between them.

[0025] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting ceramic layers in thermal barrier coatings is provided, which can be applied to... Figure 1 The following steps, 201 to 205, are used as an example of computer equipment. Step 201: Obtain multiple terahertz time-domain echoes of the thermal barrier coating under test.

[0026] The terahertz time-domain echoes are the time-domain echo responses received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. The different physical scanning positions can be the detection positions formed after the terahertz transceiver 102 moves to different positions sequentially under the control of the scanning control mechanism 104.

[0027] Understandably, when a terahertz wave is incident on the thermal barrier coating under test, it can be reflected and propagate multiple times at the interfaces between the air and the ceramic layer, between the ceramic layer and the adhesive layer, between the ceramic layer and the metal substrate, and within the ceramic layer. After receiving the echo signals formed by the above reflections or multiple propagations, the terahertz transceiver 102 can obtain the terahertz time-domain echo at the corresponding physical scan position.

[0028] Please see Figure 3 , Figure 3 Image (a) shows a schematic diagram of terahertz wave propagation in a thermal barrier coating under a single-point measurement scenario. Figure 3 (b) shows a schematic diagram of terahertz wave propagation in a thermal barrier coating under a multi-position scanning scenario. In a single-point measurement scenario, if the thermal barrier coating under test has a curved or irregular interface, some interface echoes may not be received because the reflection direction deviates from the terahertz transceiver, resulting in signal loss. In a multi-position scanning scenario, the terahertz transceiver can move laterally along the area under test and transmit terahertz waves and receive terahertz time-domain echoes at multiple physical scanning positions. The beam coverage areas corresponding to adjacent physical scanning positions can partially overlap, thereby compensating for or reducing the echo information loss that may be caused by single-point measurement under curved conditions, and obtaining multiple terahertz time-domain echoes for ceramic layer detection.

[0029] For example, the scanning control mechanism 104 can control the terahertz transceiver 102 to move along the test area of ​​the thermal barrier coating under test, and transmit terahertz waves to the thermal barrier coating under test at multiple different physical scanning positions, and receive the terahertz time-domain echo at each physical scanning position. Optionally, the multiple physical scanning positions correspond one-to-one with multiple terahertz time-domain echoes. Optionally, the terahertz transceiver 102 can transmit a conical beam to the thermal barrier coating under test, and the coverage areas of the conical beams corresponding to two adjacent physical scanning positions can at least partially overlap.

[0030] Step 202: Perform feature mapping on each terahertz time-domain echo to obtain the echo embedding vector corresponding to each physical scan position.

[0031] Feature mapping refers to the process of converting terahertz time-domain echoes into vectorized feature representations. The echo embedding vector is used to characterize the echo features related to ceramic layer detection in the terahertz time-domain echoes at the corresponding physical scan position.

[0032] For example, for any terahertz time-domain echo corresponding to a physical scan position, a computer device can perform vectorization, feature extraction, dimensionality reduction, or network mapping on the terahertz time-domain echo to obtain the echo embedding vector corresponding to that physical scan position. After feature mapping, terahertz time-domain echoes corresponding to different physical scan positions can obtain multiple echo embedding vectors, each echo embedding vector corresponding to a different physical scan position.

[0033] Step 203: Based on the spatial physical location of each physical scan position, obtain the position encoding vector corresponding to each physical scan position.

[0034] The spatial physical location is used to characterize the position of the corresponding physical scan position in the actual detection space. For example, it can be the coordinates, displacement, position number, two-dimensional coordinates, or three-dimensional coordinates of the physical scan position in a preset scanning direction. The position encoding vector can refer to the vectorized position representation obtained based on the spatial physical location, used to characterize the position attributes of the corresponding physical scan position.

[0035] For example, the computer device can encode the spatial physical location of each physical scan location based on a preset location encoding rule to obtain a location encoding vector corresponding to each physical scan location. Alternatively, the computer device can query the location encoding vector corresponding to each spatial physical location based on a preset location mapping relationship. Or, the computer device can input each spatial physical location into a location encoding model or location encoding function to obtain a location encoding vector corresponding to each physical scan location.

[0036] Optionally, the dimension of the location encoding vector can be the same as the dimension of the echo embedding vector to facilitate subsequent vector fusion.

[0037] Step 204: For each physical scan position, fuse the echo embedding vector and the position encoding vector corresponding to the physical scan position to obtain the fused echo vector corresponding to the physical scan position, so as to obtain the echo feature sequence of the thermal barrier coating under test.

[0038] For example, for any physical scan position, the computer device can obtain the echo embedding vector and the position encoding vector corresponding to the physical scan position, and perform vector addition, vector concatenation, weighted fusion or mapping fusion on the echo embedding vector and the position encoding vector to obtain the fused echo vector corresponding to the physical scan position.

[0039] Optionally, after obtaining the fused echo vectors corresponding to multiple physical scan positions, the computer device 106 can obtain the echo feature sequence of the thermal barrier coating under test based on the multiple fused echo vectors. For example, the multiple fused echo vectors can be arranged according to a preset position order of the multiple physical scan positions to obtain the echo feature sequence. The preset position order can be determined based on the acquisition order of the physical scan positions, the spatial coordinate order, the position number order, or other order that can characterize the correspondence between the physical scan positions.

[0040] Step 205: Determine the detection result of the ceramic layer in the thermal barrier coating to be tested based on the echo feature sequence.

[0041] The detection results of the ceramic layer may include ceramic layer thickness information. Optionally, the detection results of the ceramic layer may also include at least one of the following: local thickness information, thickness distribution information, interface state information, or defect detection results.

[0042] For example, a computer device can input the echo feature sequence into a preset detection algorithm to obtain the detection result of the ceramic layer in the thermal barrier coating to be tested. The preset detection algorithm can be a regression algorithm, a classification algorithm, a machine learning algorithm, a deep learning algorithm, or a combination thereof.

[0043] For example, when the detection result includes ceramic layer thickness information, the computer device can perform regression processing based on the echo feature sequence to obtain the ceramic layer thickness information; when the detection result includes defect detection results, the computer device can perform classification processing based on the echo feature sequence to obtain the detection result of whether there are defects in the ceramic layer.

[0044] In the aforementioned method for detecting ceramic layers in thermal barrier coatings, multiple terahertz time-domain echoes of the thermal barrier coating under test at various physical scanning positions are acquired. This allows the ceramic layer detection results to be determined not only based on echo information from a single position, but also through comprehensive analysis of echo information corresponding to multiple physical scanning positions. This improves the amount of information used for ceramic layer detection and the completeness of the detection area characterization, thereby enhancing the stability of the ceramic layer detection results. Furthermore, since the echo feature sequence integrates echo feature information from multiple positions and spatial physical location information, the features corresponding to each terahertz time-domain echo not only characterize the echo response at the corresponding physical scanning position but also retain the correspondence between the echo response and the actual scanning position. This reduces the possibility of echo features corresponding to different physical scanning positions being confused due to a lack of location characterization, and allows the detection process to combine the distribution of echo features at different physical scanning positions for judgment, thus improving the accuracy of the ceramic layer detection results.

[0045] In some embodiments, Figure 2The feature mapping, position encoding, feature fusion, and detection result determination processes in the illustrated embodiment can be implemented by a pre-built detection model. This detection model can be used to output the detection results of the ceramic layer in the thermal barrier coating under test based on the terahertz time-domain echoes at multiple physical scanning positions.

[0046] For example, the detection model can be a regression model, a classification model, a machine learning model, a deep learning model, or a combination thereof. For instance, the detection model may include at least one of a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), a Long Short-Term Memory (LSTM) network, a Transformer network, or a Multilayer Perceptron (MLP).

[0047] In some cases, terahertz time-domain echoes exhibit long time-series characteristics, and spatial correlations may exist between echoes corresponding to multiple physical scan locations. Using models such as convolutional neural networks and recurrent neural networks to process terahertz time-domain echoes may result in insufficient capture of the global dependencies of long-sequence signals. Therefore, in one possible implementation, the detection model can employ a Transformer-based inversion network to model the feature sequences of echoes from multiple locations through a self-attention mechanism, thereby improving the ability to utilize long-range correlation information in terahertz time-domain echoes.

[0048] For example, please refer to Figure 4 The detection model can include an embedding representation module, a position encoding module, a fusion module, and an inversion network. The embedding representation module performs feature mapping on the terahertz time-domain echo to obtain an echo embedding vector; the position encoding module obtains a position encoding vector based on the spatial physical location of the physical scan position; and the inversion network determines the detection result of the ceramic layer based on the echo feature sequence obtained by fusing the echo embedding vector and the position encoding vector.

[0049] Implementing the detection model Figure 2 Before the ceramic layer detection process in the illustrated embodiment, the detection model can be trained in advance.

[0050] In some embodiments, the computer device 106 can use a trained detection model to perform feature mapping on each terahertz time-domain echo to obtain echo embedding vectors corresponding to each physical scan position; use the trained detection model to obtain position encoding vectors corresponding to each physical scan position based on their spatial physical location; use the trained detection model to fuse the echo embedding vector and position encoding vector corresponding to each physical scan position to obtain a fused echo vector corresponding to the physical scan position, thereby obtaining the echo feature sequence of the thermal barrier coating under test; and use the trained detection model to determine the detection result of the ceramic layer in the thermal barrier coating under test based on the echo feature sequence.

[0051] The detection model can be trained based on multiple training samples and their corresponding real labels. Each training sample includes multiple terahertz time-domain echoes obtained by multi-position terahertz simulation scanning of the corresponding thermal barrier coating under the same simulation condition. The interface morphology of each thermal barrier coating is obtained by adjusting the control point positions of the Bézier curve and / or Bézier surface. The real labels include the actual local thickness of the ceramic layer in the thermal barrier coating of the corresponding sample at multiple control point positions. The detection results include the predicted local thickness of the ceramic layer in the thermal barrier coating under test at multiple control point positions.

[0052] In one possible implementation, multiple sample thermal barrier coatings with different interface morphologies can be generated by adjusting the positions of control points on the Bézier curve and / or Bézier surface. For example, the positions of the endpoint control points of the Bézier curve can be fixed, while the positions of the intermediate control points can be randomly or parametrically adjusted to generate sample thermal barrier coatings with different curvatures, degrees of undulation, or interface morphologies. For sample thermal barrier coatings with different interface morphologies, the true local thicknesses corresponding to the positions of multiple Bézier control points can be determined separately, and these true local thicknesses can be used as the true labels for the corresponding training samples. This allows the detection model to learn the mapping relationship between multi-position terahertz time-domain echoes and multiple local thickness parameters during training.

[0053] When using the trained detection model to detect the thermal barrier coating under test, the model can obtain the predicted local thickness of the ceramic layer at multiple control points based on the terahertz time-domain echoes of the thermal barrier coating at multiple physical scanning positions. The multiple predicted local thicknesses in the detection results can correspond to the multiple true local thicknesses in the training phase's true labels according to their respective positional definitions. This allows the detection model to jointly predict multiple local thickness parameters used to characterize the ceramic layer interface morphology based on the echo information from multiple physical scanning positions, thereby obtaining the thickness detection result of the ceramic layer in the thermal barrier coating under test.

[0054] The training process of the detection model is explained below.

[0055] In one exemplary embodiment, such as Figure 5 As shown, steps 501 to 506 precede step 201. Wherein: Step 501: Obtain multiple training samples and their corresponding real labels.

[0056] One training sample may include multiple terahertz time-domain echoes of the thermal barrier coating under a simulated working condition. Each terahertz time-domain echo can be the time-domain echo response obtained by performing terahertz simulation scanning on the corresponding thermal barrier coating at different simulation scanning positions.

[0057] The authenticity label includes at least the true thickness information of the ceramic layer in the thermal barrier coating of the corresponding sample.

[0058] The thermal barrier coating sample in this application embodiment can refer to a thermal barrier coating model or thermal barrier coating sample used for simulation to generate training samples. The simulation conditions may include at least one of the following: the geometric structure, material parameters, scanning parameters, incident wave parameters, and noise parameters of the sample thermal barrier coating. The simulation scanning position can be the scanning position where the terahertz wave acts on the sample thermal barrier coating during the simulation, which is related to... Figure 2 The physical scanning positions in the embodiments correspond to each other.

[0059] In one possible implementation, the computer device 106 can simulate the interface roughness, waviness, and non-uniformity of the thermal barrier coating ceramic layer through Bézier parametric modeling, thereby generating multiple training samples.

[0060] Specifically, computer device 106 performs parametric geometric modeling of the thermal barrier coating based on the Bessel parametric geometric model to obtain multiple original thermal barrier coating geometric models; adjusts the positions of the Bessel control points corresponding to the multiple original thermal barrier coating geometric models to obtain multiple sample thermal barrier coating geometric models with different interface morphologies; sets the material parameters of the multiple sample thermal barrier coating geometric models to obtain sample thermal barrier coatings under various simulation conditions; for each sample thermal barrier coating, performs terahertz simulation scanning at multiple simulation scanning positions to obtain sample terahertz time-domain echoes corresponding to each simulation scanning position; uses multiple sample terahertz time-domain echoes obtained under the same simulation condition as a training sample, and uses the real thickness information of the ceramic layer in the corresponding sample thermal barrier coating as the real label corresponding to the training sample.

[0061] The Bezier parametric geometric model can include Bezier curve models and Bezier surface models. Bezier curve models can be used to characterize the interface profile of the thermal barrier coating on a two-dimensional cross-section, while Bezier surface models can be used to characterize the interface morphology of the thermal barrier coating in three-dimensional space.

[0062] For example, to describe curved surfaces and irregular interfaces, the upper surface of a sample thermal barrier coating can be parameterized using cubic Bézier curves to obtain multiple original thermal barrier coating geometric models. A cubic Bézier curve can be represented as: ; In the formula, Representing curve parameters The corresponding interface outline position; Represents curve parameters, and ; Indicates the Bessel control point index; Indicates the first Position vectors of the Bezier control points; Indicates the first The cubic Bessel basis functions corresponding to each Bessel control point; This represents the combination coefficients corresponding to the cubic Bessel basis functions.

[0063] The computer device 106 can fix the positions of the endpoint control points in the Bézier control points and randomly or parametrically adjust the positions of the intermediate control points to obtain sample thermal barrier coating geometric models with different curvatures, different degrees of undulation, or different interface morphologies. It should be noted that this embodiment does not require pre-setting specific curvature values; the curvature can be formed by changes in the control point positions. Constructing a large number of different interface morphologies through changes in the control point positions can enhance the diversity of training samples.

[0064] Optionally, the computer device 106 can use the de Casteljaualgorithm to determine the interface contour position corresponding to the cubic Bézier curve based on multiple Bézier control points, thereby improving the numerical stability and geometric continuity of the interface contour generation process. The aforementioned cubic Bézier curve model can also be extended to a Bézier surface model to model complex three-dimensional interface morphologies.

[0065] Please see Figure 6 , Figure 6 (a) is a schematic diagram of generating the thermal barrier coating interface profile based on Bezier control points, where the horizontal axis... Indicates the lateral position along the area to be measured, and the ordinate is... Indicates the thickness direction or height direction. , , and The red curve represents the interface profile defined by the cubic Bézier curve, corresponding to four Bézier control points. Indicates the location of the reference baseline or flat base; Figure 6 (b) in the diagram illustrates the process of generating Bézier curves using the de Castelli recursive interpolation algorithm. Indicates curve parameters, In the de Castellio recursive interpolation algorithm, the first... The interpolation yields the th... There are interpolation points, among which Corresponding to the original Bezier control points, Corresponding curve parameters The generated interface outline position The red curve represents the final generated Bézier curve.

[0066] The computer device 106 can set the material parameters of multiple sample thermal barrier coating geometric models to obtain sample thermal barrier coatings under various simulation conditions.

[0067] Optionally, the material parameters may include the optical parameters corresponding to the ceramic layer, the adhesive layer, and the metal substrate, respectively. The optical parameters may include at least one of complex refractive index, absorption coefficient, complex permittivity, and conductivity.

[0068] Optionally, air, ceramic layer, and adhesive layer or metal substrate can be used as different media, and the complex permittivity and permeability of each medium can be set respectively.

[0069] Specifically, for each sample thermal barrier coating, the computer device 106 can perform terahertz simulation scanning at multiple simulation scanning positions to obtain the sample terahertz time-domain echo corresponding to each simulation scanning position.

[0070] Optionally, the computer device 106 can use a full-wave electromagnetic simulator to simulate a terahertz transceiver, moving laterally along the test area of ​​the sample thermal barrier coating, and... Each simulated scanning position sequentially emits a conical beam and receives the echo, wherein... This indicates the total number of simulated scan positions. The coverage area of ​​the conical beam can move with the translation of the terahertz transceiver, and the conical coverage areas corresponding to adjacent simulated scan positions can partially overlap to compensate for echo loss caused by single-point reception under curved surface conditions.

[0071] For example, the The cone-shaped coverage area corresponding to each simulated scan position can be represented as: ; In the formula, Indicates the first The cone-shaped coverage area corresponding to each simulated scan position; Indicates the index of the simulation scan position, and ; Represents the spatial point location vector within the cone-shaped coverage area; Indicates the first The terahertz wave incident center position vector corresponding to each simulated scanning position; Indicates the first Local normal vectors corresponding to each simulated scan position; This indicates the effective depth of the conical beam along the local normal direction; Indicates the half-cone angle of a cone-shaped beam; " represents the vector dot product;" "" represents the Euclidean norm of a vector.

[0072] At each simulated scanning location, after the terahertz wave is incident on the thermal barrier coating of the sample, it can be reflected and propagate multiple times at the interfaces between the air and the ceramic layer, the ceramic layer and the adhesive layer, the ceramic layer and the metal substrate, and within the ceramic layer. The computer device 106 can use a full-wave electromagnetic simulator to obtain the received spectrum at the corresponding simulated scanning location, and perform inverse Fourier transform and preprocessing on the received spectrum to obtain the sample terahertz time-domain echo at the corresponding simulated scanning location.

[0073] For example, for the first At each simulated scanning position, under the approximation of local normal incidence, let the medium... The complex permittivity is The propagation constant is The local ceramic layer thickness is Then the medium The propagation constant can be expressed as: ; In the formula, Indicates the media index; Indicates angular frequency; Represents the imaginary unit; Indicates medium Magnetic permeability; Indicates medium At angular frequency The complex permittivity of the following; Indicates medium At angular frequency The propagation constant under the given conditions.

[0074] In one example, medium 0, medium 1, and medium 2 can be air, a ceramic layer, and an adhesive layer or a metal substrate, respectively. At each simulated scanning location, the round-trip propagation factor of the terahertz wave within the local ceramic layer can be expressed as: ; In the formula, Indicates the first The local ceramic layer round-trip propagation factor corresponding to each simulated scan position; Indicates the ceramic layer at angular frequency The propagation constant under these conditions; Indicates the first The local ceramic layer thickness corresponding to each simulated scanning location; This represents an exponential function with the natural constant as its base. Furthermore, The real and imaginary parts can be used to characterize the attenuation and phase delay of terahertz waves during propagation within the ceramic layer, respectively.

[0075] Under the local normal incidence approximation, let and Indicates media index, Indicates medium to medium Fresnel reflection coefficient of the interface Indicates medium to medium Fresnel transmittance coefficient of the interface.

[0076] After superimposing surface echoes, main internal interface echoes, and continuous intralayer echoes, the first The received spectrum corresponding to each simulated scan position can be represented as: ; In the formula, Indicates the first The received spectrum corresponding to each simulated scan position; Indicates the transmission spectrum; Represents the natural constant; Indicates the angular frequency of air The propagation constant under these conditions; Indicates the first The air path distance corresponding to each simulated scan position; This represents the Fresnel reflection coefficient at the air-to-ceramic layer interface; This represents the Fresnel transmission coefficient at the air-to-ceramic layer interface; This represents the Fresnel transmittance coefficient at the ceramic layer-to-air interface; The Fresnel reflection coefficient represents the interface between the ceramic layer and the adhesive layer or metal substrate. This represents the Fresnel reflection coefficient at the ceramic layer-to-air interface; Indicates the first The measurement noise or numerical noise corresponding to each simulated scan position. For example, It can include random white noise.

[0077] It should be noted that the above received spectrum expression can be used as an example for simulation modeling. When using a full-wave electromagnetic simulator for simulation, the received spectrum or received time-domain signal can also be directly output by the full-wave electromagnetic simulator, and this application does not limit this.

[0078] Furthermore, computer device 106 can [address the issue of] the [other device]. The received spectrum corresponding to each simulated scan position Perform an inverse Fourier transform and preprocess the signal after the inverse Fourier transform to obtain the sample terahertz time-domain echo corresponding to the simulated scanning position. The preprocessing may include at least one of denoising, normalization, baseline correction, time alignment, and truncation of the effective time period.

[0079] In one example, computer device 106 can use multiple sample terahertz time-domain echoes obtained under the same simulation condition as a training sample, and use the actual thickness information of the ceramic layer in the thermal barrier coating of the corresponding sample as the real label of the training sample. Exemplarily, multiple sample terahertz time-domain echoes in the same training sample can form a terahertz time-domain echo matrix according to the arrangement order of the simulation scan positions. One dimension of this terahertz time-domain echo matrix corresponds to the simulation scan position, and the other dimension corresponds to the time sampling point. This training sample can be represented as: ; In the formula, The training samples are represented using a terahertz time-domain echo matrix. Indicates the first The simulated scanning position is at the [number]th [position]. Echo amplitude values ​​at each time sampling point; Indicates the first The time corresponding to each time sampling point; This represents the index of the time sampling point, and ; This represents the total number of time sampling points included in each sample terahertz time-domain echo; The dimension is The space of real matrix numbers.

[0080] In one example, the true label may include the relative permittivity of the ceramic layer and the true local thickness corresponding to the locations of multiple Bessel control points. When four Bessel control points are used to characterize the interface morphology of the ceramic layer, the true label can be represented as: ; in, This represents the true label corresponding to the training sample; Indicates the relative permittivity of the ceramic layer; , , and These represent the true local thicknesses corresponding to the four Bessel control point locations. This training sample can be directly used for subsequent joint regression with the embedding representation and Transformer inversion network to jointly estimate the relative permittivity and multiple local thickness parameters, preserving the complete temporal information and lateral correlation of each scan location, without setting intermediate imaging or peak calculation steps.

[0081] It should be noted that the true label is not limited to the above forms. When the number of control point locations is not four, the true label can include the true local thickness corresponding to other control point locations; when the ceramic layer detection result only includes thickness information, the true label can include the true local thickness corresponding to multiple control point locations; when the ceramic layer detection result also includes material parameter detection results, the true label can also include the relative permittivity of the ceramic layer or other material parameters.

[0082] Step 502: Input the multiple terahertz time-domain echoes from the training samples into the embedding representation module in the detection model to obtain the sample echo embedding vectors at each simulated scanning position.

[0083] The embedding representation module is used to perform feature mapping on the sample terahertz time-domain echoes, converting each sample terahertz time-domain echo into a corresponding sample echo embedding vector. The sample echo embedding vector is used to characterize the echo features related to ceramic layer detection in the sample terahertz time-domain echoes at the corresponding simulated scan position.

[0084] In one possible implementation, the embedding representation module can use Energy-Adaptive Window Embedding (EAWE) to perform feature mapping on each sample terahertz time-domain echo.

[0085] Specifically, the computer device 106 can use an embedded representation module to perform windowing processing on each sample terahertz time-domain echo to obtain sample echo data corresponding to multiple time sub-windows; based on the sample echo energy corresponding to the sample echo data in each time sub-window, determine the sample embedding weight corresponding to each time sub-window; wherein the sample embedding weight is positively correlated with the sample echo energy; perform vector mapping on the sample echo data in each time sub-window to obtain the sample sub-window embedding vector corresponding to each time sub-window; and perform weighted fusion of the sample sub-window embedding vectors based on the sample embedding weights corresponding to each time window to obtain the sample echo embedding vector corresponding to the sample terahertz time-domain echo.

[0086] For example, by embedding a representation module, each sample terahertz time-domain echo is divided into nearly uniform partitions. The sample echo data corresponding to each time window.

[0087] Optionally, please refer to Figure 7 The computer device 106 can detect the main echo peaks on the sample terahertz time-domain echoes by embedding a representation module, and adjust the window boundaries of a certain time sub-window when the main echo peaks are located at the window boundary of a certain time sub-window, so as to obtain an updated representation. The sample echo data corresponding to each time sub-window is used to avoid the main echo peak being truncated by the window boundary.

[0088] Regarding the first The first simulation scan position corresponds to the first The first sample terahertz time-domain echo The embedded representation module can calculate the root mean square energy of the sample echo data within a given time window. The sample echo energy corresponding to each time sub-window can be expressed as: , ; In the formula, Indicates the first The first sample terahertz time-domain echo The sample echo energy corresponding to each time sub-window; Indicates the first The first sample terahertz time-domain echo A time window; Indicates the first The number of time sampling points within each time sub-window; This indicates the total number of time-based sub-windows.

[0089] Computer device 106 can be embedded in a representation module, in conjunction with the first The corresponding terahertz time-domain echo of the sample The echo energy of each sample is used to form a sample energy vector, which is represented as: ; In the formula, Indicates the first The sample energy vector corresponding to the terahertz time-domain echo of a sample; to They represent the first The first to the second The sample echo energy corresponding to each time sub-window; This indicates transpose.

[0090] The embedding representation module may include a lightweight gating network. The computer device 106 can input the sample energy vector into the lightweight gating network to generate sample embedding weights corresponding to multiple time sub-windows. The sample embedding weights can be represented using normalized weights as follows: ; In the formula, Indicates the first The sample embedding weight vector corresponding to the sample terahertz time-domain echo; This indicates a lightweight gating network; Indicates the first The sample embedding weights corresponding to each time sub-window.

[0091] Optionally, the sample embedding weights output by the lightweight gating network are positively correlated with the sample echo energy. That is, a time window containing the main echo information and with higher energy can obtain a larger sample embedding weight, while a time window mainly composed of noise or invalid signals and with lower energy can obtain a smaller sample embedding weight.

[0092] The computer device 106 can perform vector mapping on the sample echo data within each time sub-window through the embedding representation module to obtain the sample sub-window embedding vector corresponding to each time sub-window. For example, each time sub-window can be vector-mapped using an independent linear projection matrix and a bias. Based on the sample embedding weights corresponding to each time sub-window, the embedding vectors of each sample sub-window are weighted and fused, and then subjected to layer normalization (LayerNorm) processing. The resulting sample echo embedding vector can be expressed as: ; In the formula, Indicates the first The sample echo embedding vector corresponding to each simulated scan position; Presentation layer normalization processing; Indicates the first The independent linear projection matrices corresponding to each time sub-window; Indicates the first The first sample terahertz time-domain echo Sample echo data vector within a time sub-window; Indicates the first The offset corresponding to each time sub-window; Indicates the first The sample sub-window embedding vector corresponding to each time sub-window.

[0093] During training, the embedding representation module can have its parameters updated as the detection model backpropagates. Specifically, the lightweight gating network parameters, the linear projection matrix corresponding to each time window, and the bias in the embedding representation module can be used as learnable parameters and updated based on the loss between the sample prediction results and the true labels. Optionally, if the layer normalization process includes learnable scaling parameters and learnable offset parameters, the parameters corresponding to the layer normalization process can also be updated during training.

[0094] Step 503: Input the spatial physical locations of multiple simulated scanning positions into the position encoding module to obtain the sample position encoding vector corresponding to each simulated scanning position.

[0095] The position encoding module is used to generate sample position encoding vectors based on the spatial physical location of each simulated scanning position. The sample position encoding vectors are used to characterize the position attributes of the corresponding simulated scanning positions.

[0096] In one possible implementation, the position encoding module can use Physically Informed Positional Encoding (PIPE) to obtain the sample position encoding vector corresponding to each simulation scan position.

[0097] Specifically, the computer device 106 can obtain the scanning step size and total scanning length corresponding to multiple simulated scanning positions through the position encoding module. The scanning step size represents the actual physical distance between adjacent simulated scanning positions, and the total scanning length represents the total distance covered by the multiple simulated scanning positions along a preset scanning direction. Optionally, the scanning step size can be the actual physical distance between two adjacent aperture sampling positions, and the total scanning length can be the physical length covered by the synthetic aperture scan.

[0098] Computer device 106 can determine the spatial frequency range based on the scan step size and total scan length through the position encoding module, and generate multiple discrete spatial frequencies within this range. The spatial frequency range is the range of values ​​for the spatial frequency, which is used to characterize the scale of position change along a preset scan direction. Based on the multiple discrete spatial frequencies within the spatial frequency range, the spatial physical position of each simulated scan position is encoded to obtain the physical guidance position code corresponding to each simulated scan position.

[0099] For example, regarding the first The simulated scanning position can be determined based on the first... The position index, scan step size, and discrete spatial frequency corresponding to each simulated scan position are used to construct a spatial frequency basis, represented as follows: ; ; In the formula, and They represent the first The first physical guidance position code corresponding to the simulated scan position Dimensional encoded components and the first 3D encoded components; Indicates the first The position index corresponding to each simulated scan position; An index representing the frequency in discrete space; Indicates the first A discrete spatial frequency; Indicates the scan step size.

[0100] The position encoding module can fuse the physical guidance position encoding with the position correction vector to obtain the sample position encoding vector corresponding to each simulation scan position. The position correction vector is used to correct the position representation deviation of the corresponding physical guidance position encoding.

[0101] For example, the position correction vector can be represented using a learnable absolute positional embedding (APE) residual vector. The sample position encoding vector corresponding to each simulated scan position can be represented as: ; In the formula, Indicates the first The sample position encoding vector corresponding to each simulated scan position; Indicates the first The physical guide position code corresponding to each simulated scan position; Indicates the first The learnable absolute position embedding residual vector corresponding to each simulated scan position; This indicates mixed weights.

[0102] Optionally, the hybrid weights can be preset hybrid weights or learnable hybrid weights updated during the training of the detection model. The learnable absolute position embedding residual vector can be used as model parameters updated during the training of the detection model.

[0103] In this embodiment, by determining the spatial frequency range based on the scanning step size and the total scanning length, and constructing a physical guidance position code based on the spatial physical location of each simulated scanning position, scanning geometric information and spatial frequency priors can be introduced into the sample position code vector. By fusing the physical guidance position code with the learnable absolute position embedding residual vector, the position representation deviation in the actual measurement or simulation process can be corrected while retaining the physical prior constraints, thereby providing a more accurate position representation for the subsequent fusion of the sample echo embedding vector and the sample position code vector.

[0104] Step 504: For each simulated scanning position, fuse the sample echo embedding vector and the sample position encoding vector corresponding to the simulated scanning position to obtain the sample echo feature sequence corresponding to the training sample.

[0105] Among them, the sample echo embedding vector is used to characterize the sample terahertz time-domain echo characteristics at the corresponding simulated scanning position, and the sample position encoding vector is used to characterize the spatial position attributes of the corresponding simulated scanning position.

[0106] For example, the detection model may also include a fusion module, for the first... At each simulated scanning location, the computer device 106 can use a fusion module to add the sample echo embedding vector and the sample location encoding vector corresponding to that simulated scanning location to obtain the first simulated scanning location. The sample fusion echo vector corresponds to each simulated scan position. Optionally, the computer device 106 can also obtain the sample fusion echo vector by splicing, weighted fusion, or mapping fusion.

[0107] After obtaining the sample fusion echo vectors corresponding to multiple simulated scanning positions, the computer device 106 can arrange the multiple sample fusion echo vectors according to the scanning order of the multiple simulated scanning positions to obtain the sample echo feature sequence corresponding to the training sample.

[0108] Step 505: Input the sample echo feature sequence into the inversion network to obtain the sample prediction results corresponding to the training samples.

[0109] The sample prediction results include at least the predicted thickness information of the ceramic layer in the corresponding thermal barrier coating. Optionally, the sample prediction results may also include the relative permittivity of the ceramic layer, the predicted local thickness corresponding to the locations of multiple Bessel control points, and defect probability information.

[0110] Please continue reading. Figure 4 The inversion network can include one or more enhanced Transformer blocks, global pooling layers, and a regression head. The enhanced Transformer block can include a multi-scale hierarchical self-attention (MSHA) module and a feed forward network (FFN), and the regression head can be a multilayer perceptron (MLP) regression head.

[0111] In one possible implementation, the computer device 106 can use the MSHA module to perform self-attention processing on the sample echo feature sequence to obtain a sample fine-branch feature sequence. The sample fine-branch feature sequence includes sample fine-branch features corresponding to each simulated scan position, and these features characterize the local echo detail features corresponding to the respective simulated scan position. For example, the MSHA module can perform full-resolution multi-head self-attention processing on the sample echo feature sequence within the fine branches to obtain the sample fine-branch feature sequence.

[0112] The computer device 106 can also use the MSHA module to sequentially perform downsampling, self-attention, and sequence length recovery processing on the sample echo feature sequence to obtain a coarse branch feature sequence. The coarse branch feature sequence includes coarse branch features corresponding to each simulated scan position, which characterize the spatial correlation between the corresponding simulated scan position and other simulated scan positions. For example, the MSHA module can first perform average pooling downsampling on the sample echo feature sequence in the coarse branch, then perform self-attention processing, and finally restore the sequence length to match the sample echo feature sequence through interpolation recovery processing to obtain the coarse branch feature sequence.

[0113] In one example, let the sample fine-branch feature sequence be... The sample coarse branch feature sequence is The computer device 106 can calculate the fusion weight based on the sample fine-branch feature sequence and the sample coarse-branch feature sequence through the location-related dynamic gating network in the MSHA module. The fusion weight can be expressed as: ; In the formula, This represents the fusion weights output by the dynamic gating network; and This represents the learnable weight parameters in a dynamic gating network; This represents the normalized exponential function; This represents the activation function of the Gaussian Error Linear Unit (GELU). This means concatenating the sample fine-branch feature sequences and the sample coarse-branch feature sequences along the feature dimension.

[0114] Furthermore, the fusion weights It may include the first sample fusion weights Second Sample Fusion Weight Among them, the first sample fusion weight The second sample fusion weight is used to characterize the contribution of local echo detail features in the fusion process. Used to characterize the contribution of spatial correlation features in the fusion process. Computer device 106 can use the MSHA module, based on the fusion weights of the first sample. Second Sample Fusion Weight For the sample fine branch feature sequence and sample coarse branch feature sequence Perform position-by-position weighted fusion to obtain the fused feature sequence: ; In the formula, This represents the fused feature sequence output by the MSHA module; This indicates position-by-position weighting.

[0115] Understandably, when local echo details at a certain simulated scan location are more important for thickness inversion, that simulated scan location can correspond to a larger first-sample fusion weight. This enhances the role of fine-branch information; when a certain simulated scan position needs to be combined with the overall spatial correlation information of surrounding simulated scan positions, that simulated scan position can correspond to a larger second sample fusion weight. This is to enhance the role of coarse branch information.

[0116] Optionally, please refer to Figure 8 The MSHA module may also include a residual connection, which is used to add the echo feature sequence of the input MSHA module to the fused feature sequence, and redetermine the result of the addition as the fused feature sequence to obtain the final output of the MSHA module, thereby preserving the input features and improving the stability of feature transfer.

[0117] After obtaining the fused feature sequence output by the MSHA module, the computer device 106 can perform feedforward processing on the fused feature sequence through the feedforward network in the enhanced Transformer block to obtain the feature sequence output by the enhanced Transformer block. When the inversion network includes multiple enhanced Transformer blocks, the computer device 106 can input the feature sequence output by the previous enhanced Transformer block into the next enhanced Transformer block to update the feature sequence layer by layer, and use the feature sequence output by the last enhanced Transformer block as the final sample feature sequence.

[0118] Furthermore, the computer device 106 can perform global pooling on the final sample feature sequence and input the globally pooled features into a regression head to obtain the sample prediction results corresponding to the training samples. For example, the computer device 106 can perform global average pooling on the final sample feature sequence and output the predicted thickness information of the ceramic layer through a two-layer MLP regression head.

[0119] In this embodiment, the MSHA module performs fine-branch and coarse-branch processing on the sample echo feature sequence, which can obtain the spatial correlation features between different simulated scanning positions while preserving the local echo detail features of each simulated scanning position; the location-dependent dynamic gating network determines the fusion weights corresponding to each simulated scanning position, which can adaptively balance the local echo details and spatial correlation information according to the feature situation of different simulated scanning positions.

[0120] Step 506: Based on the sample prediction results and the true labels, backpropagation is performed to update the detection model, resulting in the trained detection model.

[0121] The sample prediction result can include the predicted thickness information of the ceramic layer in the thermal barrier coating of the sample, and the true label can include the true thickness information of the ceramic layer in the thermal barrier coating of the sample. The computer device 106 can determine the training loss of the detection model based on the difference between the predicted thickness information and the true thickness information, and update the detection model through backpropagation based on the training loss.

[0122] In one possible implementation, the true thickness information includes the true local thickness corresponding to multiple Bézier control point positions, and the predicted thickness information includes the predicted local thickness corresponding to multiple Bézier control point positions. The computer device 106 can obtain a relative thickness loss based on the thickness difference between the predicted local thickness and the true local thickness corresponding to the multiple control point positions. From the multiple control point positions, control point positions where the true local thickness is less than a preset thin layer thickness threshold are determined as thin layer control point positions. The computer device 106 can obtain a thin layer penalty loss based on the thickness difference between the predicted local thickness and the true local thickness corresponding to the thin layer control point positions. Combining the relative thickness loss and the thin layer penalty loss, a training loss is determined, and the detection model is updated through backpropagation based on the training loss to obtain the trained detection model.

[0123] Optionally, the computer device 106 can update the embedded representation module, the position encoding module, the fusion module, and the learnable parameters in the inversion network of the detection model.

[0124] In one example, the training loss can be constructed using relative error and a thin-layer penalty term, and this training loss can be expressed as: ; In the formula, Indicates training loss; This indicates the total number of control point locations; The index representing the location of the control point; Indicates the first Predicted local thickness corresponding to each control point location; Indicates the first The actual local thickness corresponding to each control point location; Indicates the thin-layer weight; This indicates a preset thin-layer thickness threshold; This indicates the location of the control point where the actual local thickness is less than the preset thin layer thickness threshold.

[0125] In the above formula, the first term can be used as a way to implement the relative thickness loss, which is used to characterize the relative thickness error corresponding to multiple control point positions; the second term can be used as a way to implement the thin layer penalty loss, which is used to increase the proportion of thin layer control point positions with a real local thickness less than a preset thin layer thickness threshold in the training loss.

[0126] In this embodiment, by obtaining the relative thickness loss based on the thickness difference between the predicted local thickness and the actual local thickness corresponding to multiple control point locations, the detection model can learn the overall thickness distribution of the ceramic layer in the thermal barrier coating of the sample. By determining the control point locations where the actual local thickness is less than a preset thin layer thickness threshold as thin layer control point locations, and obtaining the thin layer penalty loss based on the thickness difference corresponding to the thin layer control point locations, the detection model's attention to the thickness error of the thin layer region can be improved. By combining the relative thickness loss and the thin layer penalty loss to perform backpropagation update on the detection model, it is beneficial to improve the detection accuracy of the trained detection model for the ceramic layer thickness, especially the thickness of the thin layer region.

[0127] In this embodiment, realistic geometric samples with controllable curvature, local undulations, and multi-layer interfaces are generated using Bézier curves or surfaces. These samples serve as the geometric basis for simulation and training data, enabling the acquisition of diverse and comprehensive training sets without requiring a large number of measured samples, significantly reducing sample preparation costs and time. Multi-position conical beam scanning is employed to partially overlap the beam coverage areas of adjacent scanning positions while preserving laterally distributed echo information. The MSHA module is used to fuse full-resolution and downsampling branches in parallel, and position-dependent dynamic gating adaptively balances local echo details and global spatial context. This allows for simultaneous modeling of short-term echo structure and long-range scan correlation, enabling the model to capture both local echo details and aggregate long-range scan context. This mitigates errors caused by single-point signal loss, multi-interface echo overlap, and image-based interface localization under curved surface conditions. Physical information position encoding injects scanning geometry and spatial frequency priors into the network. The network, combined with learnable residuals, achieves a balance between prior guidance and data adaptation. This allows the model to be constrained by physical priors in the early stages of training, while adaptively correcting measurement biases later, accelerating convergence and improving adaptability to actual measurement deviations. The introduction of energy-adaptive window embedding and thin-layer penalty loss design enables the model to achieve higher sensitivity and smaller relative errors in thin-layer (weak echo) regions, thus improving the performance shortcomings of traditional time-of-flight methods and simple regression models in thin-layer scenarios. Through downsampling and interpolation branches in the MSHA module, and adjustable layer and attention head counts, a flexible trade-off between accuracy and computational resources can be achieved, facilitating deployment on laboratory servers or embedded or edge devices in engineering sites. It is applicable to the detection of thermal barrier coatings with different curvatures, material parameters, and noise levels. It can be used for quantitative thickness inversion and parallel output of defect probability maps, possessing direct engineering application value in the non-destructive testing and life assessment of turbine blade coatings for aero-engines.

[0128] In some embodiments, after obtaining the trained detection model through the above training process, the computer device 106 can use the trained detection model to detect the thermal barrier coating under test. The trained detection model may include a trained embedding representation module, a trained position encoding module, a fusion module, and a trained inversion network. It should be noted that during the detection phase, the model parameters in the trained detection model can remain fixed, and the computer device 106 can obtain the detection results of the ceramic layer in the thermal barrier coating under test based on multiple terahertz time-domain echoes and the spatial physical locations of multiple physical scan positions.

[0129] Specifically, the computer device 106 can first acquire multiple terahertz time-domain echoes of the thermal barrier coating under test. Each terahertz time-domain echo is the time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. Multiple physical scanning positions can be set sequentially along a preset scanning direction, and multiple terahertz time-domain echoes can be arranged according to the scanning order of the corresponding physical scanning positions.

[0130] In some embodiments, the computer device 106 may employ a trained embedding representation module to perform feature mapping on each terahertz time-domain echo to obtain echo embedding vectors corresponding to each physical scan position. Specifically, for each terahertz time-domain echo, the trained embedding representation module may perform windowing processing on the terahertz time-domain echo to obtain echo data corresponding to multiple time sub-windows; based on the echo energy corresponding to the echo data within each time sub-window, determine the embedding weights corresponding to each time sub-window; perform vector mapping on the echo data within each time sub-window to obtain sub-window embedding vectors corresponding to each time sub-window; and perform weighted fusion of the sub-window embedding vectors based on the embedding weights corresponding to each time sub-window to obtain the echo embedding vector corresponding to the terahertz time-domain echo.

[0131] In one possible implementation, the embedding representation module can also adjust the embedding weights corresponding to each time sub-window based on the echo correlation information between different scanning positions, so as to reduce the suppression of effective weak echoes due to low echo energy. For each terahertz time-domain echo, the computer device 106 can determine the initial embedding weights based on the echo energy of each time sub-window, and compensate the initial embedding weights corresponding to low-energy time sub-windows based on the echo correlation between time sub-windows corresponding to adjacent physical scanning positions, to obtain the compensated embedding weights of low-energy time sub-windows, thereby determining the target embedding weights corresponding to each time sub-window. Based on the target embedding weights, the embedding vectors of each time sub-window are weighted and fused to obtain the echo embedding vector corresponding to the terahertz time-domain echo.

[0132] For example, computer device 106 can use an embedded representation module to time-align terahertz time-domain echoes corresponding to adjacent physical scan positions and calculate the similarity between echo data within the corresponding time sub-window after time alignment, thereby obtaining the echo correlation degree of the corresponding time sub-window. The time alignment can be performed based on the position of the main echo peak in each terahertz time-domain echo to reduce the impact of echo time offset between different physical scan positions on the echo correlation degree calculation.

[0133] The computer device 106 can define time windows with echo energy less than a preset energy threshold as low-energy time windows. If the echo correlation degree corresponding to a low-energy time window is greater than a preset correlation threshold, it determines a weight compensation amount based on the echo correlation degree to compensate for the corresponding initial embedding weights. For example, the sum of the initial embedding weight and the weight compensation amount can be used as the adjusted weight of the corresponding time window. The adjusted weight of the low-energy time window and the initial embedding weights of other time windows are then normalized to obtain the target embedding weights for each time window. For time windows that do not meet the weight compensation conditions, their initial embedding weights can remain unchanged and participate in the above normalization process.

[0134] In this embodiment, the initial embedding weight is positively correlated with the echo energy of the corresponding time sub-window, while the final embedding weight takes into account both the echo energy of the corresponding time sub-window and the degree of echo correlation between different physical scan positions. For time sub-windows with lower energy but higher echo correlation between adjacent physical scan positions, their contribution to the final embedding representation can be increased through weight compensation; for time sub-windows with lower energy and lower echo correlation, a lower embedding weight can be maintained to reduce the impact of noise or invalid signals on the embedding representation.

[0135] In another possible implementation, the detection model can be trained based on the true local thickness corresponding to multiple control point positions. The training loss includes a thin-layer penalty loss, which is determined based on the thickness prediction error corresponding to the control point position where the true local thickness is less than a preset thin-layer thickness threshold. During the training process of the detection model, the training loss can be determined by combining the thin-layer penalty loss, and the learnable parameters in the embedding representation module can be updated based on the training loss to adjust the embedding weights and embedding vectors corresponding to each time sub-window, thereby improving the feature extraction capability of the embedding representation module for thin-layer weak echoes.

[0136] Specifically, a pre-trained lightweight gating network can be used to generate embedding weights based on the echo energy corresponding to each time sub-window. Then, using the linear projection matrix and bias corresponding to each time sub-window, vector mapping is performed on the echo data within each time sub-window to obtain the sub-window embedding vector. Based on the embedding weights corresponding to each time sub-window, or the target embedding weights determined after echo correlation compensation, the embedding vectors of each sub-window are weighted and fused to obtain the echo embedding vector.

[0137] The computer device 106 can obtain a thin-layer penalty loss based on the thickness difference between the predicted local thickness and the actual local thickness corresponding to the thin-layer control point location, and then use the thin-layer penalty loss to perform backpropagation updates on the detection model. Specifically, the lightweight gating network parameters in the embedding representation module, the linear projection matrix corresponding to each time window, and the bias can all participate in parameter updates to adjust the embedding weights and echo feature mapping relationships for different time windows.

[0138] In this embodiment, for time sub-windows containing effective echo information of the thin layer, if their echo characteristics help reduce the thickness prediction error corresponding to the control point location of the thin layer, the thin layer penalty loss can be adjusted by backpropagation to guide the lightweight gating network and the corresponding linear projection parameters, so that the effective echo characteristics of the corresponding time sub-window play a greater role in the embedded representation. For time sub-windows that mainly contain noise or invalid signals, their interference with the thickness prediction results can be reduced through training. Combined with the aforementioned cross-location echo correlation compensation mechanism, the situation where effective weak echoes are excessively suppressed can be reduced, thereby improving the characterization ability of weak echoes in the thin layer and the thickness detection accuracy of the thin layer region.

[0139] It should be noted that during the training phase, the input to the embedding representation module is the sample terahertz time-domain echo; during the detection phase, the input to the trained embedding representation module is the terahertz time-domain echo of the thermal barrier coating under test, and the model parameters can remain fixed.

[0140] The computer device 106 can employ a trained position encoding module to obtain position encoding vectors corresponding to each physical scan position based on their spatial physical location. Specifically, the trained position encoding module can obtain the scan step size and total scan length corresponding to multiple physical scan positions; determine the spatial frequency range based on the scan step size and total scan length; encode the spatial physical location of each physical scan position based on the spatial frequency range to obtain the physical guidance position encoding corresponding to each physical scan position; and obtain the position encoding vector corresponding to each physical scan position based on the physical guidance position encoding and the position correction vector. The position correction vector can be a learnable residual vector updated during training, or it can be understood as a parameter used to correct the position representation deviation of the physical guidance position encoding.

[0141] It should be noted that the specific process by which the trained position encoding module generates the position encoding vector based on the scan step size, total scan length, spatial frequency range, physical guided position encoding, and position correction vector can be referred to the relevant description of the position encoding module processing the simulated scan position in the above training embodiment, and will not be repeated here. The difference is that the input in the detection stage is the spatial physical position of the physical scan position corresponding to the thermal barrier coating under test.

[0142] For each physical scan location, the computer device 106 can fuse the echo embedding vector and the position encoding vector corresponding to that physical scan location using a fusion module to obtain the fused echo vector corresponding to that physical scan location. After obtaining the fused echo vectors corresponding to multiple physical scan locations, the computer device 106 can arrange the multiple fused echo vectors according to the scanning order of the multiple physical scan locations to obtain the echo feature sequence of the thermal barrier coating under test.

[0143] In some embodiments, the computer device 106 can perform the following processing through a trained detection model: determining a spatial frequency range based on the scanning step size and total scanning length corresponding to multiple physical scanning positions; encoding the spatial physical position of each physical scanning position based on the spatial frequency range to obtain a position encoding vector corresponding to each physical scanning position; for each physical scanning position, fusing the echo embedding vector and the position encoding vector corresponding to the physical scanning position to obtain a fused echo vector corresponding to the physical scanning position, thereby obtaining an echo feature sequence of the thermal barrier coating under test; performing full-resolution self-attention processing on the echo feature sequence to obtain a fine branch feature sequence, and sequentially performing downsampling processing, self-attention processing, and sequence length recovery processing on the echo feature sequence to obtain a coarse branch feature sequence; determining a first fusion weight and a second fusion weight corresponding to each physical scanning position based on the fine branch feature sequence and the coarse branch feature sequence, and performing position-by-position fusion of the fine branch features and coarse branch features corresponding to each physical scanning position based on the first fusion weight and the second fusion weight to obtain a fused feature sequence; and determining the detection result of the ceramic layer in the thermal barrier coating under test based on the fused feature sequence.

[0144] The computer device 106 can use a trained inversion network to determine the detection result of the ceramic layer in the thermal barrier coating under test based on the echo feature sequence. Specifically, the trained inversion network may include one or more enhanced Transformer blocks, a global pooling layer, and a regression head. The enhanced Transformer block may include a multi-scale hierarchical self-attention module and a feedforward network.

[0145] In one possible implementation, the computer device 106 can use the MSHA module in the trained inversion network to perform self-attention processing on the echo feature sequence in the fine branches to obtain the fine branch feature sequence. The fine branch feature sequence includes the fine branch features corresponding to each physical scan position, and these fine branch features are used to characterize the local echo detail features corresponding to the respective physical scan position.

[0146] The computer device 106 can also use the MSHA module to sequentially perform downsampling, self-attention processing, and sequence length recovery processing on the echo feature sequence in the coarse branch to obtain the coarse branch feature sequence. The coarse branch feature sequence includes coarse branch features corresponding to each physical scan position, and these features characterize the spatial correlation between the corresponding physical scan position and other physical scan positions.

[0147] The computer device 106 can also use the MSHA module to combine the fine-branch feature sequence and the coarse-branch feature sequence to determine the first fusion weight and the second fusion weight corresponding to each physical scan position. The first fusion weight characterizes the contribution of the local echo detail features corresponding to the corresponding physical scan position in the fusion process, and the second fusion weight characterizes the contribution of the spatial correlation features corresponding to the corresponding physical scan position in the fusion process. Based on the first and second fusion weights corresponding to each physical scan position, the computer device 106 can perform position-by-position fusion of the fine-branch features and coarse-branch features corresponding to each physical scan position to obtain a fused feature sequence.

[0148] The computer device 106 can use a trained inversion network to sequentially perform feedforward processing, global pooling processing, and regression processing on the fused feature sequence to obtain the detection results of the ceramic layer in the thermal barrier coating under test. For example, the detection results may include thickness information of the ceramic layer in the thermal barrier coating under test. Optionally, the detection results may also include local thickness information corresponding to multiple physical scan locations or multiple control point locations, the relative permittivity of the ceramic layer, and defect detection results.

[0149] It should be noted that the specific processes of the trained inversion network in processing the echo feature sequence through fine-branching, coarse-branching, determining fusion weights, position-by-position fusion, feedforward processing, global pooling, and regression can be found in the relevant descriptions of the inversion network processing the sample echo feature sequence in the training examples above, and will not be repeated here. The difference lies in that the output of the detection stage is the detection result of the ceramic layer in the thermal barrier coating to be tested.

[0150] In this embodiment, by using a trained embedding representation module to perform feature mapping on multiple terahertz time-domain echoes of the thermal barrier coating under test, an echo embedding vector containing effective echo information can be obtained. By using a trained position encoding module to generate a position encoding vector, the correspondence between each terahertz time-domain echo and the actual physical scanning position can be preserved. By fusing the echo embedding vector and the position encoding vector to form an echo feature sequence, and using a trained inversion network to perform multi-scale self-attention processing, global pooling processing, and regression processing, local echo detail features and multi-position spatial correlation features can be comprehensively utilized, thereby improving the accuracy of the detection results of the ceramic layer in the thermal barrier coating under test.

[0151] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0152] Based on the same inventive concept, this application also provides a device for detecting ceramic layers in thermal barrier coatings to implement the above-described method for detecting ceramic layers in thermal barrier coatings. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the device for detecting ceramic layers in thermal barrier coatings provided below can be found in the limitations of the method for detecting ceramic layers in thermal barrier coatings described above, and will not be repeated here.

[0153] In one exemplary embodiment, such as Figure 9 As shown, a ceramic layer detection device in a thermal barrier coating is provided, comprising: an acquisition module 901, an embedding module 902, a positioning module 903, a fusion module 904, and a detection module 905, wherein: The acquisition module 901 is used to acquire multiple terahertz time-domain echoes of the thermal barrier coating under test; each terahertz time-domain echo is a time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. Embedding module 902 is used to perform feature mapping on each of the terahertz time-domain echoes to obtain the echo embedding vectors corresponding to each of the physical scan positions.

[0154] The position module 903 is used to obtain the position encoding vector corresponding to each of the physical scanning positions based on the spatial physical location of each of the physical scanning positions.

[0155] The fusion module 904 is used to fuse the echo embedding vector and the position encoding vector corresponding to each physical scanning position to obtain the echo feature sequence of the thermal barrier coating under test.

[0156] The detection module 905 is used to determine the detection result of the ceramic layer in the thermal barrier coating to be tested based on the echo feature sequence.

[0157] Each module in the aforementioned ceramic layer detection device for thermal barrier coatings can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0158] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting ceramic layers in thermal barrier coatings.

[0159] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0160] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for detecting ceramic layers in thermal barrier coatings provided in the first aspect of the present application.

[0161] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for detecting ceramic layers in thermal barrier coatings provided in the first aspect of the present application.

[0162] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method for detecting ceramic layers in thermal barrier coatings provided in the first aspect of the present application.

[0163] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0164] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0165] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0166] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting the ceramic layer in a thermal barrier coating, characterized in that, The method includes: Multiple terahertz time-domain echoes are obtained from the thermal barrier coating under test; each terahertz time-domain echo is a time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. Feature mapping is performed on each of the terahertz time-domain echoes to obtain the echo embedding vectors corresponding to each of the physical scan positions; Based on the spatial physical location of each physical scanning location, obtain the location encoding vector corresponding to each physical scanning location; For each physical scanning position, the echo embedding vector and the position encoding vector corresponding to the physical scanning position are fused to obtain the fused echo vector corresponding to the physical scanning position, so as to obtain the echo feature sequence of the thermal barrier coating under test. The detection result of the ceramic layer in the thermal barrier coating to be tested is determined based on the echo feature sequence.

2. The method according to claim 1, characterized in that, Before acquiring multiple terahertz time-domain echoes of the thermal barrier coating under test, the method further includes: Multiple training samples and their corresponding real labels are obtained; the training samples include multiple terahertz time-domain echoes of the thermal barrier coating under a simulated working condition; each terahertz time-domain echo is a time-domain echo response obtained by performing terahertz simulation scanning on the corresponding thermal barrier coating at different simulation scanning positions; the real labels include at least the real thickness information of the ceramic layer in the corresponding thermal barrier coating. The training samples include multiple terahertz time-domain echoes, which are input into the embedding representation module in the detection model to obtain the sample echo embedding vectors at each of the simulated scanning positions; the detection model also includes a position encoding module and an inversion network. The spatial physical locations of the multiple simulated scanning positions are input into the position encoding module to obtain the sample position encoding vector corresponding to each simulated scanning position. For each simulated scanning position, the sample echo embedding vector and the sample position encoding vector corresponding to the simulated scanning position are fused to obtain the sample echo feature sequence corresponding to the training sample; The sample echo feature sequence is input into the inversion network to obtain the sample prediction result corresponding to the training sample; the sample prediction result includes at least the predicted thickness information of the ceramic layer in the thermal barrier coating of the corresponding sample. Based on the sample prediction results and the true labels, the detection model is updated by backpropagation to obtain the trained detection model.

3. The method according to claim 2, characterized in that, The step of obtaining multiple training samples and the corresponding real labels for each training sample includes: Parametric geometric modeling of thermal barrier coatings is performed based on Bézier curves and / or Bézier surfaces to obtain multiple original thermal barrier coating geometric models. The control point positions of the Bézier curves and / or Bézier surfaces corresponding to the multiple original thermal barrier coating geometric models are adjusted to obtain multiple sample thermal barrier coating geometric models with different interface morphologies. Material parameters of multiple sample thermal barrier coating geometric models were set to obtain sample thermal barrier coatings under various simulation conditions. For each of the thermal barrier coatings of the sample, terahertz simulation scanning is performed at multiple simulation scanning positions to obtain the sample terahertz time-domain echo corresponding to each of the simulation scanning positions. Multiple terahertz time-domain echoes obtained under the same simulation condition are used as a training sample, and the actual thickness information of the ceramic layer in the thermal barrier coating of the corresponding sample is used as the real label of the training sample.

4. The method according to claim 3, characterized in that, The true thickness information includes the true local thickness corresponding to multiple control point positions; the predicted thickness information includes the predicted local thickness corresponding to multiple control point positions; the backpropagation update of the detection model based on the sample prediction results and the true labels includes: Based on the thickness difference between the predicted local thickness and the actual local thickness corresponding to the multiple control point locations, the relative thickness loss is obtained; From the multiple control point locations, determine the control point locations where the actual local thickness is less than the preset thin layer thickness threshold, and use these as the thin layer control point locations. Based on the thickness difference between the predicted local thickness and the actual local thickness corresponding to the thin layer control point location, the thin layer penalty loss is obtained. The detection model is updated by backpropagation by combining the relative thickness loss and the thin layer penalty loss.

5. The method according to claim 1, characterized in that, The step of performing feature mapping on each of the terahertz time-domain echoes to obtain the echo embedding vector corresponding to each of the physical scan positions includes: For each terahertz time-domain echo, the terahertz time-domain echo is windowed to obtain echo data corresponding to multiple time sub-windows; Based on the echo energy corresponding to the echo data within each time sub-window, the embedding weight corresponding to each time sub-window is determined; wherein, the embedding weight is positively correlated with the echo energy; Vector mapping is performed on the echo data within each time sub-window to obtain the sub-window embedding vector corresponding to each time sub-window. Based on the embedding weights corresponding to each of the time sub-windows, the embedding vectors of each sub-window are weighted and fused to obtain the echo embedding vector corresponding to the terahertz time-domain echo.

6. The method according to claim 1, characterized in that, The step of obtaining the location encoding vector corresponding to each of the physical scan locations based on their spatial physical locations includes: Obtain the scan step size and total scan length corresponding to multiple physical scan positions; the scan step size is used to characterize the actual physical distance between adjacent physical scan positions, and the total scan length is used to characterize the total distance covered by multiple physical scan positions along a preset scan direction. Based on the scanning step size and the total scanning length, a spatial frequency range is determined; the spatial frequency range is the range of values ​​for the spatial frequency, and the spatial frequency is used to characterize the scale of positional change along the preset scanning direction; Based on the spatial frequency range, the spatial physical position of each physical scanning position is encoded to obtain the physical guidance position code corresponding to each physical scanning position. Based on the physical guidance position code and position correction vector corresponding to each of the physical scanning positions, the position code vector corresponding to each of the physical scanning positions is obtained; the position correction vector is used to correct the position representation deviation of the corresponding physical guidance position code.

7. The method according to claim 1, characterized in that, The step of determining the detection result of the ceramic layer in the thermal barrier coating under test based on the echo feature sequence includes: The echo feature sequence is subjected to self-attention processing to obtain a fine branch feature sequence; wherein, the fine branch feature sequence includes fine branch features corresponding to each of the physical scan positions, and the fine branch features are used to characterize the local echo detail features corresponding to the physical scan positions; The echo feature sequence is sequentially subjected to downsampling, self-attention processing, and sequence length recovery processing to obtain a coarse branch feature sequence; wherein, the coarse branch feature sequence includes coarse branch features corresponding to each of the physical scan positions, and the coarse branch features are used to characterize the spatial correlation features between the corresponding physical scan position and other physical scan positions; By combining the fine branch feature sequence and the coarse branch feature sequence, a first fusion weight and a second fusion weight corresponding to each of the physical scan positions are determined; wherein, the first fusion weight is used to characterize the contribution of the local echo detail features corresponding to the corresponding physical scan position in the fusion process, and the second fusion weight is used to characterize the contribution of the spatial correlation features corresponding to the corresponding physical scan position in the fusion process. Based on the first fusion weight and the second fusion weight corresponding to each of the physical scanning positions, the fine branch features and coarse branch features corresponding to each of the physical scanning positions are fused position by position to obtain a fused feature sequence. The fused feature sequence is sequentially subjected to feedforward processing, global pooling processing, and regression processing to obtain the detection results of the ceramic layer in the thermal barrier coating to be tested.

8. A device for detecting ceramic layers in thermal barrier coatings, characterized in that, The device includes: The acquisition module is used to acquire multiple terahertz time-domain echoes of the thermal barrier coating under test; each terahertz time-domain echo is a time-domain echo response received after terahertz waves are emitted to the thermal barrier coating under test at different physical scanning positions. An embedding module is used to perform feature mapping on each of the terahertz time-domain echoes to obtain the echo embedding vectors corresponding to each of the physical scanning positions. The location module is used to obtain the location encoding vector corresponding to each of the physical scanning locations based on the spatial physical location of each of the physical scanning locations; The fusion module is used to fuse the echo embedding vector and the position encoding vector corresponding to each physical scanning position to obtain the fused echo vector corresponding to the physical scanning position, so as to obtain the echo feature sequence of the thermal barrier coating under test. The detection module is used to determine the detection result of the ceramic layer in the thermal barrier coating to be tested based on the echo feature sequence.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.