Nonlinear lamb wave mixing multi-task characterization method, system and electronic equipment for cracks in bridge steel plate-shaped connecting members

Abstract

This invention belongs to the field of structural health monitoring and nondestructive testing technology, specifically relating to a nonlinear Lamb wave mixing multi-task characterization method, system, and electronic equipment for cracks in bridge steel plate connection components. It involves sparsely arranging piezoelectric transducers on the component and applying dual-frequency Lamb wave excitation, extracting amplitude and phase features at the fundamental, sum, and difference frequencies, and performing reference state amplitude and phase differential calibration; constructing a physical guidance descriptor matrix from the calibrated features, and calculating path reliability weights based on path geometry; using the same path reliability weights simultaneously to generate a DAS spatial prior map (only for spatial prior) and cross-modal fusion task gating; obtaining crack center, length, direction, and crack probability enhancement maps through multi-task output, and improving characterization reliability through geometrically consistent soft rendering and deployment stage consistency verification.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring and nondestructive testing technology, specifically relating to a nonlinear Lamb wave mixing multi-task characterization method, system and electronic equipment for cracks in bridge steel plate connecting components. Background Technology

[0002] Under long-term alternating loads and complex service environments, steel plate-like connecting components of bridges (such as connecting nodes and connecting plates) are prone to the initiation of fatigue cracks and closed cracks within the plate. In their early closed state, these cracks have very weak disturbances to linear ultrasonic parameters (amplitude attenuation, arrival time difference, and phase change), often below the linear detection threshold, making them difficult to effectively identify and quantitatively characterize. Lamb waves, as guided waves propagating in plate-like structures, can cover a large detection area with a single excitation, making them suitable for large-scale scanning of steel plate-like connecting components of bridges. Furthermore, when two Lamb waves of different frequencies undergo contact nonlinear interaction at a closed crack, the received signal generates sum and difference frequency components not present in the input frequency. These nonlinear mixing components are sensitive to early closed cracks and can serve as characteristic carriers for crack identification.

[0003] However, existing technologies still face several interconnected bottlenecks in accurately characterizing cracks in steel plate-like connecting components of bridges using nonlinear mixing Lamb waves. Due to limitations in on-site installation conditions, piezoelectric transducers can typically only be sparsely arranged, resulting in a limited number of available transmit and receive paths. The signal reliability of different detection paths varies significantly due to differences in propagation distance, relative geometric relationship to the crack, coupling state, and on-site interference. Existing imaging methods often treat all paths equally, making it difficult to suppress interference from paths with low signal-to-noise ratios, severe phase mismatches, or insufficient geometric coverage, leading to artifacts at structural boundaries and inhomogeneous material regions.

[0004] In one class of existing technologies, a nonlinear mixing response is generated using dual-frequency composite excitation, and the elastic wave equation is used as a constraint term in a physical information neural network. Adaptive imaging is then used to obtain the two-dimensional distribution of damage, and the imaging results are fed back to adjust the sensor layout, forming a hardware optimization loop of detection-imaging-feedback. This type of method relies on accurate modeling of the wave equation, directly using the imaging results as the output for damage localization and measurement. Furthermore, it does not explicitly quantify the reliability of each detection path, making it difficult to suppress the contribution of paths with low signal-to-noise ratios, phase mismatches, or insufficient geometric coverage to the final result.

[0005] In another type of existing technology, the received mixed signal is subjected to Fourier transform and normalization to form a spectrum, which is then directly input into an end-to-end neural network to simultaneously provide the location, orientation, and length category of the crack in a discrete classification manner. This type of method uses the original spectrum as input, lacks interpretable features oriented towards the physical mechanism, is highly dependent on the number of samples, and uses discrete categories to represent crack geometric parameters, making it difficult to output continuous crack center coordinates, length, and orientation. Furthermore, it lacks the ability to perform geometric consistency verification between crack geometric parameters and crack probability distribution during the deployment phase.

[0006] In addition, although existing technologies have disclosed nonlinear Lamb wave mixing for crack detection, mixing-based direction recognition, reference signal-based differential processing, DAS or RAPID-type delay summation imaging, and deep learning recognition of guided wave signals, none of these technologies are sufficient to achieve multi-task characterization of cracks in bridge steel plate connecting components. Summary of the Invention

[0007] To address the technical challenges of forming a complete and mutually supportive technical chain under sparse detection paths and early-closing crack conditions, this involves: reference state amplitude-phase differential calibration; path reliability weights explicitly constructed from nonlinear mixing amplitude-phase characteristics and path geometry; the consistent effect of these weights on spatial priors and cross-modal gating; and geometric consistency constraints and deployment stage consistency verification between crack geometry parameters and crack probability distribution. This chain aims to achieve multi-task characterization of crack center location, length, direction, and probability distribution.

[0008] One objective of this invention is to propose a nonlinear Lamb wave mixing multi-task characterization method for cracks in steel plate-like connecting components of bridges, which includes the following steps: S1, piezoelectric transducers are sparsely arranged on the steel plate connecting members of the bridge. The piezoelectric transducers excite and receive Lamb waves along the sparse detection path in the member, and establish the path geometric relationship between each sparse detection path and the coordinates of the piezoelectric transducer. S2, applying a first fundamental frequency to the piezoelectric transducer With the second base frequency The dual-frequency excitation causes the excited Lamb wave to undergo nonlinear mixing interaction at the crack, generating a sum-frequency-included signal in the received signal. With difference frequency The nonlinear frequency components; S3. Perform frequency domain processing on the received signal of each detection path, extract the amplitude and phase at the fundamental frequency, sum frequency and difference frequency, and obtain the nonlinear mixing amplitude and phase characteristics of each detection path. S4, obtain the nonlinear mixing amplitude and phase characteristics of the detection path when the component is in the reference state, and perform reference state amplitude and phase differential calibration on the nonlinear mixing amplitude and phase characteristics of the test state to obtain the calibrated nonlinear mixing amplitude and phase characteristics. S5, construct a physical guidance descriptor matrix from the calibrated nonlinear mixing amplitude and phase characteristics. Each row of the physical guidance descriptor matrix corresponds to a physical guidance descriptor for a detection path. The physical guidance descriptor matrix includes amplitude scaling features, phase sine and cosine features, mixing proxy features, phase mismatch features, path geometry features, and path consistency features. S6. Calculate the path reliability weight of each detection path based on the calibrated nonlinear mixing amplitude and phase characteristics and the geometric relationship of the detection path. S7. Based on the path reliability weight, perform reliability-weighted delay summation imaging on each detection path to generate a DAS spatial prior map. The DAS spatial prior map is only used as a priori for the spatial location of the crack, and not as the final crack location result. S8, the physical guidance descriptor matrix and the DAS spatial prior map are respectively input into the feature coding branch and the image coding branch, and a fused representation is obtained through cross-modal fusion. The cross-modal fusion is then gated with the same path reliability weight. S9, the crack center location, crack length, crack direction, and crack probability enhancement map are obtained from the fused representation through multi-task output; S10, perform geometrically consistent soft rendering based on the crack center position, crack length and crack direction, so that the crack probability enhancement map is constrained by the predicted crack geometric parameters. S11, during the deployment inference phase, the geometric consistency of the crack center location, crack direction and crack size is checked. When the consistency deviation exceeds the preset threshold, a subset of reliable paths is selected according to the path reliability weight for secondary inference, or a verification prompt and low confidence indicator are output. The bridge steel plate connecting components include connecting plates, splice plates, cover plates, stiffening plates, node plates, welded connecting plates, lap plates, U-rib-panel connection areas or diaphragm connection areas in the bridge structure; the cracks include fatigue cracks, closed cracks, weld toe neighborhood cracks, hole edge cracks, plate edge cracks or bolt hole neighborhood cracks.

[0009] Another objective of this invention is to provide a nonlinear Lamb wave mixing multi-task characterization system for cracks in bridge steel plate-like connecting components, which is used to implement the above-mentioned method and includes: The excitation and acquisition module is used to apply dual-frequency excitation including a first fundamental frequency and a second fundamental frequency from sparsely arranged piezoelectric transducers and to acquire the received signals from each detection path. The frequency domain amplitude and phase extraction module is used to extract and obtain the calibrated nonlinear mixing amplitude and phase characteristics after reference state amplitude and phase differential calibration. The Physics-Guided Descriptor Constructor Module is used to construct the Physics-Guided Descriptor Matrix; The path reliability calculation module is used to calculate the path reliability weight of each detection path; the spatial prior generation module is used to generate a DAS spatial prior map based on the path reliability weights. A multi-task characterization module is used to perform cross-modal fusion and task gating of the physical guidance descriptor matrix and the DAS spatial prior map, and output the crack center location, crack length, crack direction, and crack probability enhancement map; and The geometric consistency processing module is used for geometric consistency soft rendering, geometric consistency verification, secondary inference, or review prompts.

[0010] Another object of the present invention is to provide an electronic device comprising a processor and a memory, wherein the memory stores a program, and the processor executes the program to implement the above-described method.

[0011] It is understood that the method of the present invention includes: sparsely arranging piezoelectric transducers on the steel plate-like connecting members of a bridge; exciting Lamb waves along a sparse detection path with dual-frequency excitation including a first fundamental frequency f1 and a second fundamental frequency f2, causing nonlinear mixing interaction at the crack, generating sum frequency fΣ and difference frequency fΔ components in the received signal; extracting the amplitude and phase at the fundamental frequency, sum frequency, and difference frequency for each detection path; obtaining the corresponding features of the reference state and performing reference state amplitude-phase differential calibration on the features of the state to be tested; constructing a physical guidance descriptor matrix from the calibrated nonlinear mixing amplitude-phase features, and calculating the path's potential from the calibrated nonlinear mixing amplitude-phase features and the geometric relationship of the detection path. Reliability weights are used; reliability-weighted delayed summation imaging is performed using the path reliability weights to generate a DAS spatial prior map that serves only as a spatial prior; the physical guidance descriptor matrix and the DAS spatial prior map are input into the feature encoding branch and the image encoding branch, respectively, and then fused across modally and gated using the same path reliability weights; after multi-task output, the crack center location, crack length, crack direction, and crack probability enhancement map are obtained; geometrically consistent soft rendering is used to constrain the crack probability enhancement map to the predicted crack geometric parameters; and geometric consistency verification is performed during the deployment and inference phase, and if necessary, a subset of reliable paths is selected according to the path reliability weights for secondary inference or a verification prompt is output.

[0012] Therefore, the present invention can have the following beneficial effects: Firstly, this invention limits the path reliability weight to a physical quantity explicitly constructed from the calibrated nonlinear mixing amplitude and phase characteristics (fundamental frequency signal-to-noise ratio, mixing stability, phase mismatch consistency) and path geometric effectiveness, rather than ordinary imaging weights or arbitrary attention weights learned by the network. Furthermore, the same path reliability weight is applied simultaneously to the task gating of DAS spatial prior map generation and cross-modal fusion, thereby systematically suppressing the contribution of low-reliability paths to spatial priors and multi-task outputs under sparse path conditions. Secondly, unlike existing technologies that use the elastic wave equation as a constraint of the physical information neural network, directly use the adaptive imaging results as the localization output, and adjust the sensor layout through imaging feedback, this invention replaces the reliance on precise modeling of the wave equation with reference state amplitude-phase differential calibration, and strictly limits the DAS spatial prior map to a spatial prior rather than the final localization result, so that the final location of the crack is determined by the multi-task output and geometric consistency constraints. This invention performs geometric consistency verification of position, direction and scale during the deployment inference stage, and selects a subset of reliable paths according to the path reliability weight for secondary inference or output verification prompts when the consistency is lower than the threshold, rather than optimizing the hardware by adjusting the sensor layout through feedback. Third, unlike existing end-to-end technologies that use the original spectrum as input and output crack parameters in a discrete classification manner, this invention first constructs a physical guide descriptor matrix from the calibrated nonlinear mixing amplitude and phase features, which simultaneously includes amplitude scaling, phase sine and cosine, mixing proxy, phase mismatch, path geometry, and path consistency. This gives the input features a clear physical meaning and makes them suitable for small sample conditions. This invention outputs the crack center coordinates, length values, and direction vectors represented by cos(2θ) and sin(2θ) in a continuous regression manner. It also uses geometrically consistent soft rendering to constrain the continuous crack geometric parameters and the crack probability enhancement map, thereby obtaining continuous, interpretable, and spatially consistent crack characterization results.

[0013] It should be noted that the above-mentioned beneficial effects are the result of the combined effect of the entire technology chain of the present invention, and the simple superposition of individual technical points cannot achieve the above effects. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the arrangement of bridge steel plate connecting components, sparsely arranged piezoelectric transducers, sparse detection paths and crack areas in a specific embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of the present invention in a specific embodiment, where dual-frequency excitation generates sum frequency and difference frequency through nonlinear interaction, and the nonlinear mixing amplitude and phase characteristics are extracted from the amplitude and phase. Figure 3 This is a schematic diagram of the path-frequency-amplitude-phase feature tensor in a specific embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the construction of a physically guided descriptor matrix from a feature tensor in a specific embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the generation of a DAS spatial prior graph based on path reliability weights in a specific embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of a dual-branch cross-modal multi-task representation network in a specific embodiment of the present invention; Figure 7 This is a schematic diagram of geometrically consistent soft rendering and crack probability enhancement map generation in a specific embodiment of the present invention; Figure 8 This is a block diagram representing the connection relationship of various modules of the system in a specific embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device and a computer-readable storage medium in a specific embodiment of the present invention; Figure 10 This is a schematic diagram illustrating the composition of reference state amplitude-phase differential calibration, path reliability weights, and their dual-use relationship in a specific embodiment of the present invention. Figure 11 This is a schematic diagram of the training-calibration-deployment consistency verification closed loop in a specific embodiment of the present invention; Figure 12 This is a schematic diagram comparing the reliability weighted delay summation space prior map and the geometrically consistent soft-rendered crack probability enhancement map in a specific embodiment of the present invention (Q235 steel plate simulation example). Figure 13 This is a schematic diagram of the Lamb wave dispersion curve and mixing operation point of Q235 steel plate in S0 mode in a specific embodiment of the present invention.

[0015] in, 1. Bridge steel plate connecting components; 2. Piezoelectric transducer; 3. Sparse detection path; 4. Crack region; 5. Excitation and acquisition module; 6. Nonlinear interaction region; 7. Received signal spectrum; 8. Frequency domain amplitude and phase extraction module; 9. Feature tensor construction module (path-frequency-amplitude and phase feature tensor); 10. Physically guided descriptor construction module; 11. Spatial prior graph generation module; 12. Feature coding branch; 13. Image coding branch; 14. Cross-modal fusion module; 15. Task gating module; 16. Multi-task representation module; 17. Enhanced graph generation module; 18. Result display and storage module; 19. Electronic equipment; 20. Processor; 21. Memory; 22. Program; 23. Bus; 24. Acquisition interface / communication interface; 25. Computer-readable storage medium. Detailed Implementation

[0016] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. The specific numerical values, dimensions, and exemplary formulas given in this specification are for illustrative purposes only and are not intended to limit the scope of protection of this invention, nor are they intended to limit the final engineering testing accuracy.

[0017] Incentives and data collection, such as Figure 1 As shown, several piezoelectric transducers 2 are sparsely arranged on the steel plate-shaped connecting member 1 of the bridge. The piezoelectric transducers 2 are combined in pairs to form sparse detection paths 3. The number P of sparse detection paths 3 is less than the total number of possible transmit and receive paths. The suspected crack is located in the crack region 4. For example... Figure 2 As shown, the excitation and acquisition module 5 applies dual-frequency excitation, including a first fundamental frequency f1 and a second fundamental frequency f2, to the piezoelectric transducer 2. The two excited Lamb waves undergo contact nonlinear interaction in the nonlinear interaction region 6 at the crack 4. In the received signal spectrum 7, in addition to the fundamental frequencies f1 and f2, a sum frequency f also appears. Σ With difference frequency f Δ Components. Where f1 and f2 are not equal, preferably the first fundamental frequency f1, the second fundamental frequency f2, and the sum frequency f Σ =f1+f2 and the difference frequency f Δ =|f1−f2| The four values ​​do not overlap, and f Σ f Δ All of them fall within the excitation and reception frequency band of the Lamb wave in the component, and the amplitudes of the sum frequency and difference frequency components vary with the degree of nonlinearity of the crack contact.

[0018] The frequency domain amplitude and phase extraction and reference state amplitude and phase differential calibration are as follows: the frequency domain amplitude and phase extraction module 8 performs frequency domain transformation on the received signal of each detection path and extracts the amplitude A at the fundamental frequency, sum frequency, and difference frequency. 1,k A 2,k A Σ,k A Δ,k With phase , , , Where k is the detection path number.

[0019] To eliminate the influence of common factors such as coupling differences, temperature drift, and system transfer function, the nonlinear mixing amplitude and phase characteristics of the detection path corresponding to the component in the reference state are obtained, and the amplitude and phase differential calibration of the test state characteristics is performed. For example, logarithmic amplitude differential calibration can be expressed as: ; Phase differential calibration can be expressed as: ; Where i corresponds to the first fundamental frequency, the second fundamental frequency, the sum frequency, and the difference frequency, respectively; k is the detection path number; the superscripts test and ref represent the state to be tested and the reference state, respectively; and wrap(·) is the phase wrapping function that constrains the phase to a preset principal value interval. Figure 10 As shown, Aref, phiref and Atest, phitest are obtained from the reference state signal and the state signal under test, respectively, and ΔA and Δphi are obtained from the amplitude and phase differential calibration module, which constitute the calibrated nonlinear mixing amplitude and phase characteristics.

[0020] The feature tensor, physical guidance descriptor matrix, and path reliability weights are constructed as follows: Figure 3 As shown, the feature tensor construction module 9 constructs each detection path in the frequency dimension {f1,f2,f...} Σ ,f Δ The calibrated feature organization on the channel dimension {A,phi} is the path-frequency-amplitude-phase feature tensor X.

[0021] like Figure 4 As shown, the physical guidance descriptor construction module 10 constructs a physical guidance descriptor matrix D from the feature tensor X. Each row of D corresponds to a physical guidance descriptor for a detection path, and Dd is the descriptor dimension. The physical guidance descriptor matrix includes: amplitude scaling features ai,k; phase sine and cosine features... , Mixer proxy features , Phase mismatch characteristics , Path geometric features ℓk , , mk; and the path consistency feature ck.

[0022] Specifically, the amplitude scaling feature ai,k can be obtained from the calibrated logarithmic amplitude difference Δai,k through standardization, max-min normalization, or robust scaling. For example, when using standardization, the amplitude scaling feature can be expressed as: ; Where, μ i and σ i Let be the mean and standard deviation of the i-th frequency component along each detection path, respectively, and ε be a stable term greater than 0. The phase sine and cosine characteristics are derived from the calibrated phase difference. A construction designed to avoid phase discontinuities at the boundaries of the principal value intervals, exemplarily: ; For the k-th detection path, let its transmitter coordinates be Tk=(x T,k ,yT,k The coordinates of the receiving end are Rk=(x R,k ,y R,k If the width and height of the detection area are W and H respectively, then the path length is... Path direction θ k and path midpoint feature m k They can be represented as: ; ; ; Inter-path consistency feature c k This characterizes the consistency between the phase mismatch response of the k-th detection path and other detection paths. For example, it can be constructed based on the path differences in sum-frequency phase mismatch and difference-frequency phase mismatch: ; Where j is the detection path number different from k, σ δ Let be the phase mismatch difference scale parameter. Therefore, the physical guidance descriptor for the k-th detection path can be denoted as: ; Stacking the physical guidance descriptors of each detection path row by row yields the physical guidance descriptor matrix D: ; For example, the calibrated sum-frequency mixer and difference-frequency mixer can be represented as follows: ; ; The calibrated sum-frequency phase mismatch and difference-frequency phase mismatch can be expressed as follows: ; ; Here, ε represents a stable term greater than 0. Unlike the method of directly using the original spectrogram as network input, the above-mentioned physical guidance descriptor has a clear physical meaning, which facilitates the improvement of the robustness and interpretability of the representation under small sample conditions.

[0023] Furthermore, the path reliability weight of each detection path is calculated based on the calibrated nonlinear mixing amplitude and phase characteristics and the geometric relationship of the detection path. For example... Figure 10 As shown, the path reliability weight is at least composed of the fundamental frequency signal-to-noise ratio component r. snr,k Mixing stability component r mix,k Phase mismatch consistency component r phase,k With path geometric validity component r geo,k The three determinations in the example are: ; ; Where s k Let R be the unnormalized reliability score of the k-th detection path. k Here, q represents the normalized path reliability weights, s represents the summation path number, and s represents the path reliability weights. q Let q be the unnormalized reliability score of the q-th detection path, P be the total number of detection paths, α1, α2, α3, and α4 be non-negative weight coefficients, and ε be a stable term greater than 0.

[0024] The generation of the DAS spatial prior map is as follows: Figure 5 As shown, the spatial prior map generation module 11 is based on the path reliability weight R. k Delay-sum imaging with reliability weighting is performed on each detection path 3 to obtain the DAS spatial prior map. Specifically, a candidate imaging grid is established within the detection area, and for any candidate crack location r=(x,y), the emitter T of the k-th detection path is calculated. k To candidate position r, from candidate position r to receiver R k The propagation distance, and based on the group velocity c corresponding to the selected Lamb wave mode or mixing component. g Calculate the equivalent propagation delay.

[0025] ;

[0026] Let E k (t) represents the equivalent delay response of the k-th detection path after envelope extraction at the fundamental frequency, sum frequency, or difference frequency. Then, the reliability-weighted DAS response at candidate position r can be expressed as: ; Calculate I at all candidate grid points DAS(r) The data is then normalized to generate a DAS spatial prior map. This DAS spatial prior map forms a high-response region near the candidate crack location, but it is only used as a priori input for the crack spatial location and participates in subsequent cross-modal fusion. It is not directly used as the final crack localization result through maximum response point, centroid, or threshold segmentation. The final crack localization is determined by the multi-task output and geometric consistency constraints.

[0027] Two-branch cross-modal multi-task representation, such as Figure 6As shown, feature encoding branch 12 encodes the physical guided descriptor matrix D, and image encoding branch 13 encodes the DAS spatial prior map IDAS. The two are then fused through cross-modal attention or an equivalent cross-modal weighted fusion via cross-modal fusion module 14 to obtain a fused representation. For example, the feature encoding branch can be implemented using a multilayer perceptron, a one-dimensional convolutional network, or a Transformer encoder, while the image encoding branch can be implemented using a convolutional network or a lightweight image encoder.

[0028] ; ;

[0029] Among them, F D For path-level feature representation, F I For spatial image representation. Cross-modal fusion can perform cross-modal attention by using image representation as a query and path-level feature representation as a key and value, or it can be achieved using equivalent methods such as concatenation, weighted summation, or gated fusion. For example: ; Task gating module 15 uses the same path reliability weight R k The contribution of each detection path to cross-modal fusion is weighted. For example, reliability-weighted aggregation of path-level feature representations can be performed: ; Therefore, the reliability weight R of the same path k It functions both in generating the DAS spatial prior map and in the task gating of cross-modal fusion, achieving "dual use of a single weight." The multi-task output head outputs the crack center location, crack length, crack direction, and crack probability enhancement map based on the fusion representation and the reliability-weighted path representation. ; in, The location of the crack center. The length of the crack. Let the crack direction vector be... This is a crack probability enhancement map. The specific number of network layers, channels, and activation functions of the encoder, fusion module, and output head can be adjusted equivalently according to the training sample size and deployment computing power, without changing the consistent function of this invention in generating spatial priors and gating tasks using path reliability weights.

[0030] The multi-task characterization module 16 outputs the fused characterization data via multiple tasks, obtaining the crack center location, crack length, crack direction, and crack probability enhancement map. Crack depth or equivalent damage level parameters can be used as extended outputs. Specifically, the crack center location is output as a continuous coordinate regression within the detection area, the crack length is output as a continuous numerical regression, and the crack direction is preferably output as a continuous vector regression consisting of cos(2θ) and sin(2θ) to accommodate the 180° periodicity of the crack line segment direction.

[0031] Geometrically consistent soft rendering and enhanced graph generation, such as Figure 7 As shown, the enhancement map generation module 17 performs geometrically consistent soft rendering based on the crack center position, crack direction, and crack length output by the multi-task characterization module 16, and further based on the crack depth or equivalent damage degree parameters when outputting the parameters: it constructs a differentiable two-dimensional spatial distribution with the crack center position as the center, the crack direction as the major axis direction, and the crack length as the major axis scale, and constrains the crack probability enhancement map with the two-dimensional spatial distribution, so that the high probability region of the crack probability enhancement map is consistent with the predicted crack geometric parameters in space, thereby improving the geometric consistency between the crack probability enhancement map and the crack geometric parameters.

[0032] To demonstrate the feasibility of the method of this invention in structural steel for bridges, physical simulation was performed using Q235 steel plate (6mm thick) as the object. The phase velocity and group velocity dispersion curves of the S0 mode Lamb wave were obtained by numerically solving the Rayleigh-Lamb characteristic equation, as shown below. Figure 13 As shown, based on this, the first fundamental frequency f1 = 120kHz and the second fundamental frequency f2 = 180kHz are selected, and their sum frequency f Σ =300kHz, difference frequency f Δ =60kHz, f1, f2, f Σ f Δ The four frequencies do not overlap and all fall within the excitation and reception bands of the S0 mode, thus allowing the amplitude and phase at the fundamental frequency, sum frequency, and difference frequency to be extracted from the received signal without spectral degradation. A sparse path mixing signal is generated using a semi-analytical Lamb wave propagation and crack second-order contact nonlinear scattering model, and reliability-weighted delay summation imaging is performed according to Example 4.

[0033] like Figure 12 As shown in (a), due to the limitations of sparse detection paths, the energy of the reliability-weighted delay summation prior map is relatively diffuse and contains elliptical arc artifacts, which can only provide a rough prior to the crack location and cannot directly provide the crack direction and scale; Figure 12As shown in (b), after geometrically consistent soft rendering, the high-probability regions of the crack probability enhancement map are concentrated into elliptical regions with the same center, direction, and length as the actual cracks. This confirms the beneficial effect of using the DAS spatial prior map only as a spatial prior and constraining the crack probability distribution with geometrically consistent soft rendering. The above simulation is used to verify the physical feasibility of the method. The relevant values ​​and figures are simulation examples and are not intended to limit the engineering detection accuracy and protection scope of this invention.

[0034] Geometric consistency verification and secondary inference during deployment phase, such as Figure 11 As shown, during the deployment inference phase, the crack center location, crack direction, and crack scale output by the multi-task representation module 16 are geometrically consistent. The crack center location is compared with the high-response region of the DAS spatial prior map to obtain the center deviation; the crack direction is compared with the principal axis direction of the crack probability enhancement map to obtain the direction deviation; and the crack length is compared with the scale of the crack probability enhancement map to obtain the scale deviation. When any of the center deviation, direction deviation, or scale deviation exceeds the corresponding threshold, a subset of reliable paths is selected according to the path reliability weight Rk for secondary inference, or a verification prompt and a low-confidence flag are output. It should be noted that the secondary inference and verification prompt in this embodiment are consistency processing performed at the inference and representation levels, rather than hardware optimization through feedback adjustment of the piezoelectric transducer layout.

[0035] Model training and calibration methods, such as Figure 11 As shown, the training-calibration-deployment consistency verification closed loop includes: (a) acquiring samples in the reference state and the test state; (b) generating a geometrically consistent target map corresponding to the crack geometric parameters of each sample; (c) using the physical guidance descriptor matrix and the DAS space prior map as input, jointly training a multi-task model including feature encoding branch 12, image encoding branch 13, cross-modal fusion module 14, task gating module 15 and multi-task representation module 16, and using geometrically consistent soft rendering as a constraint term during training to calibrate and update model parameters and path reliability weight calculation parameters; after deployment inference, if the deviation of geometric consistency verification exceeds the threshold, a second inference is triggered and calibration is performed accordingly, forming a training-calibration-deployment consistency verification closed loop.

[0036] like Figure 8As shown, this invention also proposes a nonlinear Lamb wave mixing multi-task characterization system for cracks in bridge steel plate connecting components, including an excitation and acquisition module 5, a frequency domain amplitude and phase extraction module 8, and a feature tensor construction module 9 connected in sequence. The feature tensor construction module 9 is connected to the physical guidance descriptor construction module 10 and the path reliability weight calculation module, which is connected to the spatial prior map generation module 11 and the task gating module 15. Both the physical guidance descriptor construction module 10 and the spatial prior map generation module 11 are connected to the multi-task characterization module 16. The multi-task characterization module 16 is connected to the enhancement map generation module 17, the consistency verification module, and the result display and storage module 18. The path reliability weight calculation module is used to calculate the path reliability weight from the calibrated nonlinear mixing amplitude and phase characteristics and the geometric relationship of the detection path. The enhancement map generation module is used to constrain the crack probability enhancement map according to the crack center position, crack length, and crack direction. The consistency verification module is used to trigger secondary inference or output verification prompts and low confidence indicators during the deployment inference phase. The functions of each module correspond to those in the aforementioned method embodiments.

[0037] like Figure 9 As shown, the present invention also provides an electronic device 19, including a processor 20 and a memory 21. The memory 21 stores a program 22. When the processor 20 executes the program 22, it implements the aforementioned method or the aforementioned model training and calibration method. The processor 20 and the memory 21 interact with a data acquisition interface / communication interface 24 via a bus 23. The data acquisition interface / communication interface 24 is connected to an external device / network. The present invention also provides a computer-readable storage medium 25, on which the program 22 is stored. When the program 22 is executed by the processor, it implements the aforementioned method or the aforementioned model training and calibration method.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A nonlinear Lamb wave mixing multi-task characterization method for cracks in steel plate connecting components of bridges, characterized in that, Includes the following steps: S1, piezoelectric transducers are sparsely arranged on the steel plate connecting members of the bridge. The piezoelectric transducers excite and receive Lamb waves along the sparse detection path in the member, and establish the path geometric relationship between each sparse detection path and the coordinates of the piezoelectric transducer. S2, applying a first fundamental frequency to the piezoelectric transducer With the second base frequency The dual-frequency excitation causes the excited Lamb wave to undergo nonlinear mixing interaction at the crack, generating a sum-frequency-included signal in the received signal. With difference frequency The nonlinear frequency components; S3. Perform frequency domain processing on the received signal of each detection path, extract the amplitude and phase at the fundamental frequency, sum frequency and difference frequency, and obtain the nonlinear mixing amplitude and phase characteristics of each detection path. S4, obtain the nonlinear mixing amplitude and phase characteristics of the detection path when the component is in the reference state, and perform reference state amplitude and phase differential calibration on the nonlinear mixing amplitude and phase characteristics of the test state to obtain the calibrated nonlinear mixing amplitude and phase characteristics. S5, construct a physical guidance descriptor matrix from the calibrated nonlinear mixing amplitude and phase characteristics. Each row of the physical guidance descriptor matrix corresponds to a physical guidance descriptor for a detection path. The physical guidance descriptor matrix includes amplitude scaling features, phase sine and cosine features, mixing proxy features, phase mismatch features, path geometry features, and path consistency features. S6. Calculate the path reliability weight of each detection path based on the calibrated nonlinear mixing amplitude and phase characteristics and the geometric relationship of the detection path. S7. Based on the path reliability weight, perform reliability-weighted delay summation imaging on each detection path to generate a DAS spatial prior map. The DAS spatial prior map is only used as a priori for the spatial location of the crack, and not as the final crack location result. S8, the physical guidance descriptor matrix and the DAS spatial prior map are respectively input into the feature coding branch and the image coding branch, and a fused representation is obtained through cross-modal fusion. The cross-modal fusion is then gated with the same path reliability weight. S9, the crack center location, crack length, crack direction, and crack probability enhancement map are obtained from the fused representation through multi-task output; S10, perform geometrically consistent soft rendering based on the crack center position, crack length and crack direction, so that the crack probability enhancement map is constrained by the predicted crack geometric parameters. S11, during the deployment inference phase, the geometric consistency of the crack center location, crack direction and crack size is checked. When the consistency deviation exceeds the preset threshold, a subset of reliable paths is selected according to the path reliability weight for secondary inference, or a verification prompt and low confidence indicator are output. The bridge steel plate connecting components include connecting plates, splice plates, cover plates, stiffening plates, node plates, welded connecting plates, lap plates, U-rib-panel connection areas or diaphragm connection areas in the bridge structure; the cracks include fatigue cracks, closed cracks, weld toe neighborhood cracks, hole edge cracks, plate edge cracks or bolt hole neighborhood cracks.

2. The method according to claim 1, characterized in that, The first fundamental frequency of the dual-frequency excitation Second fundamental frequency , and frequency With difference frequency The four are different from each other and all fall within the excitation and reception frequency band of the Lamb wave in the component; the mixing proxy features include sum-frequency mixing proxy and difference-frequency mixing proxy; The sum-frequency mixing proxy is a normalized quantity of the ratio of the calibrated sum-frequency amplitude to the product of the two fundamental frequency amplitudes; the difference-frequency mixing proxy is a normalized quantity of the ratio of the calibrated difference-frequency amplitude to the product of the two fundamental frequency amplitudes; the phase mismatch feature includes sum-frequency phase mismatch and difference-frequency phase mismatch, the sum-frequency phase mismatch is the wrapped phase difference of the calibrated sum-frequency phase and the sum of the two fundamental frequency phases, and the difference-frequency phase mismatch is the wrapped phase difference of the calibrated difference-frequency phase and the difference between the two fundamental frequency phases.

3. The method according to claim 1, characterized in that, The reference state amplitude-phase differential calibration includes: taking the logarithm of the fundamental frequency, sum frequency, and difference frequency amplitudes of the state under test and subtracting it from the logarithm of the corresponding component of the reference state to obtain the logarithmic amplitude difference; subtracting the phase of the corresponding component of the state under test from the phase of the corresponding component of the reference state and then wrapping it with the phase to obtain the phase difference; and using the logarithmic amplitude difference and the phase difference as the calibrated nonlinear mixing amplitude-phase characteristics.

4. The method according to claim 1, characterized in that, The path reliability weight is a path-level explicit weight generated prior to cross-modal fusion or task gating, and the path reliability weight of the same detection path is simultaneously used as the delay response contribution coefficient of the detection path in the DAS space prior graph, and as the feature contribution coefficient related to the detection path in cross-modal fusion or task gating. The path reliability weights are determined according to the following formula or its normalized equivalent form: ； ； in, For the first The path reliability weight of each detection path For the first The unnormalized reliability score of each detection path. To sum the path numbers, For the first The unnormalized reliability score of each detection path. This represents the number of paths detected. , , and These are the fundamental frequency signal-to-noise ratio component, the mixing stability component, the phase mismatch consistency component, and the path geometry effectiveness component, respectively. , , and These are non-negative weighting coefficients. It is a stable term that is greater than 0.

5. The method according to claim 4, characterized in that, The fundamental frequency signal-to-noise ratio component is used to characterize the signal-to-noise ratio of the first fundamental frequency and the second fundamental frequency. The mixing stability component is used to characterize the stability of the sum-frequency or difference-frequency mixing agent in adjacent acquisitions or repeated acquisitions. The phase mismatch consistency component is used to characterize the consistency of the phase mismatch feature with the phase mismatch feature of other paths. The path geometric validity component is used to characterize the coverage effectiveness of the path on the detection area or candidate crack area. The path reliability weight is not directly replaced by the attention weight in the image coding branch or feature coding branch, and does not use the residual of the nonlinear Lamb wave equation as a necessary calculation term.

6. The method according to claim 1, characterized in that, The crack direction is represented by a periodic continuous vector, which includes the cosine component of a double angle of the crack direction. Sine component The multi-task output also includes a crack depth or equivalent damage level parameter, which is used to adjust the response width, response intensity, or confidence level of the crack probability enhancement map.

7. The method according to claim 1, characterized in that, The geometrically consistent soft rendering includes: constructing a differentiable two-dimensional spatial distribution with the crack center location as the center, the crack direction as the major axis direction, and the crack length as the major axis scale, and constraining the crack probability enhancement map with this two-dimensional spatial distribution, so that the high probability region of the crack probability enhancement map is spatially consistent with the predicted crack geometric parameters; the geometric consistency verification includes: comparing the crack center location with the high response region of the DAS spatial prior map to obtain the center deviation, comparing the crack direction with the major axis direction of the crack probability enhancement map to obtain the direction deviation, and comparing the crack length with the scale of the crack probability enhancement map to obtain the scale deviation. When any of the center deviation, direction deviation, or scale deviation exceeds the corresponding threshold, the secondary inference or the verification prompt is triggered; the secondary inference does not require adjusting the piezoelectric transducer spatial layout as a necessary step, but selects a subset of reliable paths according to the path reliability weight from high to low, and regenerates the physical guidance descriptor matrix or DAS spatial prior map based on the subset of reliable paths, and then calls the same multi-task representation model to perform secondary inference.

8. The method according to claim 7, characterized in that, The training and calibration steps of the multi-task representation model include: Acquire samples in the reference state and the test state, and generate a geometrically consistent target map corresponding to the crack geometric parameters of each sample; Using the physical guidance descriptor matrix and the DAS spatial prior map as input, a multi-task model including the feature coding branch, image coding branch, cross-modal fusion, task gating, and multi-task output is jointly trained. During training, the geometrically consistent soft rendering is used as a constraint to calibrate and update the parameters of the multi-task model and the weight coefficients of the path reliability weights. The samples include those generated based on a simulation model of a structural steel component for bridges. The simulation model is used to generate nonlinear mixing received signals for the corresponding detection path under different crack center locations, crack directions, and crack lengths, and is used to generate DAS spatial prior maps and geometrically consistent target maps.

9. A nonlinear Lamb wave mixing multi-task characterization system for cracks in steel plate connecting components of bridges, used to implement the method described in any one of claims 1-8, characterized in that, include: The excitation and acquisition module is used to apply dual-frequency excitation including a first fundamental frequency and a second fundamental frequency from sparsely arranged piezoelectric transducers and to acquire the received signals from each detection path. The frequency domain amplitude and phase extraction module is used to extract and obtain the calibrated nonlinear mixing amplitude and phase characteristics after reference state amplitude and phase differential calibration. The Physics-Guided Descriptor Constructor Module is used to construct the Physics-Guided Descriptor Matrix; The path reliability calculation module is used to calculate the path reliability weight of each detection path. A spatial prior generation module is used to generate a DAS spatial prior map based on the path reliability weights. The multi-task characterization module is used to perform cross-modal fusion and task gating of the physical guidance descriptor matrix and the DAS spatial prior map, and output the crack center position, crack length, crack direction and crack probability enhancement map. as well as The geometric consistency processing module is used for geometric consistency soft rendering, geometric consistency verification, secondary inference, or review prompts.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a program, and the processor executing the program to implement the method of any one of claims 1 to 8.

Images