Method for identifying internal defects of recycled aggregate concrete

Abstract

The invention provides a method for identifying internal defects of recycled aggregate concrete, which comprises the following steps of: preparing a test piece containing prefabricated defects, collecting ultrasonic signals of the test piece, and denoising the collected ultrasonic signals to obtain the internal defects of the recycled aggregate concrete. Generating a data set based on the de-noised ultrasonic signal and preset internal defect information of the test piece, constructing an internal defect identification model, training the internal defect identification model based on the data set to obtain a trained internal defect identification model, collecting an ultrasonic signal of the to-be-identified recycled aggregate concrete, and identifying the to-be-identified recycled aggregate concrete according to the collected ultrasonic signal. And after de-noising processing is carried out on the recycled aggregate concrete, the recycled aggregate concrete is input into the trained internal defect identification model, and internal defect identification of the recycled aggregate concrete is realized. According to the method, the test piece containing the prefabricated defects is prepared, the ultrasonic signals are subjected to denoising treatment before use, noise can be removed, the accuracy can be improved, internal defect recognition of the recycled aggregate concrete can be achieved based on the internal defect recognition model, and use is convenient.

Description

Technical Field

[0001] The invention relates to the technical field of recycled aggregate concrete, and in particular to a method for identifying internal defects of recycled aggregate concrete. Background Art

[0002] With the acceleration of urbanization, the demand for concrete in the construction industry is increasing. In order to reduce costs and resource consumption, the use of recycled aggregate concrete (RAC) has gradually become a trend. Recycled aggregate concrete is an aggregate obtained by screening and crushing construction waste recycled from demolished buildings, roads and other structures, which replaces traditional natural aggregates for concrete preparation. Recycled aggregate concrete can not only effectively reduce dependence on natural resources, but also reduce the burden of construction waste on the environment, which meets the requirements of sustainable development.

[0003] However, the use of recycled aggregate concrete faces many challenges. Due to the complex sources of recycled aggregates, there may be varying degrees of pollution, uneven quality, differences in physical properties, and other problems, all of which will affect the overall performance of concrete. In particular, the surface of recycled aggregates may be attached with old concrete residues, impurities, and harmful substances. These factors may cause internal defects such as pores and cracks during the use of concrete, seriously affecting the structural safety and durability of concrete. Therefore, the technical needs for identifying and evaluating the internal defects of recycled aggregate concrete have become more urgent.

[0004] Traditional detection methods, such as visual inspection and X-ray detection, can detect concrete defects to a certain extent, but due to the particularity of recycled aggregate concrete, especially internal defects and hidden structural damage, these methods have limitations in detection accuracy and reliability. In addition, the detection requires a lot of calculations, the detection speed is relatively slow, and the efficiency is very low.

[0005] Therefore, it is necessary to design a method for identifying internal defects of recycled aggregate concrete. Summary of the invention

[0006] In order to overcome the deficiencies of the prior art, an object of the present invention is to provide a method for identifying internal defects of recycled aggregate concrete.

[0007] To achieve the above object, the present invention provides the following solutions: The present invention provides a method for identifying internal defects of recycled aggregate concrete, comprising: Step 100: preparing a test specimen with prefabricated defects and collecting ultrasonic signals of the test specimen; Step 200: performing denoising processing on the collected ultrasonic signal; Step 300: Generate a data set based on the denoised ultrasonic signal and preset internal defect information of the test specimen; Step 400: constructing an internal defect recognition model, and training the internal defect recognition model based on the data set to obtain a trained internal defect recognition model; Step 500: Collect the ultrasonic signal of the recycled aggregate concrete to be identified, perform denoising on it, and then input it into the trained internal defect recognition model to realize the internal defect recognition of the recycled aggregate concrete.

[0008] Preferably, in step 100, preparing a test specimen is specifically as follows: Experimental specimens were prepared based on the addition of expanded polystyrene particles into recycled aggregate concrete to simulate internal defects.

[0009] Preferably, the denoising process in step 200 and step 500 is specifically as follows: Based on singular value denoising, narrowband noise of the collected ultrasonic signal is removed; Based on the complementary set empirical mode decomposition adaptive threshold denoising method, high-frequency signal denoising is performed on the ultrasonic signal after singular value denoising.

[0010] Preferably, narrowband noise is removed from the collected ultrasonic signal based on singular value denoising, specifically: Collect ultrasonic signals and construct them into Hankel matrix functions; Decompose the Hankel matrix function; The signal decomposed from the singular value is selectively reconstructed through measurement quantization to form a new ultrasonic signal again to complete the denoising.

[0011] Preferably, high-frequency signal denoising is performed on the ultrasonic signal after singular value denoising based on the complementary set empirical mode decomposition adaptive threshold denoising method, specifically: The ultrasonic signal after singular value denoising is decomposed based on complementary set empirical mode decomposition adaptive threshold denoising method; Based on the autocorrelation method for judging the correlation of signals, the high-frequency and low-frequency signals of the ultrasonic signal are distinguished to obtain the high-frequency intrinsic mode function component and the low-frequency intrinsic mode function component; Extract high-frequency intrinsic mode function components and perform threshold denoising on them; The denoised high-frequency intrinsic mode function components and low-frequency intrinsic mode function components are reconstructed to complete the denoising of the ultrasonic signal.

[0012] Preferably, an internal defect recognition model is constructed, specifically: An internal defect recognition model is constructed based on a one-dimensional convolutional neural network structure, wherein the internal defect recognition model includes eight one-dimensional convolution operations, eight Relu activation functions, four pooling operations, four batch normalization layers and two fully connected operations. All operations of the internal defect recognition model are divided into four main feature extraction layers and two classification layers. The first four layers are divided into four groups of feature extraction layers, each group includes two one-dimensional convolution operations, two Relu operations, a batch normalization layer and a maximum pooling operation. The fifth layer contains one fully connected operation and one Relu operation. The last layer is a fully connected layer followed by a Softmax classifier, which maps the extracted features into defect type probability values.

[0013] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The present invention provides a method for identifying internal defects of recycled aggregate concrete, the method comprising: preparing a test specimen containing prefabricated defects, collecting ultrasonic signals of the test specimen, denoising the collected ultrasonic signals, generating a data set based on the denoised ultrasonic signals and preset internal defect information of the test specimen, constructing an internal defect recognition model, training the internal defect recognition model based on the data set to obtain the trained internal defect recognition model, collecting ultrasonic signals of recycled aggregate concrete to be identified, denoising the signals, and inputting the signals into the trained internal defect recognition model to realize internal defect recognition of recycled aggregate concrete. The present invention prepares a test specimen containing prefabricated defects, collects real ultrasonic signals for use, denoises the ultrasonic signals before use, can remove noise, improve accuracy, can realize internal defect recognition of recycled aggregate concrete based on the internal defect recognition model, and is easy to apply. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0015] Figure 1 A flow chart of a method provided by an embodiment of the present invention; Figure 2 It is a schematic diagram of expanded polystyrene particles uniformly wrapped in wet sand; Figure 3 This is a schematic diagram of the distribution of prefabrication defects inside the specimen; Figure 4 Schematic diagram of the internal defect recognition model structure built based on the one-dimensional convolutional neural network structure. DETAILED DESCRIPTION

[0016] 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.

[0017] The purpose of the present invention is to provide a method for identifying internal defects of recycled aggregate concrete, prepare test specimens containing prefabricated defects, collect real ultrasonic signals for use, and denoise the ultrasonic signals before use to remove noise and improve accuracy. Based on the internal defect recognition model, the internal defects of recycled aggregate concrete can be recognized, which is easy to use.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a method for identifying internal defects of recycled aggregate concrete, comprising: Step 100: preparing a test specimen containing prefabricated defects and collecting ultrasonic signals of the specimen; Step 200: performing denoising processing on the collected ultrasonic signal; Step 300: Generate a data set based on the denoised ultrasonic signal and preset internal defect information of the test specimen; Step 400: constructing an internal defect recognition model, and training the internal defect recognition model based on the data set to obtain a trained internal defect recognition model; Step 500: Collect the ultrasonic signal of the recycled aggregate concrete to be identified, perform denoising on it, and then input it into the trained internal defect recognition model to realize the internal defect recognition of the recycled aggregate concrete.

[0020] In step 100, a test specimen is prepared, specifically: Based on the addition of expanded polystyrene particles to recycled aggregate concrete to simulate internal defects, test specimens with prefabricated defects were prepared and described in detail as follows: Currently, there are two commonly used methods to prefabricate defects inside concrete specimens. One is to add air-entraining agents during the mixing process, and the other is to add expanded polystyrene particles to simulate initial defects. However, it is difficult to control the content and size of bubbles introduced by air-entraining agents. Therefore, expanded polystyrene particles are added to simulate initial defects. Reasons for process selection: Considering that the expanded polystyrene particles added in the above two existing construction processes are relatively high, the present invention only introduces a small amount of expanded polystyrene particles to simulate the initial pores, and the maximum amount is 8% or less. Therefore, it is necessary to adjust the above process and find a construction method suitable for concrete with a small amount of expanded polystyrene particles. In order to make the expanded polystyrene particles disperse better inside the recycled aggregate concrete specimen, the expanded polystyrene particles were first pretreated. Some sand was taken out and soaked with water and then mixed with the expanded polystyrene particles so that the expanded polystyrene particles could be evenly wrapped in the wet sand (such as Figure 2 As shown). At the same time, the method of first making cement mortar is adopted to wrap the expanded polystyrene particles by utilizing the cohesiveness of the slurry, thereby preventing the expanded polystyrene particles from floating excessively. According to the multiple preparations in the early stage of the test, the specific mixing process of the test is determined as follows: (1) Add the remaining sand and cement to the mixing pot in turn, and dry mix for 1.5 minutes; (2) Add 60% of the mixing water and all the water reducer, and stir for 1.5 minutes to make it into a slurry-like cement mortar; (3) Add the mixture of expanded polystyrene particles and sand, and stir for 1.5 minutes; (4) Add recycled coarse aggregate and the remaining mixing water, and stir for 3 minutes. The mixed mixture is poured into a plastic mold and vibrated for 20 to 22 seconds using a concrete magnetic vibration table. The expanded polystyrene particles vibrated out of the surface are shoveled into the test mold with a spatula, and the mold is evenly knocked around with a rubber hammer. The surface of the specimen is gradually flattened, and then it is ready for plastering. Cover the surface of the plastered specimen with a layer of plastic wrap, and then place it in the laboratory environment for 24 hours before removing the mold.

[0021] In order to verify the effectiveness of this process, some specimens were crushed and the crushed surfaces were soaked with water (e.g. Figure 3 As shown in the figure, it can be found that the above process can ensure that the expanded polystyrene particles (white particles) are relatively evenly dispersed inside the specimen, and no obvious floating or accumulation phenomenon is found, which can meet the requirements of subsequent tests.

[0022] In addition, the present invention also provides a finite element simulation operation, using the pressure acoustic transient module, solid mechanics module, electrostatic module and circuit module multi-physics field in the finite element software COMSOLMuitiphysics6.0 to simulate the sound field, and Abaqus can also be used to implement the above operation; First, a defective recycled aggregate concrete finite element model is established, and cracks and voids of different sizes and shapes are set in the concrete. Hard acoustic field boundary conditions are set at the upper and lower boundaries and defect boundaries of the model, and plane wave radiation boundary conditions are set at the left and right boundaries, which can also realize the collection of ultrasonic signals.

[0023] The denoising steps of step 200 and step 500 are the same, specifically: Based on singular value denoising, narrowband noise of the collected ultrasonic signal is removed; Based on the complementary set empirical mode decomposition (EMD) adaptive threshold denoising method (complementary set empirical mode decomposition (AN)), high-frequency signal denoising is performed on the ultrasonic signal after singular value denoising.

[0024] Based on singular value denoising, narrowband noise of the collected ultrasonic signal is removed, specifically: Singular value decomposition includes three processes. First, the ultrasonic signal is constructed into a Hankel matrix function; then, the Hankel matrix function is decomposed; finally, the signal decomposed by the singular value is selectively reconstructed through measurement quantization to form an ultrasonic signal again. The singular value itself is a decomposition method. Denoising mainly reflects the screening and reconstruction of the sub-signal components of the singular value decomposition. It is introduced in detail: Constructing Hankel matrix function For a one-dimensional array of partial discharge ultrasonic noise signals, set as X, it is: The Hankel function is constructed as A: The Hankel matrix function transforms the ultrasonic signal of partial discharge from a one-dimensional signal to a two-dimensional matrix; 2. Singular Value Decomposition The local discharge ultrasonic noise signal X can be expressed by the Hankel matrix function A: The mathematical geometric meaning of singular values ​​is equivalent to mapping a unit sphere with a radius of 1 into an elliptical sphere. The distance from the edge to the center of the ellipsoid, that is, the axis length, is the size of each singular value. After singular value decomposition, A can be expressed as: in, is the singular value of matrix A, U is the singular value The corresponding left singular vector, V is the right singular vector; 3. Denoising and reconstruction The original signal is subjected to singular value decomposition, and denoising and reconstruction are performed based on the singular value characteristics of each decomposed component. The principle of selection is to measure and quantify each singular value decomposition signal, and the signal is recorded as: The measurement quantification is mainly to calculate the norm 2 of the signal: Then the norm 1 of the signal is calculated as: The ratio of norm 2 to norm 1 is K. Among them, N represents the length of the signal, K value represents the measurement value of the measurement quantization, the size of each component measurement can reflect the signal characteristics, and reconstruction is performed based on the size of the component measurement of the odd decomposition.

[0025] Based on the complementary set empirical mode decomposition adaptive threshold denoising method, the high-frequency signal denoising of the ultrasonic signal after singular value denoising is performed, specifically: The ultrasonic signal after singular value denoising is decomposed based on complementary set empirical mode decomposition adaptive threshold denoising method; Based on the autocorrelation method for judging the correlation of signals, the high-frequency and low-frequency signals of the ultrasonic signal are distinguished to obtain the high-frequency intrinsic mode function component (high-frequency IMF component) and the low-frequency intrinsic mode function component (low-frequency IMF component); Among the decomposed intrinsic mode function components, the high-frequency signal contains noise, while the low-frequency component has become flat. It is believed that there is no noise component in the low-frequency component. After the high-frequency intrinsic mode function component is extracted, threshold denoising is used to ensure the integrity of the signal, so that the high-frequency signal is not completely discarded to make the final reconstructed signal, and the low-frequency signal is directly retained; The denoised high-frequency intrinsic mode function components and low-frequency intrinsic mode function components are reconstructed to complete the denoising of the ultrasonic signal.

[0026] The adaptive decomposition process of complementary set empirical mode decomposition is introduced: Complementary ensemble empirical mode decomposition adaptive, also known as adaptive noise complete empirical mode decomposition method, is an improved method of empirical mode decomposition method. It makes improvements on the method of ensemble empirical mode decomposition. The complementary ensemble empirical mode decomposition method uses positive and negative pairs of white noise to decompose the original ultrasonic signal, so that the modal aliasing problem of signal decomposition is suppressed, but the calculation efficiency is low, and the residual noise is not completely removed. The biggest difference between complementary ensemble empirical mode decomposition adaptive and ensemble empirical mode decomposition is that the white noise in the ensemble empirical mode decomposition method is directly added without adaptability and subsequent processing. After adding noise, the signal containing white noise is subjected to empirical mode decomposition, while the complementary ensemble empirical mode decomposition adaptive method adopts the method of gradually adding white noise in batches, starting with white noise, and then adding white noise again after each decomposition of the intrinsic mode function component, and it is adaptive, and it is iteratively decomposed downward in sequence. Compared with the ensemble empirical mode decomposition method, under the same number of averages, the complementary ensemble empirical mode decomposition adaptive method is superior in the completeness of signal decomposition, and the method of continuously removing noise can improve the calculation speed and efficiency. The solution method of the intrinsic mode function component first decomposed by the complementary ensemble empirical mode decomposition adaptive method is almost the same as that of the ensemble empirical mode decomposition. The signal y(t) that still contains noise is used as the original signal. N different white noises s(t) are added to this signal, and then the empirical mode decomposition is performed. Following the decomposition principle of the empirical mode decomposition method, the first intrinsic mode component is averaged as IMF1: Where i=1, 2, …, n; The first residual component obtained by removing IMF1 is: Add adaptive white noise to the signal after removing the first modal component to form a new modal signal. Repeat the decomposition of this signal to obtain IMF2: According to the above process, it can be concluded that the kth residual component is: Where k = 1, 2, ..., K; The kth IMF can be expressed as: Due to the continuous removal of extreme points for decomposition, when the residual function has no extreme points, that is, it conforms to the monotonic function, the decomposition process ends, and the original original signal function is expressed as: Among them, F k represents the kth intrinsic mode component IMF after empirical mode decomposition, is the noise coefficient. After the complementary set empirical mode decomposition is adaptively completed once, the signal obtained after removing the decomposed modal components is used as the new signal, and it continues to iterate downward in the same way. The complementary ensemble empirical mode decomposition adaptive method has better IMF decomposition results than the ensemble empirical mode decomposition method. In the case of ultrasonic signals, the complementary ensemble empirical mode decomposition adaptive method is used to decompose the signal, and then the decomposed intrinsic mode function components are decomposed by adding white noise and then removing the white noise. This method avoids the continuous existence of white noise in the decomposition process by continuously adding and removing white noise in the decomposition process, resulting in the residual noise not affecting the subsequent decomposition of intrinsic mode components, and not constituting continuity for the "pollution" of ultrasonic signal decomposition, which will minimize the negative effects of residual noise. The reconstructed signal greatly reduces the influence of modal aliasing effect and overcomes the defect of residual noise after adding white noise. The complementary ensemble empirical mode decomposition adaptive method is used to decompose the signal after filtering out narrowband noise into 11 IMFs and a residual component. Introduction to autocorrelation threshold denoising: The single-type empirical mode decomposition method uses a method of directly discarding the high-frequency components of the noise signal to denoise. In view of the denoising defect that the single-type empirical mode decomposition method cannot guarantee the integrity of the signal, the present invention uses an autocorrelation method to automatically determine the high and low frequency components of the signal, uses a threshold method to remove the noise contained in the high-frequency signal, and then synthesizes and reconstructs the high-frequency signal denoised by the threshold and the retained low-frequency signal; (1) Autocorrelation principle and determination First, in order to facilitate the calculation of the correlation degree of the signal and improve the calculation efficiency, the signal is normalized. The normalization function is: The time difference between different moments is It means that each intrinsic mode function component obtained by adaptive decomposition of the signal by complementary set empirical mode decomposition is normalized, and the autocorrelation calculation is performed on each normalized intrinsic mode function component. Taking y(t) as the noisy signal, the autocorrelation calculation formula is: Calculation Function The second norm of , this value is defined as the autocorrelation energy coefficient, and the calculation method is shown in the following formula: The autocorrelation energy coefficient represents the correlation degree of the signal at different times. The larger its value, the stronger the correlation. Introduction to threshold denoising: For the modes with autocorrelation energy coefficient less than 100, they are judged as high-frequency modes with much noise. The algorithm automatically screens and determines that the four IMF1~IMF4 are high-frequency mode components with noise. The threshold method is used to denoise them. The threshold denoising function is: According to the signal characteristics of the signal, the threshold denoising function model for processing each intrinsic mode function component can be changed to: Among them, it is necessary to restrict each IMF signal, T j is the threshold value of the intrinsic mode function component adaptively decomposed by the complementary set empirical mode decomposition in the method of the present invention, T j The values ​​are: Where N is the number of sampling points.

[0027] In step 300, a data set is generated based on the denoised ultrasonic signal and the preset internal defect information of the test specimen, specifically: A data set is generated based on the denoised ultrasonic signal and the internal defect information of the test specimen corresponding to the ultrasonic signal, wherein the internal defect information of the test specimen mainly includes the internal defect categories of the test specimen, including no defect, small circular hole, large circular hole, horizontal crack, vertical crack, oblique crack, etc.

[0028] In step 400, an internal defect recognition model is constructed, specifically: An internal defect recognition model is constructed based on a one-dimensional convolutional neural network structure, wherein the internal defect recognition model includes eight one-dimensional convolution operations, eight Relu activation functions, four pooling operations, four batch normalization operations, and two fully connected operations. All operations of the internal defect recognition model are divided into four main feature extraction layers and two classification layers. The first four layers are equipped with four groups of feature extraction layers, each of which includes two one-dimensional convolution operations, two Relu operations, one batch normalization operation, and one maximum pooling operation. The fifth layer contains one fully connected operation and one Relu operation. The last layer is a fully connected layer followed by a Softmax classifier, which maps the extracted features into defect type probability values. The schematic diagram is shown as follows: Figure 4 shown.

[0029] In each layer after the input layer in the network model constructed by the present invention, the kernel is halved by the maximum pooling layer. Here, the pooling size is set to 2 and the stride is 1, that is, the pooling fully acts on each sample data, and there is no jump phenomenon. After the input data passes through eight convolutional layers, the output data is flattened and connected to two fully connected layers. The kernel of the last layer is equal to the number of output types, which is set according to specific needs. The activation function used in the network is Relu activation function except for the activation function of the last output layer, which is Softmax.

[0030] Study its principle: The input form of convolution is two vectors or matrices, one of which is the original input and the other is the kernel or filter. When the activation function is input, a third vector or matrix is ​​output; Assuming the input is a vector f of length n and a filter g of length m, the one-dimensional convolution f*g is defined as follows: The kernel is mapped to the input tensor. For each filter, the overlapping values ​​of the filter and the input tensor are multiplied and then summed. The result of this multiplication and sum will be the output value at the center of the kernel in the input. The one-dimensional convolutional layer operation used is shown in the following formula: In the formula, y l 、b l 、f l-1 , g l-1 They represent the output of the lth layer, the bias of the lth layer, the output of the (l-1)th layer, and the filter of the (l-1)th layer respectively. The dense layer output O of the one-dimensional convolutional neural network model adopts a typical perceptron form. The calculation method is shown in the following formula: Where b, w and x represent bias, weight and input respectively; The predicted output value is completed by the Softmax activation function, and the calculation method is shown in the following formula: In the formula, O i is the confidence level of belonging to category i; The cross entropy loss is calculated as follows: In the formula, y and are the actual value and the predicted value respectively.

[0031] The amount of data in the data set and the selection of hyperparameters determine the success of the deep network. The present invention sets up training, verification and testing parts. A certain amount of ultrasound signal data set is input into the model to learn the hyperparameters. The training data is input to make the model learn the network parameters. The verification data set is used to check the generalization ability of the model learning parameters. The training and verification process will continue until the model reaches the generalization standard. The deep convolutional network is composed of a large number of parameters and some hyperparameters, including learning rate, number of layers and filter size, etc. By using tensorflow to simplify the hyperparameter learning framework, the model can learn the optimal network parameters, which can maximize the test accuracy. The adopted hyperparameter learning framework consists of three main modules and a control logic. The data set is divided in the early stage, and the training data set is first input into the model. The parameters can be learned from the training data set, and then the verification data set is used to verify the performance of the learning network. At this time, if the verification data set can achieve the expected accuracy, then save the model and apply these parameters for testing; if the result of the verification data set does not achieve the expected accuracy, then continue to learn new parameters through the network, and use the new information in the parameters. For the model recorded in this application, select the training cycle, learning rate η, and sequence length s for multiple experiments to determine that these parameters of the model have the optimal values. The difference of these hyperparameters will have different effects on network performance; Through specific experimental verification, this application selects the best training cycle value of 50, uses the Adam optimizer with an initial learning rate of 0.002, uses a data set sequence length of 3500, and a training number of 50. At this time, the model has the best effect.

[0032] The experimental environment of the present invention is Tensorflow1.14 and python3.6, using python programming language. The computer graphics card configuration is Inter(R)i5-10400F-CPU, NVIDIA Ge Force RTX 2060-GPU, and 64-bit Windows 10 operating system.

[0033] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0034] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for identifying internal defects of recycled aggregate concrete, characterized in that: include: Step 100: preparing a test specimen with prefabricated defects and collecting ultrasonic signals of the test specimen; Step 200: performing denoising processing on the collected ultrasonic signal; Step 300: Generate a data set based on the denoised ultrasonic signal and preset internal defect information of the test specimen; Step 400: constructing an internal defect recognition model, and training the internal defect recognition model based on the data set to obtain a trained internal defect recognition model; Step 500: Collect the ultrasonic signal of the recycled aggregate concrete to be identified, perform denoising on it, and then input it into the trained internal defect recognition model to realize the internal defect recognition of the recycled aggregate concrete.

2. The method according to claim 1, characterized in that In step 100, a test specimen is prepared, specifically: Based on the addition of expanded polystyrene particles into recycled aggregate concrete to simulate internal defects, experimental specimens with precast defects were prepared.

3. The method according to claim 2, characterized in that The denoising process in step 200 and step 500 is specifically as follows: Based on singular value denoising, narrowband noise of the collected ultrasonic signal is removed; Based on the complementary set empirical mode decomposition adaptive threshold denoising method, high-frequency signal denoising is performed on the ultrasonic signal after singular value denoising.

4. The method according to claim 3, characterized in that Based on singular value denoising, narrowband noise of the collected ultrasonic signal is removed, specifically: Collect ultrasonic signals and construct them into Hankel matrix functions; Decompose the Hankel matrix function; The signal decomposed from the singular value is selectively reconstructed through measurement quantization to form a new ultrasonic signal again to complete the denoising.

5. The method according to claim 4, characterized in that Based on the complementary set empirical mode decomposition adaptive threshold denoising method, the high-frequency signal denoising of the ultrasonic signal after singular value denoising is performed, specifically: The ultrasonic signal after singular value denoising is decomposed based on complementary set empirical mode decomposition adaptive threshold denoising method; Based on the autocorrelation method for judging the correlation of signals, the high-frequency and low-frequency signals of the ultrasonic signal are distinguished to obtain the high-frequency intrinsic mode function component and the low-frequency intrinsic mode function component; Extract high-frequency intrinsic mode function components and perform threshold denoising on them; The denoised high-frequency intrinsic mode function components and low-frequency intrinsic mode function components are reconstructed to complete the denoising of the ultrasonic signal.

6. The method according to claim 1, characterized in that In step 400, an internal defect recognition model is constructed, specifically: An internal defect recognition model is constructed based on a one-dimensional convolutional neural network structure, wherein the internal defect recognition model includes eight one-dimensional convolution operations, eight Relu activation functions, four pooling operations, four batch normalization layers and two fully connected operations. All operations of the internal defect recognition model are divided into four main feature extraction layers and two classification layers. The first four layers are divided into four groups of feature extraction layers, each group includes two one-dimensional convolution operations, two Relu operations, one batch normalization operation and one maximum pooling operation. The fifth layer contains one fully connected operation and one Relu operation. The last layer is a fully connected layer followed by a Softmax classifier, which maps the extracted features into defect type probability values.

Images