Recycled concrete internal defect detection method and system based on electromagnetic wave propagation

Abstract

The invention provides a recycled concrete defect detection method and system based on electromagnetic wave signal processing, and the method comprises the steps: obtaining an initial signal data set through transmitting an electromagnetic wave with a preset frequency, separating high-frequency and low-frequency components through wavelet transformation, judging defects, extracting spatial features through employing an improved U-Net network structure, and obtaining the coordinates of defect positions. And defect type probability distribution is obtained by using a neural network classification head, a classification result set is obtained, and internal defect detection of the recycled concrete is realized, so that the problems of defect position accurate identification and type description are solved. According to the method, the accuracy and efficiency of quality evaluation of the recycled concrete are remarkably improved, reliable application of sustainable building materials is promoted, and intelligentization and non-destructiveness of defect detection are achieved.

Description

Technical Field

[0001] This invention belongs to the field of concrete defect detection technology, specifically relating to a method and system for detecting internal defects in recycled concrete based on electromagnetic wave propagation. Background Technology

[0002] In the field of building materials and structural safety, research on detection techniques for internal defects in concrete is of paramount importance. As a core material in building structures, the internal quality of concrete directly affects the safety and durability of the project. This is especially true when using recycled materials, where internal defects can pose a greater threat to overall performance. Accurately identifying these hidden problems has become a crucial issue in ensuring project quality.

[0003] Currently, although various methods exist for detecting internal defects in concrete, these technologies often face challenges in terms of adaptability and accuracy. Many traditional methods struggle to effectively distinguish different types of defects when dealing with complex material compositions, especially in recycled concrete, which exhibits diverse compositions and heterogeneous structures, often resulting in unreliable test results. This limitation makes it difficult to detect potential safety hazards in a timely manner in practical engineering projects, necessitating a more targeted technological approach to address this issue.

[0004] Focusing on the technical challenges, the complexity of recycled concrete introduces unpredictability into signal propagation paths. Electromagnetic waves, as a potential detection tool, are affected by internal defects such as cracks and pores, leading to complex changes in signal reflection and transmission characteristics. Behind these changes lies another crucial issue: how to accurately distinguish the specific location and shape of defects from these subtle differences in signals. For example, when inspecting a bridge column made of recycled concrete, the electromagnetic wave signal may become blurred due to the combined effects of tiny internal cracks and larger pores, making it difficult for technicians to determine the severity and specific distribution of defects, thus affecting subsequent maintenance decisions.

[0005] Therefore, how to accurately capture changes in electromagnetic wave signals in complex materials like recycled concrete, and how to accurately identify the type and location of internal defects through these changes, has become a critical problem that urgently needs to be solved. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method and system for detecting internal defects in recycled concrete based on electromagnetic wave propagation. The aim is to accurately capture changes in electromagnetic wave signals in this complex material, recycled concrete, and to accurately identify the type and location of internal defects through these changes.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for detecting internal defects in recycled concrete based on electromagnetic wave propagation, the method comprising: Electromagnetic waves of a preset frequency are emitted onto recycled concrete, and the reflected and transmitted electromagnetic wave signal data are acquired and preprocessed to obtain an initial signal dataset. For the initial signal dataset, wavelet transform is used to separate the frequency components and obtain the high-frequency and low-frequency signal components; Based on the high-frequency and low-frequency signal components, a B-scan image is generated to identify abnormal regions and obtain a subset of defect-related signals. Based on a subset of defect-related signals, an improved U-Net network structure is used to extract spatial features and obtain the defect location coordinates. For the coordinates of the defect location, a neural network classification head is used to determine the probability distribution of the defect type and obtain a classification result set, thereby realizing the detection of internal defects in recycled concrete.

[0008] Preferably, the method for preprocessing reflected and transmitted electromagnetic wave signal data includes: Time-domain filtering: Smoothing filtering methods are used to smooth electromagnetic wave signals and remove high-frequency noise from them; Frequency domain filtering: Suppresses non-target frequency band interference in electromagnetic wave signals through bandpass filtering; Gain compensation: Time-varying gain or exponential gain is used to compensate for the attenuation of electromagnetic wave signals with depth; Background removal: The average background subtraction method is used to eliminate the effects of electromagnetic wave transmitting antenna coupling and surface reflection.

[0009] Preferably, the method for generating B-scan images based on high-frequency and low-frequency signal components, identifying abnormal regions, and obtaining defect-related signal subsets includes: The high-frequency and low-frequency signal components are arranged according to the time-space dimension to form an initial signal matrix; Based on the propagation speed of electromagnetic waves in recycled concrete, the time axis is converted into the depth axis to establish a depth-space coordinate system; Based on the initial signal matrix, signal enhancement is performed on the high-frequency signal component and the low-frequency signal component respectively; In the depth-space coordinate system, a weighted fusion method is used to generate B-scan images from the enhanced high-frequency signal components and low-frequency signal components. An adaptive thresholding algorithm is used to separate abnormal regions in B-scan images; Based on the spatiotemporal coordinates of the abnormal region, the initial signal dataset is traced back, and the A-scan signal within the corresponding time window is extracted to obtain a subset of defect-related signals.

[0010] Preferably, the method for extracting spatial features and obtaining defect location coordinates based on a subset of defect-related signals using an improved U-Net network structure includes: Convert the signals in the defect-related signal subset into electromagnetic wave feature maps; The electromagnetic wave feature map is input into the improved U-Net network structure to extract spatial features and obtain the defect probability map; Based on the defect probability graph, the coordinates of the defect location are obtained by using the maximum connected component.

[0011] Preferred, improved U-Net network architectures include: The 2D convolutions in the classic U-Net network structure are replaced by RFD modules. The RFD modules are composed of shallow residual blocks, contrast channel attention, batch normalization and skip connections coupled together. Replace the MSE loss in the classic U-Net network structure with a hybrid MSE+SSIM loss.

[0012] Preferably, contrast channel attention includes: ; in, z c It is the first c Contrast information values ​​of individual electromagnetic wave feature maps H GC (·) represents the global comparison information evaluation function. x c For the first c Electromagnetic wave characteristic map H and W The height and width of the electromagnetic wave feature map, i , j () represents the position in the electromagnetic wave characteristic map. For the first c Location in the electromagnetic wave characteristic map ( i , j The pixel value of ).

[0013] The preferred loss function for the improved U-Net network structure is: ; ; ; Among them, MSE ( x , y ) and SSIM ( x , y ) are MSE loss and SSIM loss, respectively, Loss( x , y) represents the loss function for the improved U-Net network structure. x (i,j) For input data, y (i,j) To reconstruct the data, M , N The size of the data matrix; where, µ x and m y They are x and y The mean; s x and s y Represent x , y standard deviation s xy for x and y covariance; C 1. C 2. C 3 is a constant used to avoid system errors caused by a denominator of 0.

[0014] The present invention also provides a system for detecting internal defects in recycled concrete based on electromagnetic wave propagation. The system is used to implement the aforementioned method and includes: a signal acquisition module, a frequency separation module, a defect identification module, a location acquisition module, and a classification identification module. The signal acquisition module is used to emit electromagnetic waves of a preset frequency to the recycled concrete, acquire reflected and transmitted electromagnetic wave signal data, and perform preprocessing to obtain an initial signal dataset. The frequency separation module is used to separate the frequency components from the initial signal dataset using wavelet transform to obtain high-frequency and low-frequency signal components. The defect identification module is used to generate B-scan images based on high-frequency and low-frequency signal components, identify abnormal areas, and obtain defect-related signal subsets. The location acquisition module is used to extract spatial features based on a subset of defect-related signals and employ an improved U-Net network structure to obtain the defect location coordinates. The classification and recognition module is used to determine the probability distribution of defect types by using a neural network classification head to identify the coordinates of the defect location, obtain a classification result set, and realize the detection of internal defects in recycled concrete.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses a method and system for detecting defects in recycled concrete based on electromagnetic wave signal processing. It addresses the challenge of accurately identifying and locating internal defects such as cracks and pores in recycled concrete, a problem involving the logical correlation between extracting internal interference features, defect morphology distribution, and type classification. This invention acquires an initial signal dataset by emitting electromagnetic waves at a preset frequency, uses wavelet transform to separate high-frequency and low-frequency components to identify defects, then applies an improved U-Net network structure to extract spatial features and obtain defect location coordinates. Finally, it uses a neural network classification head to obtain the probability distribution of defect types, resulting in a classification set. This enables the detection of internal defects in recycled concrete, thus solving the difficulties of accurate defect location identification and type description. This invention significantly improves the accuracy and efficiency of recycled concrete quality assessment, promotes the reliable application of sustainable building materials, and achieves intelligent and non-destructive defect detection. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method for detecting internal defects in recycled concrete based on electromagnetic wave propagation, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of an internal defect detection system for recycled concrete based on electromagnetic wave propagation, according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Example 1 like Figure 1 As shown, this invention provides a method for detecting internal defects in recycled concrete based on electromagnetic wave propagation, comprising: Electromagnetic waves of a preset frequency are emitted onto recycled concrete, and the reflected and transmitted electromagnetic wave signal data are acquired and preprocessed to obtain an initial signal dataset. For the initial signal dataset, wavelet transform is used to separate the frequency components and obtain the high-frequency and low-frequency signal components; Based on the high-frequency and low-frequency signal components, a B-scan image is generated to identify abnormal regions and obtain a subset of defect-related signals. Based on a subset of defect-related signals, an improved U-Net network structure is used to extract spatial features and obtain the defect location coordinates. For the coordinates of the defect location, a neural network classification head is used to determine the probability distribution of the defect type and obtain a classification result set, thereby realizing the detection of internal defects in recycled concrete.

[0021] The specific implementation process of this invention is as follows: S1. Emit electromagnetic waves of a preset frequency to the recycled concrete, acquire and preprocess the reflected and transmitted electromagnetic wave signal data to obtain an initial signal dataset, specifically including: When transmitting electromagnetic waves of a preset frequency to recycled concrete to obtain reflected and transmitted signals, electromagnetic waves with a frequency of 2.4 GHz can be selected as the detection method. Electromagnetic waves of this frequency have good penetration and resolution in recycled concrete materials and can effectively capture minute structural changes inside the material.

[0022] Based on the material properties of recycled concrete (such as aggregate dielectric constant, moisture content, porosity, etc.), a preset frequency electromagnetic wave (typically a pulsed wave or continuous wave in the range of 0.5–3 GHz) is selected. Low-frequency electromagnetic waves (such as 0.5–1 GHz) have strong penetrating power and are suitable for the inspection of large-volume concrete structures, but have lower resolution; high-frequency electromagnetic waves (such as 2–3 GHz) can provide more precise defect identification, but are easily affected by material attenuation. Electromagnetic wave transmission can be achieved using ultra-wideband antennas or stepped-frequency continuous wave (SFCW) radar to ensure that the signal can effectively penetrate the concrete and obtain sufficient reflection and transmission data.

[0023] The reflected and transmitted signals are captured by the receiving antenna. The original signal is susceptible to noise interference (such as system noise, multiple reflections, scattering, etc.), so preprocessing is required. Time-domain filtering: Smoothing filtering methods (such as moving average filtering or Savitzky-Golay filtering) are used to smooth electromagnetic wave signals and remove high-frequency noise from the electromagnetic wave signals; Frequency domain filtering: Suppressing non-target frequency band interference in electromagnetic signals through bandpass filtering (such as Butterworth filters); Gain compensation: Time-varying gain (TVG) or exponential gain is applied to compensate for the attenuation of electromagnetic wave signals with depth; Background removal: The average background subtraction method is used to eliminate the effects of electromagnetic wave transmitting antenna coupling and surface reflection.

[0024] The preprocessed signals constitute the initial signal dataset, which contains information such as time series, amplitude, and phase, providing high-quality input for subsequent wavelet transform and defect identification.

[0025] S2. For the initial signal dataset, wavelet transform is used to separate the frequency components, obtaining high-frequency and low-frequency signal components, specifically including: Based on the time-frequency characteristics of the recycled concrete signal in the initial signal dataset, an asymmetric, tightly supported wavelet basis function is selected to perform multi-level decomposition of the signal in the initial signal dataset, typically 5-7 levels, to ensure that low-frequency components cover the concrete thickness while retaining high-frequency information related to defects. Simultaneously, considering the heterogeneity of recycled concrete (such as differences in the dielectric constant of old aggregates), the number of wavelet decomposition levels is dynamically adjusted to avoid over-decomposition leading to feature loss.

[0026] Ultimately, high-frequency and low-frequency signal components are obtained, providing input for subsequent B-scan imaging and deep learning.

[0027] S3. Based on the high-frequency and low-frequency signal components, generate a B-scan image, identify abnormal regions, and obtain a subset of defect-related signals, specifically including: The high-frequency signal components (defect features) and low-frequency signal components (basis response) obtained by wavelet transform are converted into B-scan images, which visually present the internal structure of recycled concrete and locate defects. This process includes key steps such as signal enhancement, image synthesis, anomaly detection, and defect signal extraction, and is implemented as follows: The high-frequency and low-frequency signal components obtained by wavelet transform are arranged according to the time-space dimension to form an initial signal matrix. This step transforms the discrete time-domain signal into structured data, providing an input framework for B-scan image generation.

[0028] Based on the propagation speed of electromagnetic waves in recycled concrete, the time axis is converted into the depth axis to establish a depth-space coordinate system. This step converts the electromagnetic wave propagation time into the actual depth, thereby realizing the physical spatial location of defects.

[0029] Based on the initial signal matrix, signal enhancement is performed on both the high-frequency and low-frequency signal components: Hilbert transform is applied to the high-frequency signal component to extract its envelope, highlighting the transient reflection characteristics of the defect; logarithmic energy compression (e.g., 10^6) is applied to the low-frequency signal component. log10(·)) enhances the contrast of the substrate material.

[0030] In the depth-space coordinate system, the high-frequency signal components and low-frequency signal components after signal enhancement are weighted and fused to generate B-scan images. B-scan images take into account both defect sensitivity and structural background information.

[0031] An adaptive threshold algorithm is used to separate abnormal regions in B-scan images, and the number, area, and spatial distribution of abnormal regions are statistically analyzed through connected component analysis to achieve abnormal region detection.

[0032] Based on the spatiotemporal coordinates of the abnormal regions in the B-scan image, the initial signal dataset is traced back to extract the A-scan signal within the corresponding time window. Then, power spectral density (PSD) analysis is performed on the extracted A-scan signal to retain the energy peaks of the defect characteristic frequency bands. Principal component analysis (PCA) is used to reduce dimensionality and remove redundant frequency points. After the above processing, a subset of defect-related signals is obtained, including: signal amplitude sequence, instantaneous frequency, temporal entropy, defect center coordinates, equivalent diameter, mean reflection intensity, etc., which provides standardized input for the subsequent U-Net segmentation extraction and classification network.

[0033] S4. Based on a subset of defect-related signals, an improved U-Net network structure is used to extract spatial features and obtain the defect location coordinates. Specifically, this includes: Convert the signals in the defect-related signal subset into electromagnetic wave feature maps; The electromagnetic wave feature map is input into the improved U-Net network structure to extract spatial features and obtain the defect probability map; Based on the defect probability graph, the coordinates of the defect location are obtained by using the maximum connected component.

[0034] U-Net was originally designed for medical image segmentation. Subsequently, its "encoder-decoder" concept was adopted by various fields, leading to a series of lightweight versions for deployment in resource-constrained scenarios. The standard U-Net's encoding branch uses "convolution + max pooling + ReLU" modules alternately layer by layer to extract multi-scale features; the decoding branch symmetrically upsamples and concatenates features of the same resolution through skip connections, finally outputting a segmentation map through convolution and activation.

[0035] This invention makes improvements based on the classic U-Net network structure: 1. Replace all traditional 2D convolutions with the RFD module. The RFD module is composed of shallow residual blocks (SRB), contrast channel attention (CCA), batch normalization (BN), and skip connections (SC). Among them, SRB enhances the representation of residual features, CCA reconstructs high-fidelity details, BN accelerates convergence and suppresses gradient vanishing, and SC compensates for information loss during convolution.

[0036] Specifically, the RFD module focuses on the extraction of residual features of the effective signal: its SRB is composed of "residual branch (including weight layer) + identity branch" connected in parallel and then followed by activation unit, which significantly enhances parameter efficiency and characterization ability.

[0037] CCA is used to adjust the weights of each channel, converting the signals in the defect-related signal subset into electromagnetic wave feature maps with a spatial size of [missing information]. H × W , using c Data of electromagnetic wave characteristic maps X =[ x 1, x 2, ..., x c As the CCA input, the contrast information value is calculated using the following formula: ; in, z c It is the first c Contrast information values ​​of individual electromagnetic wave feature maps H GC (·) represents the global comparison information evaluation function. H and W The height and width of the electromagnetic wave feature map, i , j () represents the position in the electromagnetic wave characteristic map. For the first c Location in the electromagnetic wave characteristic map ( i , j The pixel value of ).

[0038] Superimposing CCA with the original input of RFD as the final output of RFD can stably enhance the structural and feature information of data details.

[0039] Instead of the ReLU activation function, use PReLU. The PReLU activation function is: ; in, x For input data, when learnable parameters α i When the value is 0, the PReLU activation function degenerates into the ReLU activation function. Using the PReLU activation function can effectively avoid gradient vanishing.

[0040] By combining the advantages of network structures such as RFD, CCA, and SRB, an improved U-Net network structure was finally obtained.

[0041] 2. Upgrade the single MSE loss to a hybrid MSE+SSIM loss, which balances accuracy and structural similarity, thereby improving both accuracy and effective signal continuity in spatial feature extraction.

[0042] The formula for calculating MSE is: ; in,x (i,j) For input data, y (i,j) To reconstruct the data, M , N This represents the size of the data matrix.

[0043] The formula for calculating SSIM is: ; in, µ x and m y They are x and y The mean; s x and s y Represent x , y standard deviation s xy for x and y covariance; C 1. C 2. C 3 is a constant used to avoid system errors caused by a denominator of 0.

[0044] The loss function of the improved U-Net network structure is: .

[0045] S5. For the defect location coordinates, a neural network classification head is used to determine the probability distribution of the defect type, obtain the classification result set, and realize the detection of internal defects in recycled concrete, specifically including: After the initial location of the defect is determined, a deep learning-based neural network classification head is used to perform refined identification and probabilistic classification of the defect type, obtain a classification result set, and realize the detection of internal defects in recycled concrete.

[0046] Specifically, the process begins by obtaining a subset of signals containing spatial features of the defects based on their location coordinates. This subset is then input into a multi-layered neural network classification head. This classification head typically uses a Global Average Pooling (GAP) layer to compress the spatial dimension, followed by 1-2 fully connected layers for feature mapping. Finally, a Softmax activation function is used to output the probability distribution of each defect type. For the multimodal characteristics of internal defects in recycled concrete (such as cracks, pores, and delamination), an attention-enhanced dual-branch classification head can be employed: one branch focuses on high-frequency scattering features at the defect edges using a self-attention mechanism (suitable for crack identification), while the other branch expands the receptive field through dilated convolutions to capture pore shadow features in low-frequency transmission signals. During the training phase, the classification head uses Focal Loss to alleviate class imbalance and introduces a transfer learning strategy, improving generalization ability in small-sample scenarios through a pre-trained physical feature encoder on a synthetic dataset. The final output classification result set contains the defect type probability vector and its confidence score at each location coordinate.

[0047] In summary, this invention provides a method for detecting defects in recycled concrete based on electromagnetic wave signal processing. It addresses the challenge of accurately identifying and locating internal defects such as cracks and pores in recycled concrete, a problem involving the logical correlation between extracting internal interference features, defect morphology distribution, and type classification. This invention acquires an initial signal dataset by emitting electromagnetic waves at a preset frequency, uses wavelet transform to separate high-frequency and low-frequency components to identify defects, then applies an improved U-Net network structure to extract spatial features and obtain defect location coordinates. Finally, it utilizes a neural network classification head to obtain the probability distribution of defect types, resulting in a classification set. This enables the detection of internal defects in recycled concrete, thus solving the difficulties of accurate defect location identification and type description. This invention significantly improves the accuracy and efficiency of recycled concrete quality assessment, promotes the reliable application of sustainable building materials, and achieves intelligent and non-destructive defect detection.

[0048] Example 2 like Figure 2 As shown, based on the same inventive concept, the present invention also provides an electromagnetic wave propagation-based internal defect detection system for recycled concrete, used to implement the method described in the foregoing embodiments. The system includes: a signal acquisition module, a frequency separation module, a defect identification module, a location acquisition module, and a classification identification module. The signal acquisition module is used to emit electromagnetic waves of a preset frequency to the recycled concrete, acquire reflected and transmitted electromagnetic wave signal data, and perform preprocessing to obtain an initial signal dataset. The frequency separation module is used to separate the frequency components from the initial signal dataset using wavelet transform to obtain high-frequency and low-frequency signal components. The defect identification module is used to generate B-scan images based on high-frequency and low-frequency signal components, identify abnormal areas, and obtain defect-related signal subsets. The location acquisition module is used to extract spatial features based on a subset of defect-related signals and employ an improved U-Net network structure to obtain the defect location coordinates. The classification and recognition module is used to determine the probability distribution of defect types by using a neural network classification head to identify the coordinates of the defect location, obtain a classification result set, and realize the detection of internal defects in recycled concrete.

[0049] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for detecting internal defects in recycled concrete based on electromagnetic wave propagation, characterized in that, The method includes: Electromagnetic waves of a preset frequency are emitted onto recycled concrete, and the reflected and transmitted electromagnetic wave signal data are acquired and preprocessed to obtain an initial signal dataset. For the initial signal dataset, wavelet transform is used to separate the frequency components and obtain the high-frequency and low-frequency signal components; Based on the high-frequency and low-frequency signal components, a B-scan image is generated to identify abnormal regions and obtain a subset of defect-related signals. Based on a subset of defect-related signals, an improved U-Net network structure is used to extract spatial features and obtain the defect location coordinates. For the coordinates of the defect location, a neural network classification head is used to determine the probability distribution of the defect type and obtain a classification result set, thereby realizing the detection of internal defects in recycled concrete.

2. The method according to claim 1, characterized in that, Methods for preprocessing reflected and transmitted electromagnetic wave signal data include: Time-domain filtering: Smoothing filtering methods are used to smooth electromagnetic wave signals and remove high-frequency noise from them; Frequency domain filtering: Suppresses non-target frequency band interference in electromagnetic wave signals through bandpass filtering; Gain compensation: Time-varying gain or exponential gain is used to compensate for the attenuation of electromagnetic wave signals with depth; Background removal: The average background subtraction method is used to eliminate the effects of electromagnetic wave transmitting antenna coupling and surface reflection.

3. The method according to claim 1, characterized in that, Methods for generating B-scan images based on high-frequency and low-frequency signal components, identifying abnormal regions, and obtaining defect-related signal subsets include: The high-frequency and low-frequency signal components are arranged according to the time-space dimension to form an initial signal matrix; Based on the propagation speed of electromagnetic waves in recycled concrete, the time axis is converted into the depth axis to establish a depth-space coordinate system; Based on the initial signal matrix, signal enhancement is performed on the high-frequency signal component and the low-frequency signal component respectively; In the depth-space coordinate system, a weighted fusion method is used to generate B-scan images from the enhanced high-frequency signal components and low-frequency signal components. An adaptive thresholding algorithm is used to separate abnormal regions in B-scan images; Based on the spatiotemporal coordinates of the abnormal region, the initial signal dataset is traced back, and the A-scan signal within the corresponding time window is extracted to obtain a subset of defect-related signals.

4. The method according to claim 1, characterized in that, Based on a subset of defect-related signals, a method for extracting spatial features and obtaining defect location coordinates using an improved U-Net network structure includes: Convert the signals in the defect-related signal subset into electromagnetic wave feature maps; The electromagnetic wave feature map is input into the improved U-Net network structure to extract spatial features and obtain the defect probability map; Based on the defect probability graph, the coordinates of the defect location are obtained by using the maximum connected component.

5. The method according to claim 4, characterized in that, The improved U-Net network architecture includes: The 2D convolutions in the classic U-Net network structure are replaced by RFD modules. The RFD modules are composed of shallow residual blocks, contrast channel attention, batch normalization and skip connections coupled together. Replace the MSE loss in the classic U-Net network structure with a hybrid MSE+SSIM loss.

6. The method according to claim 5, characterized in that, Contrast channel attention includes: ； in, z c It is the first c Contrast information values ​​of individual electromagnetic wave feature maps H GC (·) represents the global comparison information evaluation function. x c For the first c Electromagnetic wave characteristic map H and W The height and width of the electromagnetic wave feature map, i , j () represents the position in the electromagnetic wave characteristic map. For the first c Location in the electromagnetic wave characteristic map ( i , j The pixel value of ).

7. The method according to claim 6, characterized in that, The loss function of the improved U-Net network structure is: ； ； ； Among them, MSE ( x , y ) and SSIM ( x , y ) are MSE loss and SSIM loss, respectively, Loss( x , y ) represents the loss function for the improved U-Net network structure. x (i,j) For input data, y (i,j) To reconstruct the data, M , N The size of the data matrix; where, µ x and μ y They are x and y The mean; σ x and σ y Represent x , y standard deviation σ xy for x and y covariance; C 1. C 2. C 3 is a constant used to avoid system errors caused by a denominator of 0.

8. A system for detecting internal defects in recycled concrete based on electromagnetic wave propagation, said system being used to implement the method described in any one of claims 1-7, characterized in that, The system includes: a signal acquisition module, a frequency separation module, a defect identification module, a location acquisition module, and a classification identification module; The signal acquisition module is used to emit electromagnetic waves of a preset frequency to the recycled concrete, acquire reflected and transmitted electromagnetic wave signal data, and perform preprocessing to obtain an initial signal dataset. The frequency separation module is used to separate the frequency components from the initial signal dataset using wavelet transform to obtain high-frequency and low-frequency signal components. The defect identification module is used to generate B-scan images based on high-frequency and low-frequency signal components, identify abnormal areas, and obtain defect-related signal subsets. The location acquisition module is used to extract spatial features based on a subset of defect-related signals and employ an improved U-Net network structure to obtain the defect location coordinates. The classification and recognition module is used to determine the probability distribution of defect types by using a neural network classification head to identify the coordinates of the defect location, obtain a classification result set, and realize the detection of internal defects in recycled concrete.

Images