Concrete structure damage detection system and method based on percussion sound and acceleration information

Abstract

The invention relates to a concrete structure damage detection system and method based on percussion sound and acceleration information, the system comprises a percussion hammer, a microphone, an acceleration sensor, a data acquisition instrument and a processor, and the method comprises the following steps: collecting the percussion sound and acceleration information of a concrete structure by using the percussion hammer, the microphone and the acceleration sensor; carrying out data processing on the collected percussion sound and acceleration information to construct a data set; training, verifying and testing the constructed neural network architecture by using the data set, and establishing a damage state detection model; and obtaining percussion sound and acceleration information of the current concrete structure, inputting the percussion sound and acceleration information into the damage state detection model after data processing, and outputting to obtain a damage state detection result. Compared with the prior art, the damage state of the concrete structure can be accurately, conveniently and nondestructively detected in real time, and the damage state can be classified and identified.

Description

Technical Field

[0001] The present invention relates to the technical field of building structure damage detection, and in particular to a concrete structure damage detection system and method based on percussion sound and acceleration information. Background Art

[0002] In the field of construction, concrete structures are large in number and widespread. After decades of service, the performance of many structures has begun to deteriorate. In order to detect potential problems in a timely manner and take appropriate repair measures, the safety status detection of concrete structures has become crucial.

[0003] The existing methods for damage detection of concrete structures mainly include: visual inspection, ultrasonic testing, drilling testing, etc. Among them, visual inspection is one of the most commonly used detection methods. By carefully observing the appearance of the concrete structure, cracks, deformations, bulges and other problems can be found. This method is simple and intuitive, and can be used to preliminarily evaluate the overall condition of the structure, but it cannot accurately and timely detect subtle and internal defects, and it is time-consuming and labor-intensive. It is even more difficult to grasp the structural performance in complex environments;

[0004] Ultrasonic testing is a non-destructive testing method that uses the difference in the propagation speed of ultrasound in different materials to detect defects inside concrete. By measuring the propagation time and reflection intensity of ultrasound, cracks, cavities, etc. in concrete can be determined. This method can provide relatively accurate internal information, but it requires professional equipment and specially trained technicians to operate. In addition, it is still difficult to directly determine the damage of the structure based on the results of ultrasonic detection alone.

[0005] Drilling testing evaluates the quality and durability of the structure by drilling samples into the concrete structure and conducting laboratory analysis and testing. This includes measuring indicators such as the strength, density, and compressive strength of the concrete. However, this method is destructive and can only characterize the local material performance characteristics of the inspected area.

[0006] In summary, most of the existing detection methods are unable to obtain the current status of concrete structures in a timely, accurate, convenient and non-destructive manner, and it is also difficult to classify and determine the damage status, which is not conducive to the rapid performance evaluation and reinforcement and repair of large-scale urban houses. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a concrete structure damage detection system and method based on percussion sound and acceleration information, which can detect the damage status of the concrete structure in real time, accurately, conveniently and non-destructively, and can classify and identify the damage status.

[0008] The object of the present invention can be achieved by the following technical solutions: A concrete structure damage detection system based on percussion sound and acceleration information, comprising a percussion hammer, a microphone, an acceleration sensor, a data acquisition instrument and a processor, wherein the microphone and the acceleration sensor are respectively connected to the data acquisition instrument, the data acquisition instrument is connected to the processor, and the percussion hammer is used to percuss the concrete structure;

[0009] The microphone and acceleration sensor are used to collect vibration information during the percussion process, including percussion sound information and acceleration information;

[0010] The data collector is used to transmit the collected vibration information to the processor;

[0011] The processor is provided with a pre-trained damage detection model, and the vibration information is input into the damage detection model, and the corresponding damage state detection result is output.

[0012] A concrete structure damage detection method based on percussion sound and acceleration information comprises the following steps:

[0013] S1. Collect the percussion sound and acceleration information of the concrete structure and perform data processing to construct a data set, wherein the data processing process includes cleaning, denoising, normalization, fast Fourier transform, feature extraction using MFCC, and damage state association;

[0014] S2. Divide the data set into a training set, a validation set, and a test set according to a preset ratio, and use the training set, the validation set, and the test set to train, validate, and test the constructed neural network architecture in turn to establish a damage status detection model;

[0015] S3. Obtain the percussion sound and acceleration information of the current concrete structure, input the data into the damage state detection model after data processing, and output the damage state detection result.

[0016] Furthermore, the step S1 includes the following steps:

[0017] S11, tapping a specific position on the concrete structure for a set number of times using a percussion method, while collecting acceleration information and sound information;

[0018] S12, after cleaning, denoising and normalizing the collected acceleration information and sound information, fast Fourier transform is performed respectively, and the corresponding MFCC (Mel-Frequency Cepstral Coefficients) is extracted;

[0019] S13, associating the MFCC corresponding to the acceleration information and the data label corresponding to the damage state of the specific position of the concrete structure to construct an acceleration information data set;

[0020] The MFCC corresponding to the sound information and the data labels corresponding to the damage status of specific positions of the concrete structure are associated to construct a sound information dataset.

[0021] Furthermore, the MFCC extraction process in step S12 includes: processing the spectrum obtained after fast Fourier transform through a Mel filter bank, the Mel filter bank includes a plurality of filters distributed on different Mel scales, and the output of each filter is the sum of the power spectrum of the frequency range covered by it;

[0022] Then, the output of the Mel filter bank is logarithmized, and DCT (Discrete cosine transmission) is performed on the output of the Mel filter bank after logarithmization, and the first N coefficients of the DCT result are selected as MFCC.

[0023] Furthermore, the data labels corresponding to the damage status of the specific positions of the concrete structure in step S13 include four damage categories: no damage, slight damage, moderate damage, and severe damage.

[0024] Furthermore, the damage-free category corresponds to no cracks or only a few tiny cracks;

[0025] The minor damage category corresponds to certain fine cracks with a width of less than 1 mm;

[0026] The moderate damage category corresponds to crack width greater than 1 mm or a small amount of concrete spalling;

[0027] The severe damage category corresponds to obvious concrete spalling or exposed steel bars.

[0028] Furthermore, the step S2 comprises the following steps:

[0029] S21, respectively using the training set, test set, and verification set of acceleration information and sound information to train, test, and verify the constructed neural network architecture to obtain a first deep learning model and a second deep learning model;

[0030] S22. Use the first deep learning model and the second deep learning model together as a damage status detection model.

[0031] Furthermore, the constructed neural network architecture includes a convolutional layer, a batch normalization layer, a maximum pooling layer and a fully connected layer, wherein the association module of the convolutional layer + batch normalization layer + pooling layer is repeated four times in the entire network, each association module is linearly connected, and finally a fully connected layer realizes the linear association of the previous layer and outputs the classification result: expressed as classification scores of different injury categories, and the highest score is taken as the classification result; the loss function of the neural network is the negative log-likelihood loss function (Negative Log-Likelihood Loss, NLL Loss).

[0032] Furthermore, step S3 specifically adopts the same method as step S1 to obtain the percussion sound and acceleration information of the current concrete structure, and then performs windowing, framing and fast Fourier transformation on the percussion sound and acceleration information to obtain a spectrum diagram, and performs filtering, logarithm taking and DCT processing on the spectrum to obtain corresponding MFCC data, and inputs the MFCC data into the damage status detection model to output the damage status detection result.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] The present invention relates to a concrete structure damage detection system, which includes a percussion hammer, a microphone, an acceleration sensor, a data acquisition instrument and a processor, wherein the microphone and the acceleration sensor are respectively connected to the data acquisition instrument, which is connected to the processor, and the corresponding position of the concrete structure is percussed by the percussion hammer, and the vibration information during the percussion process is collected by the microphone and the acceleration sensor, including the percussion sound information and the acceleration information, and then the collected vibration information is transmitted to the processor by the data acquisition instrument, and the vibration information is input into the damage detection model using the pre-trained damage detection model set in the processor, and the corresponding damage state detection result can be output. In this way, the percussion method and the deep learning classification model algorithm are combined, and the damage state detection of the concrete structure can be carried out in real time, accurately, conveniently and non-destructively, and the classification and identification of the damage state can be carried out.

[0035] The design of the present invention adopts the percussion method to knock a set number of times at a specific position on the concrete structure, and collects acceleration information and sound information at the same time, and performs fast Fourier transform on the collected acceleration information and sound information respectively, extracts the corresponding MFCC, and uses MFCC as the characteristics of sound and acceleration signals to form a data set to train a neural network model, which can achieve high classification accuracy and does not require high computing power. It has been verified that it only takes about 3 to 4 seconds to predict 360 sound signals. It has the advantages of high efficiency and low computing power requirements, and can be deployed on mobile devices for real-time reasoning detection and classification.

[0036] The present invention uses the crack width within the entire node range as a quantitative basis for judging the degree of node damage, and considers concrete spalling and exposed steel bars, and divides the damage category labels into four categories: no damage (no cracks or only a few fine cracks), slight damage (certain fine cracks with a width less than 1 mm), moderate damage (crack width greater than 1 mm or a small amount of concrete spalling) and severe damage (obvious concrete spalling or exposed steel bars), so that the damage state detection model obtained by subsequent training can accurately output the corresponding damage category results.

[0037] The present invention uses the acceleration information data set and the sound information data set to train the constructed neural network architecture respectively, and obtains the first deep learning model and the second deep learning model; the first deep learning model and the second deep learning model are used together as damage status detection models for practical application. The two complement each other and can improve the effectiveness of on-site detection. The result outputs of the two are used together as the basis for structural damage status assessment, and can also avoid distortion caused by factors such as data acquisition environment and equipment when one of the deep learning models is actually applied, thereby fully ensuring the reliability of the damage status detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0039] Figure 2 A schematic diagram of a concrete structure model in an embodiment;

[0040] Figure 3a~3d It is a schematic diagram of the damage results of the concrete structure in the embodiment;

[0041] Figure 4 It is a diagram of the arrangement of the knocking nodes in the embodiment;

[0042] Figure 5 The waveform diagram of the sound information collected in the embodiment;

[0043] Figure 6 It is a waveform diagram of acceleration information collected in the embodiment;

[0044] Figure 7 A schematic diagram of a beam for collecting percussion vibration information in an embodiment;

[0045] Figure 8 This is the relationship between the Mel scale and the Hertz scale;

[0046] Fig. 9 Schematic diagram of the process of extracting the corresponding MFCC from a percussion sound information;

[0047] Fig.10 A schematic diagram of the definition of a tap point data label in an embodiment;

[0048] Figures 11a-11b This is a schematic diagram corresponding to the embodiment without damage;

[0049] Figures 12a to 12b This is a schematic diagram corresponding to a slight injury in the embodiment;

[0050] Figures 13a-13b This is a schematic diagram corresponding to moderate damage in the embodiment;

[0051] Figures 14a-14b This is a schematic diagram corresponding to severe injury in the embodiment;

[0052] Fig.15 A schematic diagram of a neural network architecture constructed in an embodiment;

[0053] Figures 16a-16b The parameter change trend in the neural network model training based on sound information in the embodiment;

[0054] Fig.17 The confusion matrix obtained by using the MFCC of the sound information as the data set in the embodiment;

[0055] Figures 18a-18b The parameter change trend in the neural network model training based on acceleration information in the embodiment;

[0056] Fig.19 MFCC is a confusion matrix obtained by using the MFCC of acceleration information as a data set in the embodiment. DETAILED DESCRIPTION

[0057] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Example

[0059] The percussion method is to analyze the sound characteristics produced by percussion and determine whether the structure is damaged based on the changes in the sound. This method has the advantages of being simple and easy to use, and will not cause any damage to the original structure, does not require any embedded parts, and does not require cumbersome wires. It also has obvious advantages when detecting some structural damage that is difficult to observe directly, such as deep cracks or internal cavities. The percussion method is not just a simple tapping and listening to the sound, it involves complex acoustic principles and measurement techniques. When performing structural damage detection based on percussion,

[0060] An accelerometer and a microphone with a sampling frequency of about 50 Hz are selected as data acquisition devices. In this implementation case, the microphone sampling frequency is set to 44.1 Hz, and the acquisition frequency during acceleration data acquisition is set to 1024 Hz. First, the percussion points and acquisition points are arranged according to the size and type of the structure, and then the sensor is installed on the surface of the structure acquisition point, and the sound and acceleration vibration signals are generated by knocking the surface of the structure with a percussion hammer. Next, the collected signals are denoised and normalized using a preprocessing algorithm, and then features are extracted. For sound and acceleration signals, we use Mel-frequency cepstral coefficients as acoustic features for deep learning model training. The extracted features are input into the pre-trained convolutional neural network model for damage classification.

[0061] The collected sound signal usually needs to specify the sound information that needs to be processed; in some cases, the time domain signal of the original sound can be directly used as its sound feature without any processing, but in most cases, in order to ensure the accuracy, reliability and robustness of the detection results, it is usually necessary to process the original sound signal to a certain extent, such as dividing the audio signal according to the tapping situation, removing environmental noise, adjusting the sound amplitude, and then performing a fast Fourier transform (FFT) from the time domain to the frequency domain, and then using the Mel-Frequency Cepstral Coefficients (MFCC) to extract features.

[0062] Mel-frequency cepstral coefficients are a method used in sound processing to represent the short-time power spectrum of sound. It is a linear cosine transform of the logarithmic power spectrum based on the nonlinear Mel frequency scale. MFCC contains a series of coefficients. Mel-frequency cepstral coefficients are widely used in many audio processing tasks such as speech recognition and music information retrieval. Fast Fourier transform is an algorithm for calculating discrete Fourier transform or its inverse transform. Fourier transform can convert a signal from its original domain (usually time or space) to a representation in the frequency domain.

[0063] In engineering practice and scientific research, FFT is widely used in digital signal processing; a typical application is in audio signal processing, where FFT can be used to convert audio signals from the time domain to the frequency domain, making it easier to observe the spectral characteristics of the signal, such as the intensity of each frequency component in the audio signal. The application of percussion in building structure damage detection shows great potential and prospects.

[0064] To this end, this scheme design applies percussion method and deep learning technology to concrete structure damage detection, and proposes a concrete structure damage detection system based on percussion sound and acceleration information, including a percussion hammer, a microphone, an acceleration sensor, a data acquisition device and a processor. The microphone and the acceleration sensor are respectively connected to the data acquisition device, and the data acquisition device is connected to the processor. The percussion hammer is used to tap the concrete structure;

[0065] The microphone and the acceleration sensor are used to collect vibration information during the tapping process, including percussion sound information and acceleration information;

[0066] The data logger is used to transmit the collected vibration information to the processor;

[0067] A pre-trained damage detection model is set in the processor. The vibration information is input into the damage detection model, and the corresponding damage status detection result is output.

[0068] Based on the above system, a concrete structure damage detection method based on percussion sound and acceleration information is implemented, such as Figure 1 As shown, the following steps are included:

[0069] S1. Collect the percussion sound and acceleration information of the concrete structure and perform data processing to construct a data set, wherein the data processing process includes cleaning, denoising, normalization, fast Fourier transform, feature extraction using MFCC, and damage state association;

[0070] S2. Divide the data set into a training set, a validation set, and a test set according to a preset ratio, and use the training set, the validation set, and the test set to train, validate, and test the constructed neural network architecture in turn to establish a damage status detection model;

[0071] S3. Obtain the percussion sound and acceleration information of the current concrete structure, input the data into the damage state detection model after data processing, and output the damage state detection result.

[0072] This embodiment applies the above technical solution to a scaled model of a damaged reinforced concrete frame structure to study the detection and classification capabilities of the solution for the damage state of the concrete structure, and collects data again after the damaged nodes are reinforced to determine the reinforcement effect.

[0073] In this embodiment, Figure 2As shown in the figure, the concrete structure model is a five-story reinforced concrete frame 1:3 scale structure, the prototype floor height is 3m, the total height of the structure is 15m; the plane size is 10.8m×7.2m, the X direction has two spans, each span is 5.4m, and the Y direction has a single span, the span is 7.2m. The model length similarity coefficient is selected as 1 / 3. According to this similarity coefficient, the total height of the model is 5.0 meters (including the base), the plane size of the model is 3.6m×3.0m, and the beams, slabs and columns are made of C30 commercial concrete, which is poured in batches and layers according to the number of structural layers.

[0074] After the scaled model was tested on a simulated earthquake vibration table with multiple seismic wave conditions equivalent to a rare earthquake, obvious cracks appeared at the structural beam-column joints, and large pieces of concrete peeled off at some joints. The final structure has some obvious damage (mainly concentrated at the beam-column joints). Figure 3a~3d As shown, Figure 3a It corresponds to the five-story beam-column node on the B axis. Figure 3b The corresponding node is a 2-axis five-story beam-column node. Figure 3c The corresponding node is the second-floor beam-column node of the B-axis. Figure 3d It corresponds to the first-layer beam-column node of the C-axis.

[0075] After that, the percussion information is collected: the percussion method is used to knock a certain number of times at specific locations on a certain number of beams in the frame structure, and the acceleration information and sound information are collected at the same time, and the deep learning classification model algorithm is used to determine the degree and location of damage. For the collected sound information, the fast Fourier transform is performed on it, and its Mel frequency cepstrum coefficients are extracted, and its performance is analyzed. After the first data collection, data processing, network training and performance verification, all damaged nodes are reinforced, and the sound and acceleration data are collected using the same method. The previously trained neural network is applied to the new data for classification and prediction, and the quality of the reinforcement and the response of the reinforced structure are evaluated.

[0076] For each beam used for testing, it is divided into 4 parts along the length direction, with a total of 5 knocking nodes, namely the left end point, the left 1 / 4 span point, the mid-span point, the right 1 / 4 span point and the right end point. In order to facilitate the knocking of the concrete at the beam-column node, the knocking point is uniformly set to the position of half the height of the frame facade beam, such as Figure 4 As shown, Figure 4 In the figure, DK indicates a single-span beam, P (Percussion) indicates the percussion point, T (Testing) indicates the microphone sound pickup point; 2 indicates 2-axis, 2M and 1M indicate the 1 / 4 span closest to the 2-axis and the 1 / 4 span closest to the 1-axis respectively, M (Middle) indicates the mid-span; 5F indicates the 5th floor, and the same applies to the floors below the 5th floor. The blue (mid-span of the beam) indicates the percussion point, and the yellow (top of the beam) indicates the sound pickup and acceleration measurement point.

[0077] Each tapping position is tapped 60 times, and the rate is controlled at about 1 second / time, that is, the data collection of one tapping position is completed in about 60 seconds. In order to avoid interference from the laboratory environment noise, the best effect can be obtained by selecting the evening to late night time for the test. When tapping, the same model of hammer body and hammer head should be used, and the tapping should be performed with as close a force as possible; at the same time, the tapping interval should be controlled as much as possible, and tools such as a metronome can be used as an aid. In order to ensure that the percussion data is collected without duplication or omission, each tapping point needs to be numbered and the tapping order needs to be determined.

[0078] The equipment used in this embodiment mainly includes: a percussion hammer, a laptop computer, a small wireless microphone, a data acquisition instrument (INV3062 acquisition instrument of Dongfang Institute), and a three-axis acceleration sensor. The acceleration sensor is adhered to the upper end of the beam with epoxy resin glue for adhesion and fixation.

[0079] The sampling frequency of the microphone is set to 44.1kHz. A metronome is used to assist in adjusting the tapping frequency during tapping. The nodes are tapped in a predetermined order. The tapping number needs to be recorded after each tapping, and the data record needs to be checked for normality and completeness. Figure 5 The percussion sound waveform of a node collected by the microphone (the first 12 times are intercepted, corresponding to DK-2-1F, that is, the 2-axis node of one layer in a single span direction); due to the length of the data, only the first 12 percussion waveforms of the entire waveform are intercepted. Similarly, Figure 6 The acceleration information is collected by the accelerometer at the same position, and the waveforms of the middle 6 taps are captured.

[0080] According to the damage conditions of different beams and considering the need to select beams with the same size, support conditions and node form, a total of 8 beams were collected from the two frames of the C-axis and A-axis, with 5 percussion points on each beam (excluding the nodes), and data was collected for a total of 40 percussion points; each node was percussed 60 times, and the size of all data in units of sound signals was 60×40=2400, which was the case for both acceleration and sound data. Figure 7 As shown, according to the following tapping arrangement diagram, the number of nodes at the beam end to be tested is 16, Figure 7 The medium blue dotted box represents the four beams on the rear side (A-axis) of the elevation view.

[0081] Mel-frequency cepstral coefficients (MFCC) are then used as acoustic features for subsequent deep learning model training.

[0082] The basis of MFCC is the Mel scale, which is a scale that has a nonlinear relationship with the Hertz scale, but is linearly related to the pitch directly perceived by the human ear, but a logarithmic relationship. This means that at lower frequencies, a small change in the Hertz value will result in a large change in the Mel scale, while at higher frequencies, a large change in the Hertz value may only result in a small change in the Mel scale. This property makes the Mel scale better simulate the human ear's perception of sound, because the human ear is more sensitive to lower pitches than to higher pitches. The conversion relationship between the Mel scale and the Hertz scale is shown in the following formula and Figure 8 (Depend on Figure 8 It can be seen that for high-frequency sound signals, the Mel scale is less sensitive to its changes than for low-frequency sound signals):

[0083]

[0084] Where m is the frequency on the Mel scale and f is the frequency on the Hertz scale. The process of extracting MFCC from an audio can be simply divided into the following steps:

[0085] 1. Pre-emphasize, frame (in this embodiment, every 512 sample points is set as one frame) and add window to the audio;

[0086] 2. Apply fast Fourier transform to each separated frame to obtain the spectrum image of each frame;

[0087] 3. Process the spectrum through a Mel filter bank (usually 20 to 40 filters, distributed on the Mel scale), where the output of each filter is the sum of the power spectra of the frequency range it covers;

[0088] 4. Take the logarithm of the output of the filter bank to simulate the human ear's perception of volume;

[0089] 5. Finally, perform discrete cosine transform (DCT) on the logarithmic filter bank output. Usually the first 12-13 coefficients of the DCT result are selected as MFCC.

[0090] This embodiment uses a Python sound and music processing tool library called librosa to automatically detect and extract each beat (i.e., the audio of a single tap) in each audio file, and stores each tap audio separately, and then extracts the Mel frequency cepstral coefficients according to the above steps. The steps of the whole process are shown in Fig. 9 Each frame of the sound signal (each frame contains 512 sound samples) can derive a set of MFCC parameters, with 13 parameters in each set. Considering that the audio sampling rate is 44.1kHz, the total number of all parameters is:

[0091] 2400×44100 / 512×13=2,687,343

[0092] Among them, 2400 is the number of taps in the entire test.

[0093] In order to apply deep learning technology to the detection of beam damage, nodes with different degrees of damage need to be classified and assigned corresponding deep learning labels. Considering that most of the cracks in the structure appear at the nodes, this embodiment assumes that there is no damage in the middle of the beam and column, and the damage is all concentrated at the beam-column nodes. Existing studies have used the following appearance standards to classify the damage of concrete structures that have experienced earthquakes into five levels:

[0094] Level 1: No damage, with appearance characteristics of no cracks or only some minor cracks;

[0095] Level 2: Slight damage, with appearance features of some fine cracks and a small number of wide cracks;

[0096] Level 3: Moderate damage, with the appearance of concrete peeling off the surface;

[0097] Level 4: Severe damage, with the appearance of large amounts of spalling and exposed steel bars;

[0098] Level 5: Near collapse, with the appearance of core concrete crushed and steel bars yielding.

[0099] In this embodiment, there is no level five damage, so only the first four damage categories are considered. The crack width within the entire node range is used as a quantitative basis for judging the degree of damage to the node, and considering concrete spalling and exposed steel bars, it is divided into four categories: no damage (no cracks or only a few fine cracks), slight damage (certain fine cracks, width less than 1mm), moderate damage (crack width greater than 1mm or a small amount of concrete spalling) and severe damage (obvious concrete spalling or exposed steel bars). The damage classification results of this embodiment according to the above classification method include: There are 5 percussion points in the no damage level, corresponding to 2 nodes (no damage example such as Figures 11a-11b shown);

[0100] There are 8 percussion points for the minor injury level, corresponding to 4 nodes (for example, minor injury Figures 12a to 12b shown);

[0101] There are 17 percussion points for moderate injury, corresponding to 6 nodes (for example, Figures 13a-13b shown);

[0102] There are 10 percussion points for severe injury, corresponding to 4 nodes (for example, Figures 14a-14b shown).

[0103] In deep learning, each piece of data needs to have a corresponding label for deep learning model training or accuracy verification. In this embodiment, the deep learning label is the degree of damage; however, there is no damage at the percussion points that are not at the beam-column nodes in the experiment, so the degree of damage to which it belongs needs to be defined. For the convenience of research, this embodiment corresponds the degree of damage corresponding to the two percussion points on the left side of a beam (at the node and at 1 / 4 of the left span) to the beam-column node on its left side, and vice versa; for the mid-span percussion points, for safety reasons in engineering practice, the one with the greater degree of damage at the two beam ends is taken as its damage degree label, such as Fig.10 shown.

[0104] According to the above classification method, 10 of the 40 percussion points are severely injured, 17 are moderately injured, 8 are slightly injured, and 5 are not injured. The injury information is bound to the corresponding audio signal by using a script to automatically generate a configuration file that stores the correspondence between audio and labels, so as to proceed to the next step. After preparing the data and assigning the labels, in order to prevent overfitting and underfitting, this embodiment randomly divides the entire data set into a training set (Training dataset), a validation set (Validation dataset) and a test set (Test dataset) in a ratio of 7:1.5:1.5, and selects a suitable neural network for training.

[0105] When building a neural network architecture, since sound is one-dimensional information, all layers in the neural network are also one-dimensional. Considering the computational complexity of the classification task and the low requirements on the network depth and information extraction capabilities, this embodiment builds a one-dimensional convolutional neural network structure including a convolutional layer, a batch normalization layer, a maximum pooling layer, and a fully connected layer, with reference to the existing convolutional neural network architecture dedicated to speech recognition. Their functions are as follows:

[0106] Convolutional Layer: As the core structure of convolutional neural networks, it realizes local perception of input data by sliding a small window (called a convolution kernel or filter) on the input data and performing calculations. Each convolution kernel can detect a specific feature or pattern in the input data, such as edges, colors, textures, etc.

[0107] Batch Normalization Layer: It mainly accelerates neural network training by standardizing the input data of each batch, reducing the complexity of initialization weights, and making the network more robust to different initialization weights.

[0108] Max Pooling Layer: It is a downsampling technique that reduces the computational complexity of the neural network and prevents overfitting by sliding a small window over the input data and then taking the maximum value in each window as the output.

[0109] Fully Connected Layer: Mainly used to process high-level features in the network. Fully connected layers are usually located in the last few layers or output layers in the entire network architecture, and are used to achieve the final mapping from feature space to target space.

[0110] The network built Fig.15 As shown in the figure, the module of convolution layer + batch normalization layer + pooling layer is repeated four times in the whole network, and each module is linearly connected; only the number of data channels is different among them. Finally, a fully connected layer realizes the linear association of the previous layer and outputs the classification result (expressed as the classification scores of three structural damage categories, and the highest score is taken as the classification result); the loss function of the neural network is the negative log-likelihood loss function (Negative Log-Likelihood Loss, NLL Loss).

[0111] This embodiment uses a sound signal data set to train the neural network on the one hand, and uses an acceleration signal data set to train the neural network on the other hand. When the sound signal data set is used to train the neural network, the loss value decreases normally in the initial stage, but the accuracy does not begin to increase significantly until 10 rounds of training. After about 30 minutes and 60 rounds of training (since the frequency of loss value recording is higher during the training process, the horizontal axis range shows about 350, but it actually corresponds to 60 rounds), the loss value and accuracy (the ratio of the number of correct predictions to the number of the entire test set) have stabilized and converged. The parameter image of the whole process is shown in the figure. Figures 16a-16b shown.

[0112] Table 1 shows the performance of different precision, recall and F-1 scores for three different damages. It can be seen that the performance of several indicators is satisfactory. Fig.17 The confusion matrix obtained by running the final trained model on the test set is shown. It can be observed that the accuracy is high, the number of classification errors is very low, the model performance is good, and different feature information can be accurately extracted; and the classification prediction task of the test data set consisting of 360 percussion signals only took 3.5 seconds, with low computing power requirements, and can be widely deployed on computers of different performances, and even achieve real-time injury detection and classification.

[0113] Table 1

[0114]

[0115] When using acceleration signal data sets for neural network training, although acceleration signals are different from sound signals, they are similar mechanical wave signals, so the same strategy of extracting their MFCC parameters as the final training data can be adopted. The acquisition frequency of acceleration data is 1024Hz, so the number of parameters extracted according to the MFCC extraction strategy for sound is:

[0116] 2400×1024 / 512×13=62,400

[0117] Prepare the entire dataset in the same proportion of training, validation, and test sets. Since the dataset is small, the training only took about 15 minutes to reach convergence. The changes in accuracy and loss value are shown in the figure below. Figures 18a-18b shown.

[0118] Table 2 shows its three performance indicators; Fig.19 The confusion matrix obtained by testing the model on the acceleration test set; similar to the neural network obtained using sound signals as the dataset, the network has good accuracy.

[0119] Table 2

[0120]

[0121] Comparing the results of the two deep learning models, it can be seen that the acceleration signal's ability to reflect the intrinsic characteristics of the structure is slightly worse than that of the sound signal. The reason may be that the acceleration sensor is not firmly fixed on the structure, resulting in the inability to accurately capture the vibration signal generated by the knocking, thereby reducing the ability to distinguish between two signals with similar degrees of damage. But in general, the accuracy of the two methods is very high, and they can complement each other in practical applications. Both have broad application prospects. Using the two deep learning models together as a damage state detection model can improve the effectiveness of on-site detection. Using the output of the two results together as the basis for structural damage state assessment can also avoid the distortion caused by factors such as data acquisition environment and equipment when one of the deep learning models is actually applied.

[0122] In order to verify the effectiveness of this solution, this embodiment performs structural reinforcement and compares the damage state identification results before and after reinforcement. First, the structural node reinforcement and data are re-acquired. In order to further verify the feasibility of this solution, the damaged node positions of the frame structure are reinforced and repaired, and extended experimental research is carried out. The reinforcement method mainly involves chiseling out the concrete at the cracked nodes, and then using the existing composite modified cement-based repair mortar material to re-form and repair, wait for solidification, and retain 3 standard test blocks made of mortar for measuring their strength. After solidification is completed, repeat the aforementioned percussion test steps, use the newly collected data as the test dataset, and apply the previously trained neural network model to predict its degree of damage.

[0123] After reinforcement, repeat the test process before reinforcement, and perform sampling operation at the same sampling point with the same sampling method; when sampling, ensure that the same hammer head, hammer force and hammer position as the previous test are used, and record acceleration and sound signals at the same time; because the data collected from the reinforced structure is not used to train the neural network, each tapping point is only tapped 15 times; after conversion into the corresponding Mel cepstral coefficients, the total amount of data collected is as follows:

[0124] For sound data (sampling rate 44.1kHz):

[0125] 15×40×44100 / 512×13=671,836

[0126] For acceleration data (sampling rate 1024Hz):

[0127] 15×40×1024 / 512×13=15,600

[0128] The collected and processed data were input into the trained deep learning classification model to obtain the predicted classification results; 11 of the 16 nodes were predicted by the model to be undamaged and 5 to be slightly damaged.

[0129] The damage comparison of each node before and after reinforcement is shown in Table 4:

[0130] Table 4

[0131]

[0132]

[0133] In summary, this scheme combines percussion method, Mel-frequency cepstral coefficients and deep learning technology to study and explore the application of damage classification of frame structures. In the embodiment, relying on the test structure - a five-story scaled frame model, as well as the simulated earthquake motion and damage it has experienced, it is designed to calculate the Mel-frequency cepstral coefficients of the percussion sound and the acceleration signal according to certain rules and form them into a data set, and then correspond the damage conditions of the corresponding frame nodes to the data set; then, the MFCCs of the sound signal and the MFCCs of the acceleration signal are used to train a small convolutional neural network, and its performance is tested and verified; finally, the damaged nodes of the prototype structure are reinforced and repaired, and the sound data is recollected from the repaired structure to apply the deep learning model obtained in the previous article to predict the new degree of damage, and compared with the degree of damage of the corresponding node before reinforcement. Through the above research, the following conclusions can be drawn:

[0134] (1) Using MFCC as the features of sound and acceleration signals to form a data set to train a neural network model can achieve high classification accuracy without requiring high computing power. It only takes about 3 to 4 seconds to predict 360 sound signals. The high efficiency and low computing power requirements enable it to be deployed on mobile devices for real-time reasoning, detection and classification.

[0135] (2) In this embodiment, the convolutional neural network model trained by using the MFCC data set calculated from the sound signal has slightly better performance than the model obtained from the acceleration signal, and is about 3 percentage points higher in terms of overall accuracy (the ratio of the number of correct predictions to the total number of test sets);

[0136] (3) After the frame nodes were reinforced, they were retested, and the results showed that the damage to most nodes was significantly reduced, which was consistent with the actual situation. The response of the repaired structure was collected using a high-precision accelerometer and the dynamic characteristics were analyzed. Compared with before reinforcement, the natural frequency was significantly increased, which was the result of the improvement of the structural stiffness, proving the effectiveness of the reinforcement and repair work.

[0137] In addition, the lightweight characteristics of the model are worth further exploitation. Thanks to the rapid improvement in hardware equipment performance and the significant increase in portability, real-time structural damage detection based on this solution has become possible. The development of stable, accurate and robust real-time detection algorithms will add powerful intelligent tools to the field of structural damage detection.

Claims

1. A concrete structure damage detection system based on percussion sound and acceleration information, characterized in that: The percussion hammer comprises a percussion hammer, a microphone, an acceleration sensor, a data acquisition instrument and a processor, wherein the microphone and the acceleration sensor are respectively connected to the data acquisition instrument, the data acquisition instrument is connected to the processor, and the percussion hammer is used to percuss the concrete structure; The microphone and acceleration sensor are used to collect vibration information during the percussion process, including percussion sound information and acceleration information; The data acquisition device is used to transmit the collected percussion information to the processor; The processor is provided with a pre-trained damage detection model, and the vibration information is pre-processed and then input into the damage detection model, and the corresponding damage state detection result is output.

2. A concrete structure damage detection method based on percussion sound and acceleration information, characterized in that: The following steps are involved: S1. Collect the percussion sound and acceleration information of the concrete structure and perform data processing to construct a data set, wherein the data processing process includes cleaning, denoising, normalization, fast Fourier transform, feature extraction using MFCC, and damage state association; S2. Divide the data set into a training set, a validation set, and a test set according to a preset ratio, and use the training set, the validation set, and the test set to train, validate, and test the constructed neural network architecture in turn to establish a damage status detection model; S3. Obtain the percussion sound and acceleration information of the current concrete structure, input the data into the damage state detection model after data processing, and output the damage state detection result.

3. A concrete structure damage detection method based on percussion sound and acceleration information according to claim 2, characterized in that: The step S1 comprises the following steps: S11, tapping a specific location on the concrete structure for a set number of times using the percussion method, while collecting sound and acceleration information; S12, after cleaning, denoising and normalizing the collected sound and acceleration information, fast Fourier transform is performed respectively, and the corresponding MFCC is extracted; S13, associating the MFCC corresponding to the sound information and the data label corresponding to the damage state of the specific position of the concrete structure to construct a sound information data set; The MFCC corresponding to the acceleration information and the data labels corresponding to the damage status of specific positions of the concrete structure are associated to construct the acceleration information dataset.

4. A concrete structure damage detection method based on percussion sound and acceleration information according to claim 3, characterized in that: The MFCC extraction process in step S12 includes: processing the spectrum obtained after fast Fourier transform through a Mel filter bank, the Mel filter bank includes a plurality of filters distributed on different Mel scales, and the output of each filter is the sum of the power spectrum of the frequency range covered by it; Then, the output of the Mel filter bank is logarithmized, and the output of the Mel filter bank after logarithmization is subjected to DCT processing, and the first N coefficients of the DCT result are selected as MFCC.

5. The concrete structure damage detection method based on percussion sound and acceleration information according to claim 3, characterized in that: The data labels corresponding to the damage status of the specific positions of the concrete structure in step S13 include four damage categories: no damage, slight damage, moderate damage, and severe damage.

6. A concrete structure damage detection method based on percussion sound and acceleration information according to claim 5, characterized in that: The damage-free category corresponds to no cracks or only a few tiny cracks; The minor damage category corresponds to certain fine cracks with a width of less than 1 mm; The moderate damage category corresponds to crack width greater than 1 mm or a small amount of concrete spalling; The severe damage category corresponds to obvious concrete spalling or exposed steel bars.

7. A concrete structure damage detection method based on percussion sound and acceleration information according to claim 3, characterized in that: The step S2 comprises the following steps: S21, respectively using the training set, test set, and verification set of acceleration information and sound information to train, test, and verify the constructed neural network architecture to obtain a first deep learning model and a second deep learning model; S22. Use the first deep learning model and the second deep learning model together as a damage status detection model.

8. The concrete structure damage detection method based on percussion sound and acceleration information according to claim 7, characterized in that: The neural network architecture constructed includes a convolutional layer, a batch normalization layer, a maximum pooling layer and a fully connected layer, wherein the association module of the convolutional layer + batch normalization layer + pooling layer is repeated four times in the entire network, each association module is linearly connected, and finally a fully connected layer realizes the linear association of the previous layer and outputs the classification result: the classification score of different injury categories, and the highest score is taken as the classification result; the loss function of the neural network is the negative log-likelihood loss function.

9. The method for detecting concrete structure damage based on percussion sound and acceleration information according to claim 3, characterized in that: The step S3 specifically adopts the same method as step S1 to obtain the percussion sound and acceleration information of the current concrete structure, and then performs windowing, framing and fast Fourier transformation on the percussion sound and acceleration information to obtain a spectrum diagram, and performs filtering, logarithm taking and DCT processing on the spectrum to obtain corresponding MFCC data, and inputs the MFCC data into the damage state detection model to output the damage state detection result.

Images