Nondestructive detection system and method for defects of concrete structure

Abstract

The invention relates to the technical field of concrete, in particular to a nondestructive testing system and method for defects of a concrete structure. The method comprises the following steps: firstly, arranging an ultrasonic detector on concrete needing to be detected; then, transmitting an ultrasonic signal, receiving an echo signal transmitted by the concrete, taking the echo signal as original signal data, and collecting multiple groups of original signal data; and then, carrying out data processing on the original signal data. Then, a defect real-time identification model MP-FA-GRU is constructed, and the defect real-time identification model MP-FA-GRU comprises a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module comprises three feature extraction paths; the multi-path feature fusion module comprises a frequency domain attention module and a gating loop module. And finally, carrying out performance test on the defect real-time identification model MP-FA-GRU by utilizing the test set, and verifying the validity. The model used by the method has excellent robustness, and the accuracy of concrete defect detection can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of concrete, and in particular to a nondestructive detection system and method for concrete structure defects. Background Art

[0002] Concrete structures are widely used in key projects such as bridges, high-rise buildings, and tunnels. When substandard concrete or aggregates are used, or the mix ratio is inaccurate, it will lead to insufficient concrete strength or poor durability. In addition, when concrete is exposed to harsh environments for a long time, such as humidity, extreme temperature changes, and chemical corrosion, it will accelerate the deterioration of concrete, causing cracks and other defects. Other reasons for concrete defects include cracks and deformation when the concrete structure is subjected to actual loads that exceed the design load.

[0003] Nondestructive testing can detect internal defects and deterioration without destroying the structure, preventing potential structural failure and thus protecting public safety. By detecting and identifying defects in concrete at an early stage, repairs and reinforcements can be carried out in time to prevent problems from worsening, extending the service life of the structure and reducing long-term maintenance and repair costs.

[0004] Nondestructive testing technologies include acoustic emission technology, parameter analysis methods such as amplitude and ring count, and frequency domain analysis methods. Acoustic emission technology is a method based on monitoring the acoustic wave signals generated by the release of internal stress in the material. This technology can capture relevant signals when tiny damage occurs inside the concrete. Although this technology can provide detailed information about the internal defects of the structure, it faces difficulties in data processing in practical applications. Due to the complexity of the internal destruction process of concrete structures, the acoustic emission signals generated are usually very large and dense, which brings great challenges to data processing and analysis. The principles of parameter analysis methods such as amplitude and ring count are relatively simple, easy to implement and operate, but there are also some obvious shortcomings in practical applications. Since these methods mainly rely on simple parameters, they often cannot fully reflect the true damage state of concrete structures. They are easily disturbed by the external environment, such as temperature changes, mechanical vibrations, etc., and these interference factors may lead to inaccurate data. The frequency domain analysis method uses the spectral characteristics of the signal to detect defects in concrete structures. This method has strong noise resistance and can perform effective detection under more complex environmental conditions. However, frequency domain analysis also has its own limitations. For example, the frequency domain analysis method has poor consistency in results, which means that the test results of the same structure at different times or under different environmental conditions may be different, resulting in wrong judgments. In summary, although the existing non-destructive testing technologies for concrete structures have their own advantages, they also generally have some problems. The main problem is that the accuracy and real-time performance of the existing technologies are not high enough.

[0005] Therefore, a nondestructive detection system and method for concrete structure defects are proposed. Summary of the invention

[0006] The object of the present invention is to provide a nondestructive detection system and method for concrete structure defects, so as to improve the detection accuracy and efficiency of concrete structure defects. First, an ultrasonic detector is arranged on the concrete to be inspected, covering the concrete area with defects; then, an ultrasonic signal is emitted, and an echo signal propagated by the concrete is received, and the echo signal is used as the original signal data to collect multiple groups of original signal data; wherein, the original signal data includes data of known defect types and data of unknown defect types, and the data of the known defect types are constructed as a training data set, and are divided into a training set, a validation set and a test set in proportion, and the data of the unknown defect types are constructed as a defect data set to be identified; then, the original signal data is processed, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data; then, a real-time defect recognition model MP-FA-GRU is constructed, and the real-time defect recognition model MP-FA-GRU includes a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes 3 feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module GRU; finally, the real-time defect recognition model MP-FA-GRU is performance tested using the test set to verify its effectiveness.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A nondestructive detection system for concrete structure defects, comprising:

[0009] A deployment unit deploys ultrasonic detectors on the concrete to be tested to ensure that the defective concrete area is covered;

[0010] A data acquisition unit transmits an ultrasonic signal, receives an echo signal propagated by the concrete, and collects multiple sets of original signal data as original signal data; wherein the original signal data includes data of known defect types and data of unknown defect types, the data of known defect types are constructed as a training data set, and are divided into a training set, a verification set, and a test set in proportion, and the data of unknown defect types are constructed as a defect data set to be identified;

[0011] a data processing unit, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data;

[0012] The MP-FA-GRU construction unit of the defect real-time recognition model includes a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes three feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module GRU;

[0013] The testing unit uses the test set to perform performance testing on the real-time defect recognition model MP-FA-GRU to verify its effectiveness.

[0014] Furthermore, in the data processing unit, an adaptive filter is used to perform preliminary denoising on the multiple groups of original signal data to reduce the impact of noise based on real-time signal characteristics; a wavelet transform is used to perform multi-scale analysis on the multiple groups of original signal data to capture features at different scales in the multiple groups of original signal data, help identify and separate clutter, and further reduce noise in the multiple groups of original signal data; and the normalization normalizes the denoised signal data to a uniform range.

[0015] Furthermore, in the data processing unit, the specific implementation method of weighted fusion of denoised data is:

[0016] Assume that the original signal data is y[n]=x[n]+d[n]; where x[n] is the actual signal and d[n] is the noise;

[0017] Use an adaptive filter to perform preliminary denoising on y[n], and get in, is the noise estimated by the adaptive filter;

[0018] At the same time, wavelet transform is used to obtain y[n] 2 [n] = W -1 (max(W[y[n]])-λ,0)·sign(W[y[n]]); where W is the wavelet transform; λ is the threshold;

[0019] Then, y 1 [n] and y 2 [n] weighted fusion, we get:

[0020]

[0021] Among them, f 1 and f 2 They are respectively 1 [n] and y 2 The function that models the features of [n].

[0022] Furthermore, in the defect real-time identification model MP-FA-GRU construction unit, the input of the defect real-time identification model MP-FA-GRU is the result obtained by preliminary denoising, the result obtained by further denoising and the result of weighted fusion of the two; the output is the defect type of the concrete structure; after reading the input data, the defect real-time identification model MP-FA-GRU captures the global features, detail features and complementary features of each path feature through the multi-path feature extraction module; then, through the multi-path feature fusion module, the results of the three path outputs in the multi-path feature extraction module are fused; finally, the defect type is output through the output layer;

[0023] The multi-path feature extraction module includes three paths, the first path reads the result obtained by preliminary denoising, the second path reads the result obtained by further denoising, and the third path reads the result of weighted fusion of the two; the first path and the second path have the same structure, firstly, extract features through convolution blocks; then perform nonlinear mapping through tanh activation function; then, use the maximum pooling layer to capture significant features, and the input of the first path is connected to the maximum pooling layer through a residual connection layer; then, pass through a convolution block and tanh activation function, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, perform feature normalization through a BatchNorm layer;

[0024] The structure of the third path is similar to that of the first and second paths. First, features are extracted through a convolution block; then nonlinear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features; then, another convolution block and a tanh activation function are passed through, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, features are normalized through a BatchNorm layer.

[0025] Furthermore, the multipath feature fusion module includes a frequency domain attention module and a gated loop module;

[0026] In the frequency domain attention module, the output features of the first path, the second path and the third path are first concatenated; then, the prominent features in the initial frequency domain features are captured through a linear activation function layer, a tanh activation function layer and another linear activation function layer to obtain frequency domain features; then, the frequency domain features are converted into weight distribution vectors through a softmax layer, and are element-by-element multiplied with the initial frequency domain features; then, the multiplied result is subjected to an inverse fast Fourier transform IFFT to obtain time domain features;

[0027] In the gated recurrent module, the time domain features are passed through three gated recurrent units GRU to obtain the time dependency in the time domain features.

[0028] A nondestructive detection method for concrete structure defects, comprising:

[0029] The ultrasonic detector is placed on the concrete to be inspected, covering the concrete area with defects;

[0030] Transmitting an ultrasonic signal, receiving an echo signal propagated by concrete, taking the echo signal as original signal data, and collecting multiple sets of original signal data; wherein the original signal data includes data of known defect types and data of unknown defect types, constructing the data of known defect types as a training data set, and dividing it into a training set, a verification set, and a test set in proportion, and constructing the data of unknown defect types as a defect data set to be identified;

[0031] Performing data processing on the raw signal data, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data;

[0032] Constructing a real-time defect recognition model MP-FA-GRU, the real-time defect recognition model MP-FA-GRU includes a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes three feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module GRU;

[0033] The test set is used to perform performance testing on the real-time defect recognition model MP-FA-GRU to verify its effectiveness.

[0034] Furthermore, an adaptive filter is used to perform preliminary denoising on the multiple groups of original signal data to reduce the impact of noise based on real-time signal characteristics; a wavelet transform is used to perform multi-scale analysis on the multiple groups of original signal data to capture features at different scales in the multiple groups of original signal data, help identify and separate clutter, and further reduce noise in the multiple groups of original signal data; and the normalization normalizes the denoised signal data to a uniform range.

[0035] Furthermore, the specific implementation method of weighted fusion of denoised data is:

[0036] Assuming the original signal data is y[n], use an adaptive filter to perform preliminary denoising on y[n] to obtain y 1 [n], use wavelet transform to get y[n] 2 [n]; then, y 1 [n] and y 2 [n] weighted fusion, we get:

[0037]

[0038] Among them, f 1 and f 2 They are respectively 1 [n] and y 2 The function that models the features of [n].

[0039] Furthermore, the input of the real-time defect recognition model MP-FA-GRU is the result obtained by preliminary denoising, the result obtained by further denoising and the result of weighted fusion of the two; the output is the defect type of the concrete structure; after reading the input data, the real-time defect recognition model MP-FA-GRU captures the global features, detail features and complementary features of each path feature through the multi-path feature extraction module; then, through the multi-path feature fusion module, the results of the three path outputs in the multi-path feature extraction module are fused; finally, the defect type is output through the output layer;

[0040] The multi-path feature extraction module includes three paths, the first path reads the result obtained by preliminary denoising, the second path reads the result obtained by further denoising, and the third path reads the result of weighted fusion of the two; the first path and the second path have the same structure, firstly, extract features through convolution blocks; then perform nonlinear mapping through tanh activation function; then, use the maximum pooling layer to capture significant features, and the input of the first path is connected to the maximum pooling layer through a residual connection layer; then, pass through a convolution block and tanh activation function, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, perform feature normalization through a BatchNorm layer;

[0041] The structure of the third path is similar to that of the first and second paths. First, features are extracted through a convolution block; then nonlinear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features; then, another convolution block and a tanh activation function are passed through, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, features are normalized through a BatchNorm layer.

[0042] Furthermore, the multipath feature fusion module includes a frequency domain attention module and a gated loop module;

[0043] In the frequency domain attention module, the output features of the first path, the second path and the third path are first concatenated; then, the prominent features in the initial frequency domain features are captured through a linear activation function layer, a tanh activation function layer and another linear activation function layer to obtain frequency domain features; then, the frequency domain features are converted into weight distribution vectors through a softmax layer, and are element-by-element multiplied with the initial frequency domain features; then, the multiplied result is subjected to an inverse fast Fourier transform IFFT to obtain time domain features;

[0044] In the gated recurrent module, the time domain features are passed through three gated recurrent units GRU to obtain the time dependency in the time domain features.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention uses adaptive filter and wavelet transform for denoising, and fuses the results of the two denoising. The adaptive filter can effectively smooth the noise and adapt to the noise characteristics, while the wavelet transform can better capture details and local features. Weighted fusion can provide more comprehensive signal features, thereby improving the accuracy of defect detection and reducing errors caused by the insufficiency of a single denoising method.

[0047] 2. The present invention uses a multi-path feature extraction module, which can fully capture the defect characteristics of concrete from different feature perspectives by processing the results of preliminary denoising, further denoising and weighted fusion respectively. The first and second paths extract features at different denoising stages, which helps to retain more detailed information; the third path combines the advantages of the two. The final feature fusion improves the accuracy and robustness of defect detection.

[0048] 3. The present invention uses a multi-path feature fusion module, which can improve the accuracy of defect detection and the ability to capture details through the combination of the frequency domain attention module and the gated loop module. The frequency domain attention module enhances the focus on frequency domain features through FFT and IFFT processing, while the gated loop module uses GRU to capture the time dependency of time domain features, thereby better understanding the dynamic changes of the signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A system structure diagram of a nondestructive detection system for concrete structure defects of the present invention;

[0050] Figure 2 It is a network structure diagram of the real-time defect recognition model MP-FA-GRU of the present invention;

[0051] Figure 3 Schematic diagram of the structure of the gated recurrent unit GRU of the present invention;

[0052] Figure 4 The present invention is a method flow chart of a method for nondestructive detection of concrete structure defects. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] In order to improve the detection accuracy and efficiency of concrete structure defects, the present invention provides a nondestructive detection system and method for concrete structure defects. In order to illustrate the function of the present invention, the effectiveness of the present invention will be described from the following examples.

[0055] Embodiment 1

[0056] In order to accurately detect the types of structural defects in concrete roads, a road maintenance team used a concrete structural defect non-destructive detection system. Figure 1 , which is a system structure diagram of a nondestructive detection system for concrete structure defects.

[0057] A nondestructive detection system for concrete structure defects includes a deployment unit, which deploys ultrasonic detectors on concrete to be tested to ensure that the defective concrete area is covered. The deployment method uses a flat measurement method.

[0058] A nondestructive detection system for concrete structural defects also includes a data acquisition unit, which transmits ultrasonic signals, receives echo signals propagated by concrete, and collects multiple groups of original signal data as original signal data. For concrete with internal structural defects, the high-frequency component in its echo signal is relatively reduced, while the low-frequency component is relatively increased, and the main frequency value of the received wave decreases, that is, its acoustic parameter value is significantly different from the acoustic parameter value of normal concrete without internal structural defects.

[0059] Among them, the original signal data includes data of known defect types and data of unknown defect types. The data of known defect types are constructed as a training data set and divided into a training set, a verification set and a test set in a ratio of 7:2:1. The data of unknown defect types are constructed as a defect data set to be identified.

[0060] A nondestructive detection system for concrete structure defects also includes a data processing unit, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data.

[0061] Among them, an adaptive filter is used to perform preliminary denoising processing on the multiple groups of original signal data to reduce the impact of noise according to real-time signal characteristics; a wavelet transform is used to perform multi-scale analysis on the multiple groups of original signal data to capture the characteristics of the multiple groups of original signal data at different scales, help identify and separate clutter, and further reduce the noise in the multiple groups of original signal data; the normalization normalizes the denoised signal data to a unified range, that is, within the interval of 0-1.

[0062] The data processing module achieves efficient noise suppression and feature extraction through adaptive filters and wavelet transforms. The final normalization step ensures the consistency and comparability of data processing and improves the accuracy of subsequent analysis and model training.

[0063] Furthermore, in the data processing unit, the specific implementation method of weighted fusion of denoised data is:

[0064] Assume that the original signal data is y[n]=x[n]+d[n]; where x[n] is the actual signal and d[n] is the noise;

[0065] Use an adaptive filter to perform preliminary denoising on y[n], and get in, is the noise estimated by the adaptive filter;

[0066] At the same time, wavelet transform is used to obtain y[n] 2 [n] = W -1 (max(W[y[n]])-λ,0)·sign(W[y[n]]); where W is the wavelet transform; λ is the threshold;

[0067] Then, y 1 [n] and y 2 [n] weighted fusion, we get:

[0068]

[0069] Among them, f 1 and f 2 They are respectively 1 [n] and y 2 The function that models the characteristics of [n]. 1 and f 2 It can be defined as a linear function, a square function, a combined characteristic function, and a nonlinear function.

[0070] If it is in the form of a linear function, it can be expressed as f 1 =a 1 y 1 [n]+b 1 y2 [n]+c 1 and f 2 =a 2 y 1 [n]+b 2 y 2 [n]+c 2 , where a 1 、a 2 、b 1 、b 2 、c 1 and c 2 are the parameters of the model;

[0071] If it is in the form of a square function, it can be expressed as where α 1 , α 2 , β 1 , β 2 , γ 1 and γ 2 is the parameter of the model; if it is in the form of a combined characteristic function, it can be expressed as f 1 =δ 1 y 1 [n]+ε 1 y 2 [n]+ζ 1 ·(y 1 [n]·y 2 [n]) and f 2 =δ 2 y 1 [n]+ε 2 y 2 [n]+ζ 2 ·(y 1 [n]·y 2 [n]), where δ 1 ,δ 2 , ε 1 , ε 2 , 1 and 2 are the parameters of the model;

[0072] If it is in the form of a nonlinear function, it can be expressed as and in ω 1 and ω 2 are the parameters of the model.

[0073] The weighted fusion operation combines the denoising results of adaptive filters and wavelet transforms to take advantage of both and balance their shortcomings. The adaptive filter effectively suppresses real-time noise, while the wavelet transform captures details at multiple scales. Weighted fusion helps to further reduce noise, retain important features, and enhance detection accuracy and robustness, thereby improving the overall performance of concrete defect detection.

[0074] A nondestructive detection system for concrete structure defects also includes a defect real-time identification model MP-FA-GRU construction unit, including a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes three feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module.

[0075] Among them, refer to Figure 2 , is the network structure diagram of the real-time defect recognition model MP-FA-GRU. The input of the real-time defect recognition model MP-FA-GRU is the result obtained by preliminary denoising, the result obtained by further denoising and the result of weighted fusion of the two; the output is the defect type of the concrete structure; after reading the input data, the real-time defect recognition model MP-FA-GRU uses the multi-path feature extraction module to capture the global features, detail features and complementary features of each path feature; then, the multi-path feature fusion module is used to fuse the results of the three path outputs in the multi-path feature extraction module; finally, the defect type is output through the output layer.

[0076] Furthermore, the multi-path feature extraction module includes three paths, the first path reads the result obtained by preliminary denoising, the second path reads the result obtained by further denoising, and the third path reads the result of weighted fusion of the two; the structure of the first path and the second path is the same, first, the feature is extracted by a convolution block; then, non-linear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features, and the input of the first path is connected to the maximum pooling layer through a residual connection layer; then, it passes through a convolution block and a tanh activation function, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, the feature is normalized through a BatchNorm layer; the structure of the third path is similar to that of the first path and the second path, first, the feature is extracted by a convolution block; then, non-linear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features; then, it passes through a convolution block and a tanh activation function, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, the feature is normalized through a BatchNorm layer. Combining these steps, three outputs of three paths are obtained.

[0077] In this embodiment, the processed data is one-dimensional time series data, so one-dimensional convolution is used; through the combination of convolution blocks, tanh activation functions and maximum pooling layers, the module not only enhances the expressive power of features, but also improves the training stability and generalization ability of the model through residual connections and BatchNorm layers.

[0078] The multi-path feature extraction module processes the denoising results and weighted fusion results through three paths respectively, effectively extracting multi-level features. The first and second paths extract key features from different denoising results, and the third path processes the fusion results, ensuring comprehensive feature capture. The combination of convolutional blocks, tanh activation functions, maximum pooling layers, and BatchNorm layers improves the feature extraction capability and normalization effect, making the final feature representation more accurate.

[0079] Furthermore, the multi-path feature fusion module includes a frequency domain attention module and a gated loop module.

[0080] In the frequency domain attention module, the output features of the first path, the second path and the third path are first concatenated; then, the features are subjected to fast Fourier transform FFT to obtain initial frequency domain features; then, the prominent features in the initial frequency domain features are captured through a linear activation function layer, a tanh activation function layer and another linear activation function layer to obtain frequency domain features; then, the frequency domain features are converted into weight distribution vectors through a softmax layer, and are element-wise multiplied with the initial frequency domain features; then, the multiplied result is subjected to an inverse fast Fourier transform IFFT to obtain time domain features.

[0081] Among them, the fast Fourier transform FFT is used for time-frequency domain conversion, making the frequency components in the time series signal easier to identify and analyze in the frequency domain. Certain defects or anomalies may cause obvious changes within a specific frequency range, and these changes are more prominent in the frequency domain characteristics.

[0082] In the gated recurrent module, the time domain features are passed through three gated recurrent units GRU to obtain the time dependency in the time domain features. The schematic diagram of the structure of the gated recurrent unit GRU is as follows: Figure 3 As shown in Figure 2. The gated recurrent unit GRU has two gates, namely the reset gate and the update gate. The reset gate determines how to combine new input information with previous memory, and the update gate defines the amount of previous memory saved to the current time step. These two gating mechanisms can preserve information in long-term sequences without being cleared over time or removed because they are not relevant to prediction.

[0083] Furthermore, the output layer includes a convolution block, a BatchNorm layer and a softmax layer, wherein the softmax layer is used to output the classification result.

[0084] The multi-path feature fusion module improves the accuracy and timeliness of feature fusion through frequency domain attention and gated recurrent units. The frequency domain attention module enhances the distinguishing ability of feature representation, ensures the prominence of important features, and reduces noise interference. The gated recurrent unit (GRU) captures the dependencies in the time series, maintains long-term memory, and helps improve the model's sensitivity to time changes. This combination improves the expressiveness of features and the overall performance of the model.

[0085] A nondestructive detection system for concrete structure defects also includes a testing unit, which uses the test set to perform performance testing on the real-time defect recognition model MP-FA-GRU to verify its effectiveness.

[0086] Specifically, the performance of the real-time defect recognition model MP-FA-GRU and the existing defect recognition models are compared on the test set. The existing defect recognition models used include GRU, Bi-LSTM and Attention-based RNNs. The results of the comparative experiment are shown in Table 1.

[0087] Table 1 Comparative experimental results

[0088] Network Model Recognition accuracy GRU 0.949 Bi-LSTM 0.947 Attention-based RNNs 0.941 MP-FA-GRU 0.956

[0089] As can be seen from Table 1, the real-time defect recognition model MP-FA-GRU used in this embodiment shows excellent defect classification performance in the test set, with an accuracy rate of 0.956, while the classification accuracy rates obtained by other existing defect recognition models based on the test machine are still lower than that of MP-FA-GRU.

[0090] The trained real-time defect recognition model MP-FA-GRU is used to perform defect recognition on the defect data set to be identified, realizing the practical application of defect recognition.

[0091] In this embodiment, a nondestructive detection system for concrete structure defects is used to achieve accurate defect identification through efficient ultrasonic detection and advanced MP-FA-GRU model. The system's layout unit ensures full coverage of the detection area, and the data acquisition unit collects sufficient training and identification data to improve the reliability of model training. The data processing unit improves the signal quality through multi-stage denoising and normalization. The MP-FA-GRU model combines multi-path feature extraction and fusion modules to enhance the feature extraction and fusion capabilities, ultimately improving the accuracy and real-time performance of defect identification. The test unit verifies the effectiveness of the system, making the system highly accurate and reliable for practical applications.

[0092] Embodiment 2

[0093] In order to determine the defect types of concrete roads in the city, the municipal department of a certain city used a non-destructive detection method for concrete structure defects. Figure 4 , which is a method flow chart of a nondestructive detection method for concrete structure defects.

[0094] Reference Figure 4 In step S01, ultrasonic detectors are placed on the concrete to be inspected, covering the concrete area with defects.

[0095] Reference Figure 4 In step S02, an ultrasonic signal is emitted, an echo signal propagated by concrete is received, the echo signal is used as original signal data, and multiple groups of original signal data are collected.

[0096] Among them, the original signal data includes data of known defect types and data of unknown defect types. The data of known defect types are constructed as a training data set and divided into a training set, a verification set and a test set in proportion. The data of unknown defect types are constructed as a defect data set to be identified.

[0097] Reference Figure 4 In step S03, data processing is performed on the original signal data, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data.

[0098] Among them, an adaptive filter is used to perform preliminary denoising processing on the multiple groups of original signal data to reduce the impact of noise according to real-time signal characteristics; a wavelet transform is used to perform multi-scale analysis on the multiple groups of original signal data to capture the characteristics of the multiple groups of original signal data at different scales, help identify and separate clutter, and further reduce the noise in the multiple groups of original signal data; the normalization normalizes the denoised signal data to a unified range, that is, within the interval of 0-1.

[0099] Furthermore, the specific implementation method of weighted fusion of denoised data is:

[0100] Assume that the original signal data is y[n]=x[n]+d[n]; where x[n] is the actual signal and d[n] is the noise;

[0101] Use an adaptive filter to perform preliminary denoising on y[n], and get in, is the noise estimated by the adaptive filter;

[0102] At the same time, wavelet transform is used to obtain y[n] 2 [n] = W -1(max(W[y[n]])-λ,0)·sign(W[y[n]]); where W is the wavelet transform; λ is the threshold;

[0103] Then, y 1 [n] and y 2 [n] weighted fusion, we get:

[0104]

[0105] Among them, f 1 and f 2 They are respectively 1 [n] and y 2 The function that models the features of [n].

[0106] Reference Figure 4 In step S04, a real-time defect recognition model MP-FA-GRU is constructed, wherein the real-time defect recognition model MP-FA-GRU includes a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes three feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module.

[0107] Among them, the input of the defect real-time identification model MP-FA-GRU is the result obtained by preliminary denoising, the result obtained by further denoising and the result of weighted fusion of the two; the output is the defect type of the concrete structure; after reading the input data, the defect real-time identification model MP-FA-GRU uses the multi-path feature extraction module to capture the global features, detail features and complementary features of each path feature; then, through the multi-path feature fusion module, the results of the three path outputs in the multi-path feature extraction module are fused; finally, the defect type is output through the output layer;

[0108] The multi-path feature extraction module includes three paths, the first path reads the result obtained by preliminary denoising, the second path reads the result obtained by further denoising, and the third path reads the result of weighted fusion of the two; the first path and the second path have the same structure, firstly, extract features through convolution blocks; then perform nonlinear mapping through tanh activation function; then, use the maximum pooling layer to capture significant features, and the input of the first path is connected to the maximum pooling layer through a residual connection layer; then, pass through a convolution block and tanh activation function, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, perform feature normalization through a BatchNorm layer;

[0109] The structure of the third path is similar to that of the first and second paths. First, features are extracted through a convolution block; then nonlinear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features; then, another convolution block and a tanh activation function are passed through, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, features are normalized through a BatchNorm layer.

[0110] Furthermore, the multipath feature fusion module includes a frequency domain attention module and a gated loop module;

[0111] In the frequency domain attention module, first, the output features of the first path, the second path and the third path are spliced; then, the features are subjected to fast Fourier transform FFT to obtain initial frequency domain features; then, the prominent features in the initial frequency domain features are captured through a linear activation function layer, a tanh activation function layer and another linear activation function layer to obtain frequency domain features; then, the frequency domain features are converted into weight distribution vectors through a softmax layer, and are element-by-element multiplied with the initial frequency domain features; then, the multiplied result is subjected to inverse fast Fourier transform IFFT to obtain time domain features;

[0112] In the gated recurrent module, the time domain features are passed through three gated recurrent units GRU to obtain the time dependency in the time domain features.

[0113] Reference Figure 4 In step S05, the real-time defect recognition model MP-FA-GRU is subjected to a performance test using a test set to verify its effectiveness.

[0114] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A nondestructive detection system for concrete structure defects, characterized in that: include: A deployment unit deploys ultrasonic detectors on the concrete to be tested to ensure that the defective concrete area is covered; A data acquisition unit transmits an ultrasonic signal, receives an echo signal propagated by the concrete, and collects multiple sets of original signal data as original signal data; wherein the original signal data includes data of known defect types and data of unknown defect types, the data of known defect types are constructed as a training data set, and are divided into a training set, a verification set, and a test set in proportion, and the data of unknown defect types are constructed as a defect data set to be identified; a data processing unit, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data; The MP-FA-GRU construction unit of the defect real-time recognition model includes a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes three feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module; The testing unit uses the test set to perform performance testing on the real-time defect recognition model MP-FA-GRU to verify its effectiveness.

2. A nondestructive detection system for concrete structure defects according to claim 1, characterized in that: In the data processing unit, an adaptive filter is used to perform preliminary denoising on the multiple groups of original signal data to reduce the impact of noise according to real-time signal characteristics; a wavelet transform is used to perform multi-scale analysis on the multiple groups of original signal data to capture the features at different scales in the multiple groups of original signal data, help identify and separate clutter, and further reduce the noise in the multiple groups of original signal data; the normalization normalizes the denoised signal data to a uniform range.

3. A nondestructive detection system for concrete structure defects according to claim 1, characterized in that: In the data processing unit, the specific implementation method of weighted fusion of denoised data is: Assume that the original signal data is y[n]=x[n]+d[n]; where x[n] is the actual signal and d[n] is the noise; Use an adaptive filter to perform preliminary denoising on y[n], and get in, is the noise estimated by the adaptive filter; At the same time, wavelet transform is used to obtain y2[n] = W -1 (max(W[y[n]])-λ,0)·sign(W[y[n]]); where W is the wavelet transform; λ is the threshold; Then, y1[n] and y2[n] are weightedly fused to obtain: Among them, f1 and f2 are functions that model the features of y1[n] and y2[n] respectively.

4. A nondestructive detection system for concrete structure defects according to claim 1, characterized in that: In the defect real-time identification model MP-FA-GRU construction unit, the input of the defect real-time identification model MP-FA-GRU is the result obtained by preliminary denoising, the result obtained by further denoising and the result of weighted fusion of the two; the output is the defect type of the concrete structure; after reading the input data, the defect real-time identification model MP-FA-GRU captures the global features, detail features and complementary features of each path feature through the multi-path feature extraction module; then, through the multi-path feature fusion module, the results of the three path outputs in the multi-path feature extraction module are fused; finally, the defect type is output through the output layer; The multi-path feature extraction module includes three paths, the first path reads the result obtained by preliminary denoising, the second path reads the result obtained by further denoising, and the third path reads the result of weighted fusion of the two; The first path and the second path have the same structure. First, features are extracted through a convolution block; then nonlinear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features, and the input of the first path is connected to the maximum pooling layer through a residual connection layer; then, a convolution block and a tanh activation function are passed, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, features are normalized through a BatchNorm layer; The structure of the third path is similar to that of the first and second paths. First, features are extracted through a convolution block; then nonlinear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features; then, another convolution block and a tanh activation function are passed through, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, features are normalized through a BatchNorm layer.

5. A nondestructive detection system for concrete structure defects according to claim 4, characterized in that: The multipath feature fusion module includes a frequency domain attention module and a gated loop module; In the frequency domain attention module, firstly, the output features of the first path, the second path and the third path are concatenated; then, the features are subjected to fast Fourier transform (FFT) to obtain initial frequency domain features; Then, the prominent features in the initial frequency domain features are captured through a linear activation function layer, a tanh activation function layer and another linear activation function layer to obtain frequency domain features; then, the frequency domain features are converted into weight distribution vectors through a softmax layer, and are multiplied element by element with the initial frequency domain features; then, the multiplied result is subjected to an inverse fast Fourier transform IFFT to obtain time domain features; In the gated recurrent module, the time domain features are passed through three gated recurrent units GRU to obtain the time dependency in the time domain features.

6. A method for nondestructive detection of concrete structure defects, characterized in that: include: The ultrasonic detector is placed on the concrete to be inspected, covering the concrete area with defects; Transmitting an ultrasonic signal, receiving an echo signal propagated by concrete, taking the echo signal as original signal data, and collecting multiple sets of original signal data; wherein the original signal data includes data of known defect types and data of unknown defect types, constructing the data of known defect types as a training data set, and dividing it into a training set, a verification set, and a test set in proportion, and constructing the data of unknown defect types as a defect data set to be identified; Performing data processing on the raw signal data, including preliminary denoising, further denoising, normalization and weighted fusion of denoised data; Constructing a real-time defect recognition model MP-FA-GRU, the real-time defect recognition model MP-FA-GRU includes a multi-path feature extraction module and a multi-path feature fusion module; the multi-path feature extraction module includes three feature extraction paths; the multi-path feature fusion module includes a frequency domain attention module and a gated loop module; The test set is used to perform performance testing on the real-time defect recognition model MP-FA-GRU to verify its effectiveness.

7. A method for nondestructive detection of concrete structure defects according to claim 6, characterized in that: An adaptive filter is used to perform preliminary denoising on the multiple groups of original signal data to reduce the impact of noise based on real-time signal characteristics; a wavelet transform is used to perform multi-scale analysis on the multiple groups of original signal data to capture features at different scales in the multiple groups of original signal data, help identify and separate clutter, and further reduce noise in the multiple groups of original signal data; and the normalization normalizes the denoised signal data to a uniform range.

8. A method for nondestructive detection of concrete structure defects according to claim 6, characterized in that: The specific implementation method of weighted fusion of denoised data is: Assuming that the original signal data is y[n], an adaptive filter is used to perform preliminary denoising on y[n] to obtain y1[n], and a wavelet transform is used on y[n] to obtain y2[n]; then, y1[n] and y2[n] are weightedly fused to obtain: Among them, f1 and f2 are functions that model the features of y1[n] and y2[n] respectively.

9. A method for nondestructive detection of concrete structure defects according to claim 6, characterized in that: The input of the real-time defect recognition model MP-FA-GRU is the result obtained by preliminary denoising, the result obtained by further denoising and the result of weighted fusion of the two; the output is the defect type of the concrete structure; after reading the input data, the real-time defect recognition model MP-FA-GRU uses the multi-path feature extraction module to capture the global features, detail features and complementary features of each path feature; then, the multi-path feature fusion module is used to fuse the results of the three path outputs in the multi-path feature extraction module; finally, the defect type is output through the output layer; The multi-path feature extraction module includes three paths, the first path reads the result obtained by preliminary denoising, the second path reads the result obtained by further denoising, and the third path reads the result of weighted fusion of the two; the first path and the second path have the same structure, firstly, extract features through convolution blocks; then perform nonlinear mapping through tanh activation function; then, use the maximum pooling layer to capture significant features, and the input of the first path is connected to the maximum pooling layer through a residual connection layer; then, pass through a convolution block and tanh activation function, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, perform feature normalization through a BatchNorm layer; The structure of the third path is similar to that of the first and second paths. First, features are extracted through a convolution block; then nonlinear mapping is performed through a tanh activation function; then, a maximum pooling layer is used to capture significant features; then, another convolution block and a tanh activation function are passed through, and the output of the maximum pooling layer is connected to the tanh activation function through a residual connection layer; finally, features are normalized through a BatchNorm layer.

10. A method for nondestructive detection of concrete structure defects according to claim 9, characterized in that: The multipath feature fusion module includes a frequency domain attention module and a gated loop module; In the frequency domain attention module, the output features of the first path, the second path and the third path are first concatenated; then, the prominent features in the initial frequency domain features are captured through a linear activation function layer, a tanh activation function layer and another linear activation function layer to obtain frequency domain features; then, the frequency domain features are converted into weight distribution vectors through a softmax layer, and are element-by-element multiplied with the initial frequency domain features; then, the multiplied result is subjected to an inverse fast Fourier transform IFFT to obtain time domain features; In the gated recurrent module, the time domain features are passed through three gated recurrent units GRU to obtain the time dependency in the time domain features.

Images