Intelligent detection method for internal cavity of grouting body based on impact echo and deep learning

Abstract

The invention discloses a grouting body internal cavity intelligent detection method based on impact echoes and deep learning. The method comprises the steps of physical modeling and defect simulation, data acquisition and preprocessing, model construction and training and the like. By establishing the grouting layer defect feature database and training the intelligent recognition model, high-precision and automatic recognition of the internal defects of the grouting layer can be realized, and an innovative, efficient and generalizable technical approach is provided for solving the problem of synchronous grouting quality detection.

Description

Technical Field

[0001] This invention relates to the field of tunnel and underground engineering technology, and in particular to an intelligent detection method for internal cavities in grouting bodies based on impact echo and deep learning. Background Technology

[0002] The information disclosed in this background section is intended only to enhance understanding of the overall background of the invention and is not necessarily to be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

[0003] During shield tunnel construction, after the tunnel segments are assembled, there are often gaps between them and the surrounding soil. To fill these gaps, stabilize the segment structure, suppress ground deformation, and achieve waterproofing, simultaneous grouting is usually required. The quality of the grouting layer directly affects the long-term safety of the tunnel structure, its waterproof sealing performance, and the stability of surface structures.

[0004] However, due to the complex interplay of factors such as grouting pressure, grout flow characteristics, and geological conditions, hidden defects such as voids, looseness, delamination, or water abundance can easily occur within the grouting layer. These defects not only weaken the load-bearing capacity of the lining but can also lead to stress concentration and cracking in the tunnel segments, further forming seepage channels and seriously threatening the operational safety and service life of the tunnel. Currently, the main methods for detecting the effectiveness of simultaneous grouting include: (1) Ground Penetrating Radar (GPR): Although this method is widely used in engineering inspection, electromagnetic waves are severely attenuated in reinforced concrete and are interfered with by dense steel mesh, making it difficult to effectively penetrate and accurately identify defects in the grouting layer, and its ability to identify defect types is limited.

[0005] (2) Core drilling method: Although the results are intuitive and reliable, it is a destructive testing method with low efficiency and high cost. It can only reflect the local situation and cannot achieve large-scale and continuous quality evaluation.

[0006] (3) Impact-Echo (IE): This is a non-destructive testing technique based on the principle of stress wave propagation. It is traditionally used to measure the thickness of concrete or shallow defects. However, in actual engineering, the signals acquired by the impact-echo method often have a low signal-to-noise ratio and complex characteristics. Traditional methods based on frequency domain analysis or experience interpretation are difficult to accurately adapt to complex field conditions, and the stability and accuracy of the test results are difficult to guarantee. Summary of the Invention

[0007] To address the problems of the aforementioned impact echo method, this invention proposes an intelligent detection method for internal voids in grouting bodies based on impact echo and deep learning. By establishing a database of grouting layer defect features and training an intelligent recognition model, it can achieve high-precision and automated identification of internal defects in the grouting layer, providing an innovative, efficient, and scalable technical approach to solving the problem of synchronous grouting quality inspection. Specifically, the technical solution of this invention is as follows.

[0008] A smart detection method for internal voids in grouting bodies based on impact echo and deep learning includes the following steps: (1) Construct a physical model of the hardened grouting material containing preset defects to simulate void defects of different burial depths and sizes. Then, place detection lines on the surface of the physical model.

[0009] (2) The original echo signal of the physical model is obtained by using the impulse echo method. Then calculate the Energy cumulative distribution function Obtain an effective analysis time window This is used as the unified input range for all subsequent feature extraction steps.

[0010] (3) The acquired original echo signal Empirical Mode Decomposition (EMD), Singular Spectral Analysis (SSA), and Wavelet Thresholding (WTD) are performed sequentially for denoising, resulting in a purified signal. .

[0011] (4) To enhance the sensitivity to the reflection characteristics of the de-energized interface, the purification signal is... Normalization and Hilbert transform are performed to obtain the signal envelope spectrum. With instantaneous phase And further, the phase delay curve was obtained. .

[0012] (5) Regarding the purification signal Signal envelope spectrum Continuous wavelet transforms were performed using hybrid structures of Morlet and Mexican Hat wavelets, and the two results were then fused by frequency band weighting to obtain composite time-frequency features. It has multi-resolution features and can be used as input for subsequent deep learning models.

[0013] (6) The purification signal Signal envelope spectrum Phase delay curve Composite time-frequency characteristics Composition of joint eigenvectors: The feature tensor F after this processing has high correlation and low noise features, and can be directly input into the multimodal multi-scale attention fusion deep learning model (MSA-CNN-LSTM) constructed in this invention for defect identification.

[0014] (7) The feature tensor F contains The input is fed into the time-frequency feature extraction branch of the multimodal multi-scale attention fusion deep learning model (MSA-CNN-LSTM), and the output is multimodal features. The feature tensor F contains , , The inputs are fed into the temporal feature extraction branch of the deep learning model described above, and the output is multimodal features. .

[0015] (8) The multimodal features , The input is fed into the adaptive feature weighted fusion (AFWM) module of the deep learning model to obtain adaptive weights. After weighted summation and fusion, the fusion characteristics are obtained. After being processed by the classification head, it achieves high-precision discrimination of different defect categories, thereby realizing the identification and detection of defects in the inspected object.

[0016] Furthermore, in step (1), the preset defect in the model experiment can be simulated by a hollow acrylic box, whose wave impedance is small and the difference between it and the grouting material is small, so the reflected wave at the contact surface of the two is weak, and therefore the influence of this part of the reflected wave on the experiment can be ignored.

[0017] Furthermore, in step (2), the following formula is used for the above process. Energy cumulative distribution function: ,in For time, For the original echo signal at time The instantaneous amplitude.

[0018] Furthermore, in step (2), by obtaining the... The first derivative curve automatically identifies the moment when the energy growth rate drops to a set threshold. This moment is taken as the termination point of the main reflected energy. The effective analysis time window of the signal is thus dynamically determined. The This is a margin parameter. This time window can be automatically adjusted according to the excitation energy, the structural thickness of the physical model, and its material properties to avoid the loss of effective information or the introduction of noise due to time truncation.

[0019] Further, in step (3), the joint denoising includes the following steps: S3-1: Empirical Mode Decomposition (EMD): This process converts the original echo signal... The signal is decomposed into several intrinsic mode functions (IMFs), and the IMF1 containing high-frequency noise is removed to obtain the reconstructed signal of the intrinsic mode function components.

[0020] S3-2: Singular Spectrum Analysis (SSA): The inherent mode function components are reconstructed into a signal. Based on the eigenvalue decomposition of the trajectory matrix, the first k mode components corresponding to the main singular values ​​are retained to obtain a smooth reconstructed signal.

[0021] S3-3: Wavelet Thresholding Denoising (WTD): The smoothed reconstructed signal is then subjected to a soft threshold. , ( (where N is the noise standard deviation and N is the signal length) further reduces random high-frequency interference in the frequency domain. The signal obtained by the three-level joint denoising strategy of EMD-SSA-WTD (Empirical Mode Decomposition-Singular Spectrum Analysis-Wavelet Thresholding) proposed in this invention has a higher signal-to-noise ratio and retains the main reflection waveform characteristics, which can effectively avoid interference from on-site noise and instruments.

[0022] Furthermore, in step (4), the normalization and Hilbert transform are performed using the following formula: .in, For Hilbert transform operators, The imaginary unit ( The imaginary part of the complex signal is constructed using this method. The envelope spectrum and phase features of the signal after the above processing exhibit higher discriminative power in defect identification. The phase delay curve... It is used to reflect multipath reflection and changes in burial depth.

[0023] Furthermore, in step (5), the transform coefficients of the continuous wavelet transform (CWT) for: .in, It represents the complex conjugate of the mother wavelet. It is a time variable. , These represent the scale parameter and translation parameter, respectively; optionally, , .

[0024] Furthermore, in step (5), the composite time-frequency characteristics are calculated using the following formula: ,in For calculation using Morlet wavelets CWT Coefficient matrix. For calculation using MexicanHat wavelets CWTCoefficient matrix; , Here, are weighting coefficients, and a and b represent the scale parameter and translation parameter, respectively. Optionally, the weighting coefficients are determined based on the energy distribution. , ,and .

[0025] Furthermore, in step (7), the time-frequency feature extraction branch is based on a fused convolutional neural network (CNN) structure, employing multi-scale residual convolution and introducing a time-frequency attention mechanism. This invention extracts feature information at different time-frequency resolutions through parallel multi-scale convolution kernels and assigns adaptive weights in the two-dimensional time-frequency plane, highlighting defect-sensitive time intervals and frequency bands, thereby obtaining information-rich time-frequency weighted features.

[0026] Furthermore, in step (7), the temporal feature extraction branch is obtained by using a Long Short-Term Memory (LSTM) network combined with a temporal attention mechanism. Based on modeling the temporal signal sequence, this invention automatically highlights the key moment response related to the defect in the impact signal through attention weights, suppresses redundant noise periods, and thus obtains more discriminative temporal weighted features.

[0027] Furthermore, step (7) also includes transmitting the purification signal. Amplitude spectrum obtained by Fast Fourier Transform (FFT) The input is fed into the spectral feature extraction branch of the multimodal multi-scale attention fusion deep learning model (MSA-CNN-LSTM), and the output is multimodal features. This is to supplement the frequency domain energy distribution characteristics. Preferably, the frequency domain energy distribution characteristics are supplemented. With the , The fused features are obtained by inputting them into the adaptive feature weighting fusion module (AFWM). .

[0028] Furthermore, the spectral feature extraction branch is obtained using a lightweight convolutional neural network (Light CNN) combined with global average pooling (GAP). This invention, through the above processing, achieves the extraction of amplitude and energy distribution features in the frequency domain space, thus supplementing the deficiencies in spectral information provided by the time domain and time-frequency branches.

[0029] Furthermore, in step (8), the fusion feature The calculation formula is: .

[0030] Furthermore, in step (8), the Classification Head uses the Softmax activation function at the output end to achieve high-precision discrimination of different defect categories.

[0031] Furthermore, step (8) also includes: using the uncertainty estimation module in the multimodal multiscale attention fusion deep learning model (MSA-CNN-LSTM) to calculate the variance of the prediction result, and using it as the uncertainty index of the model output to quantify the credibility of the classification result in order to assist engineering decision-making.

[0032] Compared with the prior art, the present invention has at least the following beneficial technical effects: (1) Adaptive signal interception and multi-level noise reduction: This invention uses an adaptive time window driven by energy accumulation distribution and a joint EMD-SSA-WTD noise reduction strategy to automatically identify the effective reflection range and suppress high-frequency noise, thereby significantly improving the signal-to-noise ratio and making the reflected waveform clearer and more stable. This is because the joint noise reduction strategy of this invention is a "layer-by-layer stripping of noise chain" oriented towards the characteristics of impact echo signals. It realizes a specific noise reduction sequence of structural layering, trend correction and local refinement, effectively avoiding the destruction of reflected wave information and significantly improving the signal-to-noise ratio.

[0033] (2) Multi-source feature synergistic enhancement: Based on traditional time-frequency analysis, this invention introduces envelope and phase delay features after Hilbert transform to form a three-dimensional fusion feature space of time domain, frequency domain and phase, which improves the ability to distinguish between shallow and deep void reflections.

[0034] (3) Extraction of time-frequency features of composite wavelet: This invention adopts energy weighted fusion of Morlet and Mexican Hat wavelets, which significantly improves the resolution balance problem of different frequency bands and realizes multi-scale visualization and adaptive enhancement of the de-reflection signal.

[0035] (4) Multimodal and multiscale attention fusion recognition: This invention uses a multimodal network structure composed of time domain (LSTM + temporal attention), time frequency (CNN + time frequency attention) and optional spectrum (LightCNN + GAP) branches, and the adaptive feature weighted fusion module (AFWM) realizes dynamic weight allocation. The model can automatically adjust the feature contribution for different defect types.

[0036] (5) Improved intelligence and reliability: By introducing an uncertainty estimation module based on Monte Carlo Dropout technology, this invention can not only output the defect category, but also provide the prediction variance as an uncertainty assessment index, providing a confidence basis for engineering inspection, and realizing traceability of results and controllable risks.

[0037] In summary, this invention forms a complete technology chain from signal acquisition, preprocessing, feature extraction to multimodal intelligent recognition, which is superior to existing methods in terms of detection accuracy, robustness and engineering feasibility. Attached Figure Description

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0039] Figure 1 This is a schematic diagram of the void distribution preset in the physical model for the following embodiments.

[0040] Figure 2 This is a schematic diagram of the distribution of measuring lines and measuring points on the upper surface of the physical model in the following embodiments.

[0041] Figure 3 The following is a flowchart of the preprocessing of the impulse echo signal in the embodiments below.

[0042] Figure 4 The following is a schematic diagram of the structure of the multimodal multiscale attention fusion deep learning model (MSA-CNN-LSTM) in the following embodiments. Detailed Implementation

[0043] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0045] For ease of description, the terms "up," "down," "left," and "right" appearing in this invention only indicate that they correspond to the up, down, left, and right directions in the accompanying drawings. They do not limit the structure and are merely for the purpose of describing the invention and simplifying the description. They do not indicate or imply that the device or component referred to needs to have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. The detection method of the present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0046] A smart detection method for internal voids in grouting bodies based on impact echo and deep learning includes the following steps: Step (1): Construct a physical model of the hardened grouting material containing pre-defined defects to simulate voids at different burial depths and sizes. This includes the following steps: S1-1: A steel rectangular box with a length of 1.5 m, a width of 1 m, and a height of 40 cm is used as a mold, with an open top. S1-2: Five hollow acrylic boxes are pre-placed at the bottom of the mold, wherein: acrylic boxes one to three are cubes with a hollow side length of 20 cm, and their embedment depths (i.e., vertical depths) are 32 cm, 24 cm, and 12 cm respectively. Acrylic box four has a hollow dimension of 30 cm in length and 20 cm in width, and acrylic box five has a hollow dimension of 40 cm in length and 20 cm in width. The embedment depth of acrylic boxes four and five is 32 cm. Figure 1 As shown.

[0047] S1-3: After arranging the hollow acrylic box, use cement-water glass double liquid grout as the grouting material to pour it into the mold until it is full. Then, after hardening, remove the mold and cure for 28 days to obtain the physical model.

[0048] S1-4: Arrange a survey line network on the upper surface of the physical model, with its outer edge 15cm from the boundary and its interior divided into a 5cm×5cm grid to form a survey point array, such as... Figure 2 As shown.

[0049] Step (2), Signal Acquisition: According to Figure 2 The measured point and line array shown uses a 28 mm diameter vibrating hammer to excite elastic waves. An accelerometer is placed in close contact with the upper surface of the model to receive the echo. The sampling frequency of the sensor is set to 200 kHz to ensure that high-frequency components in the impact echo signal can be captured. The sampling duration for each monitoring point is 40 ms. During testing, the accelerometer is placed in close contact with the upper surface of the model, positioned 5 cm to the right of the impact point of the vibrating hammer. To ensure the representativeness of the sample data and the sufficiency of model training, 10 samples are collected from each monitoring point, resulting in a total of 3750 samples, covering dense areas and various void conditions.

[0050] Step (3): Use the following formula (1) to process the original echo signal obtained in the above steps. Accumulated energy distribution: (1); in, For time, For the original echo signal at time The instantaneous amplitude. Then, by calculating the... The first derivative line automatically identifies the moment when the energy growth rate drops to the 5% threshold. This moment is taken as the termination point of the main reflected energy. The effective analysis time window of the signal is thus dynamically determined. The The margin parameter is 0.5 ms. The effective analysis time window... This will serve as the unified input range for all subsequent feature extraction steps (including EMD-SSA-WTD denoising, Hilbert envelope and phase calculation, CWT composite wavelet analysis, and FFT spectrum extraction), ensuring that each feature module processes only the effective segments containing the main reflection information, avoiding background noise interference, and improving feature stability and recognition accuracy.

[0051] Step (4), as Figure 3 As shown, the joint denoising of the acquired original echo signal specifically includes the following steps: S4-1: Empirical Mode Decomposition (EMD): The original echo signal is decomposed into several intrinsic mode functions (IMFs), and IMF1 containing high-frequency noise is removed to obtain the reconstructed signal of the intrinsic mode function components.

[0052] S4-2: Singular Spectrum Analysis (SSA): The inherent mode function components are reconstructed into a signal. Based on the eigenvalue decomposition of the trajectory matrix, the first k mode components corresponding to the main singular values ​​are retained to obtain a smooth reconstructed signal.

[0053] S4-3: Wavelet Thresholding Denoising (WTD): Apply a soft threshold to the smoothed reconstructed signal. , ( (where N is the noise standard deviation and N is the signal length) further weakens random high-frequency interference in the frequency domain.

[0054] Step (5): To enhance sensitivity to the reflection characteristics of the de-energized interface, the purification signal... Perform normalization and Hilbert transform, such as Figure 3 As shown. Specifically, the normalization and Hilbert transform are performed using the following equation (2): (2); in, For Hilbert transform operators, The imaginary unit ( This is used to construct the imaginary part of a complex signal. From this, the signal envelope spectrum is obtained. With instantaneous phase The phase delay curve was further calculated. .

[0055] Step (6): To comprehensively characterize the time-frequency characteristics of the signal, the purified signal is... Signal envelope spectrum Continuous wavelet transform is performed using a hybrid structure of Morlet and Mexican Hat wavelets, respectively, with the transform coefficients... for: .in, It represents the complex conjugate of the mother wavelet. It is a time variable. a , b These represent the scale parameter and the translation parameter, respectively. a =8~54, corresponding to an effective frequency range of 3~20 kHz, b=5μs.

[0056] The proposed Morlet–MexicanHat wavelet hybrid structure is not a traditional single wavelet basis function, but rather a composite wavelet adaptively constructed based on the energy distribution of the purified signal at different scales. Specifically, the Morlet wavelet has high frequency resolution, making it suitable for extracting steady-state features of deep reflections; the MexicanHat wavelet has strong temporal localization capabilities, making it suitable for characterizing shallow burst reflection waves. This invention calculates the energy weights at each scale. and Construct the following composite wavelet transform: The two results are then fused by frequency band weighting to obtain the composite time-frequency characteristics. Specifically, the composite time-frequency characteristics are calculated using the following formula (3): (3); in, For calculation using Morlet wavelets CWT Coefficient matrix. For calculation using MexicanHat wavelets CWT Coefficient matrix; , Here, a and b represent the weighting coefficients, respectively, and the scale parameter and translation parameter. , Determined based on energy distribution, and .

[0057] The Morlet-MexicanHat wavelet hybrid structure constructed in this embodiment achieves joint enhancement of shallow and deep reflection information across the entire time-frequency scale domain.

[0058] Step (7): The purification signal Signal envelope spectrum Phase delay curve Composite time-frequency characteristics Composition of joint eigenvectors: This is then used in the corresponding branch of the subsequent multimodal multiscale attention fusion network.

[0059] Step (8), Reference Figure 4 , in the feature tensor F The input is fed into the time-frequency feature extraction branch of the multimodal multi-scale attention fusion deep learning model (MSA-CNN-LSTM), and the output is multimodal features. The time-frequency feature extraction branch is based on a fused convolutional neural network (CNN) structure, employing multi-scale residual convolution and introducing a time-frequency attention mechanism.

[0060] Step (9), Reference Figure 4 , in the feature tensor F , , The common inputs are fed into the temporal feature extraction branch of the learning model described above, and the output is multimodal features. The temporal feature extraction branch is obtained by combining a Long Short-Term Memory (LSTM) network with a temporal attention mechanism.

[0061] The multimodal features , The input is fed into the adaptive feature weighted fusion (AFWM) module of the deep learning model to obtain adaptive weights. After weighted summation and fusion, the fusion characteristics are obtained. The calculation formula is as follows: .

[0062] The fusion features The recognition result is output through two fully connected layers (128 and 64 neurons) and a classifier (the output uses the Softmax activation function). At the same time, the uncertainty estimation module Monte Carlo Dropout (20 times) in the multimodal multi-scale attention fusion deep learning model (MSA-CNN-LSTM) is used to calculate the prediction variance, which is used as the uncertainty index of the model output to quantify the credibility of the classification result.

[0063] This embodiment also uses traditional CNN or LSTM single-branch models for comparative experiments. The results show that the CNN model with single modality input has an accuracy of about 94.4% in the empty-out detection, the LSTM model has an accuracy of about 92.3%, and the detection method in this embodiment has an accuracy of over 96.8%, while having stronger robustness and generalization ability, verifying the effectiveness and superiority of the technical solution of this invention.

[0064] In another embodiment, the above-mentioned intelligent detection method for internal voids of grouting bodies based on impact echo and deep learning further includes: in step (8), the purification signal... Amplitude spectrum obtained by Fast Fourier Transform (FFT) The input is fed into the spectral feature extraction branch of the multimodal multi-scale attention fusion deep learning model (MSA-CNN-LSTM), and the output is multimodal features. This is to supplement the frequency domain energy distribution characteristics. The spectral feature extraction branch is obtained by using a lightweight convolutional neural network (Light CNN) combined with the global average pooling (GAP) method. Then, this... With the , The fused features are obtained by inputting them into the adaptive feature weighting fusion module (AFWM). .

[0065] In another implementation, to ensure the stability of the detection results, the aforementioned intelligent detection method for internal voids in grouting bodies based on impact echo and deep learning also incorporates a signal quality factor-based approach. Dynamic parameter adjustment mechanism: (4); In equation (4), Energy in the main reflection zone; Energy in the non-reflection region; To prevent the use of tiny constants with a denominator of zero (in this embodiment, the value is taken as...) Specifically, the aforementioned The effective analysis time window in step (3) above The signal energy within. The aforementioned For this effective analysis time window The energy within the wave mainly consists of wake noise and environmental background interference.

[0066] like (By setting a threshold), the system will automatically adjust the sampling time length and the effective analysis time window. Or noise reduction parameters to improve signal quality and recognition reliability. This mechanism achieves adaptive optimization of the impulse echo signal, enabling the detection system to have intelligent feedback and self-learning capabilities. Specifically, in cases where high signal quality is required, the aforementioned... It can be set to 30, in project use. It can generally be set to 20.

[0067] Finally, it should be noted that any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention. Although specific embodiments of this invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this invention are still within the scope of protection of this invention.

Claims

1. A method for intelligent detection of internal cavities in a grouting body based on an impact echo and deep learning, characterized in that, Comprise the following steps: (1) Constructing a physical model of grouting material hardening body containing preset defects to simulate different buried depths and sizes of void defects; then arranging detection lines on the surface of the physical model; (2) The original echo signal of the physical model is obtained by using the impulse echo method. Then calculate the Energy cumulative distribution function Obtain an effective analysis time window This is used as the unified input range for all subsequent feature extraction steps; (3) obtaining the original echo signal The experience mode decomposition, singular spectrum analysis and wavelet threshold joint denoising are sequentially performed, and a purified signal is obtained after completion ; (4) normalizing and Hilbert transforming the cleaned signal to obtain a signal envelope spectrum and an instantaneous phase to obtain a phase delay curve ; (5) purifying the signal , signal envelope spectrum The Morlet wavelet and the Mexican Hat wavelet are mixed to perform continuous wavelet transform, and two results obtained are fused according to frequency bands to obtain a composite time-frequency feature ; (6) the purifying signal , signal envelope spectrum , phase delay curve , composite time-frequency feature into a joint feature vector: ; (7) inputting the feature tensor F in the time-frequency feature extraction branch of the multi-modal multi-scale attention fusion deep learning model, and outputting multi-modal features ; inputting the feature tensor F in the time-frequency feature extraction branch of the multi-modal multi-scale attention fusion deep learning model, and outputting multi-modal features ; inputting the feature tensor F in the time-frequency feature extraction branch of the multi-modal multi-scale attention fusion deep learning model, and outputting multi-modal features , , ; inputting the feature tensor F in the time-frequency feature extraction branch of the multi-modal multi-scale attention fusion deep learning model, and outputting multi-modal features ; (8) The multimodal features , The input is fed into the adaptive feature weighting fusion module of the deep learning model, and then adaptive weights are obtained. After weighted summation and fusion, the fusion characteristics are obtained. ; After being processed by the classification head, high-precision discrimination of different defect categories is realized, and the identification and detection of defects in the detected object are realized.

2. The method according to claim 1, wherein, In step (2), the energy accumulation distribution function of the echo signal is calculated using the following equation: wherein is the time, is the instantaneous amplitude of the original echo signal at time t.​ Optionally, in step (2), by obtaining the The first derivative curve automatically identifies the moment when the energy growth rate drops to a set threshold. This moment is taken as the termination point of the main reflected energy; the effective analysis time window of the signal is thus dynamically determined. The This is the margin parameter.

3. The method of claim 1, wherein the method is characterized by, In step (3), the joint denoising comprises the following steps: S3-1: Empirical Mode Decomposition: This involves converting the original echo signal... The signal is decomposed into several intrinsic mode functions, and the IMF1 containing high-frequency noise is removed to obtain the reconstructed signal of the intrinsic mode function components; S3-2: Singular spectrum analysis: reconstruct the signal of the intrinsic mode function component, based on the eigenvalue decomposition of the trajectory matrix, retain the first k mode components corresponding to the main singular values, and obtain a smooth reconstructed signal; S3-3: Wavelet threshold denoising: the smoothed reconstructed signal is adopted soft threshold Further weaken the random high frequency interference in the frequency domain; wherein, σ is the noise standard deviation, N is the signal length.

4. The method of claim 1, wherein the method is characterized by, In step (4), the normalization and Hilbert transform are performed using the following equation: ; where, is the Hilbert transform operator, is the imaginary unit used to construct the imaginary part of a complex signal.

5. The method of claim 1, wherein the method is characterized by, In step (5), the transform coefficients of the continuous wavelet transform (CWT) are are: ; wherein denotes the complex conjugate of the mother wavelet, is a time variable; a b denote a scale parameter and a translation parameter, respectively; optionally, the a = 8 ~ 54, b = 5 ~ 20 μs.​ 6. The method of claim 1, wherein the method is characterized by, In step (5), the composite time-frequency feature is calculated using the following formula: wherein is a coefficient matrix calculated using a Morlet wavelet; CWT is a coefficient matrix calculated using a Mexican Hat wavelet; CWT , are weight coefficients, and a and b represent scale parameters and translation parameters, respectively.​​ Optionally, the energy distribution is determined based on the , , and .

7. The method of claim 1, wherein the method is characterized by, In step (7), the time-frequency feature extraction branch is based on a fusion convolutional neural network (CNN) structure, adopts multi-scale residual convolution, and introduces a time-frequency attention mechanism; the present application extracts feature information at different time-frequency resolutions through parallel multi-scale convolution kernels, and allocates adaptive weights in the two-dimensional time-frequency plane to highlight the time interval and frequency band sensitive to defects, thereby obtaining information-rich time-frequency weighted features; Optionally, in step (7), the time-domain feature extraction branch is obtained by using a long short-term memory network combined with a time sequence attention mechanism. Optionally, in step (7), further comprising: inputting the purified signal into a feature extraction branch of the multi-modal multi-scale attention fusion deep learning model, and outputting multi-modal features The amplitude spectrum obtained by the fast Fourier transform The frequency spectrum features input into the multi-modal multi-scale attention fusion deep learning model, and output multi-modal features ; preferably, the The multi-modal features and the , are input into the adaptive feature weighting fusion module to obtain the fusion features ; Optionally, the spectral feature extraction branch is obtained by using a lightweight convolutional neural network combined with a global average pooling method.

8. The method of claim 1, wherein the method is characterized by, In step (8), the fusion feature The calculation formula is: .

9. The method of claim 1, wherein the method is characterized by, In step (8), the classification head uses a Softmax activation function at the output end to realize high-precision discrimination of different defect categories.

10. The method of claim 1-9, wherein, In step (8), it also includes the step of calculating the variance of the prediction result by using the uncertainty estimation module in the multi-modal multi-scale attention fusion deep learning model, which is used as the uncertainty index of the model output; Optionally, in step (1), the preset defects in the model experiment are simulated by using a hollow acrylic box.

Images