FRP-concrete interface stripping early warning method based on acoustic emission and deep learning technology

Abstract

The invention discloses an FRP-concrete interface stripping early warning method based on acoustic emission and deep learning technologies, and belongs to the technical field of FRP-concrete interface stripping detection. The method comprises the following steps: periodically acquiring dynamic acoustic emission signals through a plurality of acoustic emission sensors arranged on a to-be-monitored interface area of the FRP-concrete structure, and constructing a first data set; constructing a background signal dynamic threshold for performing signal anomaly identification on the data in the first data set, and extracting the data marked as an anomaly identification result to construct a second data set; preprocessing the data in the second data set to obtain a plurality of target signals, and performing time domain feature extraction to obtain time frequency feature parameters; inputting the time-frequency characteristic parameters into a stripping state prediction model based on deep learning, and outputting to obtain the stripping state of the current FRP-concrete interface; and sending a corresponding early warning signal according to the stripping state. According to the invention, real-time monitoring and accurate early warning of FRP-concrete interface stripping can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of FRP-concrete interface peeling detection, and particularly relates to a method for warning FRP-concrete interface peeling based on acoustic emission and deep learning technologies. Background Technique

[0002] FRP (fiber reinforced composite material) has been widely used in the field of concrete structure reinforcement due to its advantages of high strength, corrosion resistance, light weight, etc. However, the interfacial bonding performance between FRP and concrete is a key factor affecting the reinforcement effect, and interface peeling is one of the common failure forms. The occurrence of interface peeling will cause FRP to be unable to effectively transfer loads, reduce the bearing capacity and durability of the structure, and even may lead to structural safety accidents. Therefore, how to accurately monitor FRP-concrete interface peeling has become a difficult problem. [[ID=IO]]

[0003] To solve the above problems, the prior art usually adopts methods such as visual inspection, tapping method and ultrasonic inspection for interface peeling detection, but these methods usually have the disadvantages of low detection efficiency (visual inspection), strong subjectivity (tapping method) and insensitivity to micro peeling (ultrasonic inspection), and it is difficult to achieve accurate monitoring of interface peeling. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method for warning FRP-concrete interface peeling based on acoustic emission and deep learning technologies to solve the above technical problems.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: A method for warning FRP-concrete interface peeling based on acoustic emission and deep learning technologies, comprising: Periodically collecting dynamic acoustic emission signals through a plurality of acoustic emission sensors arranged on the area of the to-be-monitored interface of the FRP-concrete structure, and constructing a first data set; Constructing a background signal dynamic threshold for identifying signal anomalies in the data in the first data set, and extracting the data marked as the anomaly identification result to construct a second data set; Preprocessing the data in the second data set, obtaining a number of target signals for time domain feature extraction, and obtaining time-frequency feature parameters; Inputting the time-frequency feature parameters into a peeling state prediction model based on deep learning, and outputting the peeling state of the current FRP-concrete interface; and sending a corresponding warning signal according to the peeling state.

[0006] Further, a method for warning FRP-concrete interface peeling based on acoustic emission and deep learning technologies further comprises: There are at least three acoustic emission sensors deployed on the interface area to be monitored of the FRP-concrete structure, and there is at least one acoustic emission sensor at both ends and the middle position of the interface area to be monitored.

[0007] Furthermore, constructing the background signal dynamic threshold includes: Step 1: When the interface to be monitored of the FRP-concrete structure is in a normal state, collect the acoustic emission signals during each peak working period and non-peak working period through the deployed acoustic emission sensors, and construct multiple background signal sets and a second background signal set; among them, each background signal set corresponds to a peak working period, and each second background signal set corresponds to a non-peak working period. Step 2: Conduct outlier analysis on the signal frequencies in any one of the background signal sets to obtain the concentrated recorded values of the signal frequencies, extract the maximum value of the concentrated recorded values, and set it as the initial background signal threshold; at the same time, conduct outlier analysis on the signal frequencies in any one of the second background signal sets to obtain the second concentrated recorded values of the signal frequencies, extract the maximum value of the second concentrated recorded values, and set it as the second background signal threshold. Step 3: Repeat Step 2 until the threshold acquisition for all background signal sets and all second background signal sets is completed, and obtain several initial background signal thresholds and second background signal thresholds. Step 4: Mark each initial background signal threshold with the peak working period to obtain several peak background signal thresholds; and extract the maximum value of all second background signal thresholds to obtain the flat background signal threshold. Step 5: Obtain the peak background signal threshold and the flat background signal threshold, and construct the background signal dynamic threshold applicable to any time period.

[0008] Furthermore, constructing the background signal dynamic threshold also includes: S201: When it is impossible to determine whether the interface to be monitored of the FRP-concrete structure is in a normal state, collect the acoustic emission signals through the deployed acoustic emission sensors to construct multiple background signal sets and a second background signal set. S202: Input any one of the background signal sets or the second background signal set into the initial enhancement system, and perform filtering and enhancement processing through the signal enhancement module in the initial enhancement system to obtain a signal enhancement sample; among them, the signal enhancement module includes multiple amplification filter units, and each amplification filter unit corresponds to a sensitive feature frequency band of a background signal type. S203. Input the signal enhancement samples into the shared encoder of the initial enhancement system to extract sensitive features in different sensitive feature frequency bands and input them into the recognition module of the initial enhancement system for classification. Obtain the background signal classification result, mark the corresponding sensitive feature frequency bands, and obtain the marked data; S204. Obtain the background signal type corresponding to the acquisition time period of the signal enhancement samples, perform similarity analysis on the category and quantity distribution with the marked data. If the analysis result meets the preset similarity condition, construct a new background signal set or a new second background signal set according to the marked data; S205. For the new background signal set or the new second background signal set, perform the operations in steps two to five to construct a background signal dynamic threshold applicable to any time period.

[0009] Furthermore, the signal anomaly recognition for the data in the first dataset includes: Obtain the periodic acquisition marks of the first dataset; wherein, the periodic acquisition marks include the working peak period time marks and the non - working peak period time marks; Based on the periodic acquisition marks, determine the target value of the background signal dynamic threshold, and screen the data in the first dataset according to the target value to obtain the data with signal frequency greater than the target value, and mark the anomaly recognition result.

[0010] Furthermore, the pre - processing includes removing high - frequency interference signals from the data in the second dataset and performing signal enhancement operations on the data in the second dataset after the removal operation.

[0011] Furthermore, the time - frequency feature parameters include: time - domain features, frequency - domain features, and time - frequency domain features; The time - domain features include peak value, rise time, duration, and ringing count; The frequency - domain features include main frequency and energy spectral density; The time - frequency domain features include the central frequency of the time - frequency energy concentration region, energy proportion, time span, and entropy value of the time - frequency energy distribution.

[0012] Furthermore, the construction process of the peeling state prediction model includes: Obtain the training data; wherein, the training data includes several feature data groups with peeling state marks, and each feature data group includes several time - frequency feature parameters; Construct an initial convolutional neural network model, use the time - frequency feature parameters as the input and the peeling state mark as the output to train the initial convolutional neural network model until the preset training conditions are met, and obtain the peeling state prediction model.

[0013] The beneficial effects of the present invention are as follows: The present invention provides an FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology, which combines acoustic emission technology and deep learning technology. Among them, the acoustic emission technology can collect acoustic emission signals generated during the interface peeling process in real time, reflecting the dynamic process of interface peeling; the deep learning technology can automatically extract the deep characteristics of the signal, and improve the recognition and classification accuracy of the interface peeling state. It is beneficial to solve the problems of low detection efficiency, strong subjectivity and low early warning accuracy in existing detection methods, and realizes real-time monitoring and accurate early warning of FRP-concrete interface peeling.

[0014] Other advantages, objectives, and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or may be taught by those skilled in the art from the practice of the present invention. The purposes and other advantages of the present invention may be realized and obtained through the structures particularly pointed out in the written description and the accompanying drawings.

[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a method for early warning of FRP-concrete interface delamination based on acoustic emission and deep learning technology in an embodiment of the present invention; Figure 2 This is a flow chart of the process of constructing a dynamic threshold value of a background signal in an FRP-concrete interface delamination early warning method based on acoustic emission and deep learning technology in an embodiment of the present invention; Figure 3 This is another flow chart of the process of constructing the dynamic threshold of the background signal in an FRP-concrete interface delamination early warning method based on acoustic emission and deep learning technology in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0018] like Figure 1 As shown, the present invention proposes an FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology, comprising: S101, periodically collecting dynamic acoustic emission signals by multiple acoustic emission sensors arranged on the interface area to be monitored of the FRP-concrete structure to construct a first data set; S102. Construct a dynamic threshold for the background signal to identify signal anomalies in the data in the first dataset, and extract the data marked as the anomaly identification result to construct a second dataset; S103. Preprocess the data in the second dataset to obtain a number of target signals for time-domain feature extraction to obtain time-frequency feature parameters; S104. Input the time-frequency feature parameters into the peeling state prediction model based on deep learning, and output the peeling state of the current FRP-concrete interface; S105. Send a corresponding warning signal according to the peeling state; The beneficial effects of the above technical solution are as follows: Compared with the traditional detection method, the peeling state detection by acoustic emission technology can achieve real-time warning, that is, signal anomalies can also be detected at the initial stage of peeling. At the same time, the specific peeling development stage can be judged by the change of signal parameters. Combining with the peeling state prediction model based on deep learning, it can avoid the subjectivity of manual judgment of signal parameter changes, further improve the prediction efficiency of the peeling development state, and improve the warning accuracy of the peeling state.

[0019] In one embodiment, a method for warning of FRP-concrete interface peeling based on acoustic emission and deep learning technology further includes: there are at least three acoustic emission sensors arranged on the interface area to be monitored of the FRP-concrete structure, and there is at least one acoustic emission sensor at both ends and the middle position of the interface area to be monitored; The working principle and beneficial effects of the above technical solution are as follows: A plurality of acoustic emission sensors are arranged in the interface area to be monitored of the FRP-concrete structure, and the sensors are evenly distributed to ensure that the acoustic emission signals generated during the interface peeling process can be comprehensively collected. The number of sensors arranged is determined according to the size, shape and easily peeled position of the interface to be detected, but usually not less than three, and they are arranged at both ends and the middle position of the interface area to be monitored respectively to realize the positioning and source identification of the acoustic emission signals during the monitoring process. Optionally, if there is an easily peeled situation at the center position in the interface area to be monitored of the current FRP-concrete structure, four acoustic emission sensors are evenly arranged in the interface area to be monitored of the FRP-concrete structure, with one arranged at each end and two symmetrically arranged at the middle position to ensure accurate monitoring and positioning of the peeling state and position.

[0020] In one embodiment, constructing the dynamic threshold of the background signal includes: Step 1: When the monitored interface of the FRP-concrete structure is in a normal state, the acoustic emission sensors that have been arranged are used to collect the acoustic emission signals during each peak working period and non-peak working period respectively, and multiple background signal sets and a second background signal set are constructed; among them, each background signal set corresponds to a peak working period, and each second background signal set corresponds to a non-peak working period; Step 2: Perform outlier analysis on the signal frequencies in any one of the background signal sets to obtain the concentrated recorded values of the signal frequencies, extract the maximum value of the concentrated recorded values, and set it as the initial background signal threshold; at the same time, perform outlier analysis on the signal frequencies in any one of the second background signal sets to obtain the second concentrated recorded values of the signal frequencies, extract the maximum value of the second concentrated recorded values, and set it as the second background signal threshold; Step 3: Repeat Step 2 until the threshold values are obtained for all background signal sets and all second background signal sets, and a number of initial background signal thresholds and second background signal thresholds are obtained; Step 4: Mark each initial background signal threshold with the peak working period to obtain a number of peak background signal thresholds; and extract the maximum values of all second background signal thresholds to obtain the flat background signal threshold; Step 5: Obtain the peak background signal threshold and the flat background signal threshold, and construct a dynamic background signal threshold applicable to any time period; The working principle of the above technical solution is as follows: Figure 2As shown, when the interface to be monitored of the FRP-concrete structure is in a normal state, the acoustic emission sensors arranged are used to collect the acoustic emission signals during each peak working period and non-peak working period respectively, and multiple background signal sets and multiple second background signal sets are constructed to reflect the acoustic emission signals such as mechanical vibrations generated by the current concrete environment and the inherent vibrations of the material under normal conditions. It should be noted that each time of collection needs to be consistent with the collection time in the corresponding collection period in the first data set to ensure that the acquired data can accurately reflect the background state during real-time monitoring. In addition, multiple background signal sets and multiple second background signal sets can include collections of repeated periods, that is, for the same peak working period or non-peak working period, there are multiple simultaneous collections on different dates. For example, for the morning peak period from 7:00 to 9:00, during this period, the traffic volume of external vehicles is large, and the generated mechanical vibrations are numerous and complex. The corresponding acoustic emission signals are collected in batches during the 7:00-9:00 period on March 21, March 22, and March 25 respectively. The acoustic emission signals include signals such as mechanical vibrations caused by environmental reasons and the inherent vibrations of the material, and are used to construct the background signal set;After obtaining a sufficient amount of background signal data, traversal analysis is performed on the background signal set and the second background signal set respectively. Among them, for the background signal set, since there are obvious differences in the main environmental signals faced during each peak working period. For example, the mechanical vibrations caused by vehicles during the morning rush hour, the mechanical vibrations caused by factory machines during working hours, etc. Therefore, for different peak working periods, their initial background signal thresholds are determined respectively. For example, outlier analysis is performed on the signal frequencies in the first background signal set to obtain the concentrated recorded values of the signal frequencies, preventing a small number of high-frequency signals such as vehicle collisions and braking from interfering with the overall signal. Finally, the maximum value of the concentrated recorded values is extracted and set as the initial background signal threshold. It should be noted that this background signal set corresponds to a peak working period. Assuming that this background signal set corresponds to the period from 7:00 to 9:00, all the acoustic emission signals collected during the period from 7:00 to 9:00 on March 21st, March 22nd, and March 25th belong to this background signal set and are used for outlier analysis. Finally, the maximum value is obtained as the critical value, that is, the initial background signal threshold. Finally, it is marked according to the peak working period to obtain the peak background signal threshold corresponding to this period. It should be noted that the reason for using the maximum value as the critical value is that the acoustic emission signals generated during the stripping period are all high-frequency signals. Therefore, using the maximum value as the critical value can reduce the interference of subsequent background signals on the abnormal recognition result compared with using the average value method. At the same time, the purpose of performing outlier analysis is to make the obtained data more representative, improve the reliability of the obtained initial background signal threshold, and reduce the interference of a small number of high-frequency signals such as vehicle collisions and braking on the overall signal. Finally, the processing of N background signal sets is gradually completed to obtain N peak background signal thresholds, where N refers to the number of peak working periods; for non-peak working periods, the mechanical vibrations caused by environmental reasons do not have an obvious bias. Therefore, for the data during this period, the total second background signal threshold for non-peak working periods is determined according to all the collected second background signal sets. For example, outlier analysis is performed on the signal frequencies in the first second background signal set to obtain the second concentrated recorded value of the signal frequencies, extract the maximum value of the second concentrated recorded value, and set it as the second background signal threshold. Repeat this operation for each second background signal set until the processing of the Mth second background signal set is completed to obtain M second background signal thresholds. Finally, the maximum value of all the second background signal thresholds is extracted to obtain the average background signal threshold. It should be noted that M represents the number of second background signal sets. In addition, compared with directly performing total outlier analysis on the data in all the second background signal sets and extracting the maximum value, performing separate outlier analysis on the data collected in the same batch is more accurate in removing abnormal data points; finally, based on the peak background signal threshold and the average background signal threshold, a background signal dynamic threshold applicable to any time period is constructed; The beneficial effects of the above technical solution are as follows: Through the above technical solution, background signals are collected under the normal working state of the structure, and a dynamic threshold of the background signal is constructed for subsequent judgment of abnormal signals that can reflect the peeling state. Compared with the traditional technical means of using a fixed high-frequency acquisition frequency or a fixed signal threshold, it overcomes the problem that it is difficult to adapt to mechanical vibrations, normal vibrations of the structure, and other factors that cause signal fluctuations due to external environmental reasons, which are likely to cause misjudgment or missed judgment. When it can be determined that the monitored interface of the current FRP-concrete structure is in a normal state, compared with the technical means of directly using fixed-frequency feature recognition, the recognition method of the technical solution provided in this application is simpler and does not require a complex feature recognition process. At the same time, a dynamic threshold of the background signal applicable to any time period is set, so that the signal abnormality determination standard can be automatically matched according to the differences in objective factors such as the environmental background signal in different time periods, ensuring that at any moment, the abnormal signal recognition method proposed in this application can accurately identify abnormal signals, which is beneficial to improving the real-time performance and accuracy of interface peeling warning.

[0021] As Figure 3 shown, in one embodiment, constructing the dynamic threshold of the background signal further includes: S201. When it is impossible to determine whether the monitored interface of the FRP-concrete structure is in a normal state, acoustic emission signals are collected through the arranged acoustic emission sensors to construct a plurality of background signal sets and a second background signal set; S202. Input any one of the background signal sets or the second background signal set into the initial enhancement system, and perform filtering and enhancement processing through the signal enhancement module in the initial enhancement system to obtain signal enhancement samples. Among them, the signal enhancement module includes a plurality of amplification filter units, and each amplification filter unit corresponds to a sensitive feature frequency band of a background signal type; S203. Input the signal enhancement samples into the shared encoder of the initial enhancement system to extract sensitive features of different sensitive feature frequency bands and input them into the recognition module of the initial enhancement system for classification, obtain the background signal classification result, mark the corresponding sensitive feature frequency band, and obtain marked data; S204. Obtain the background signal type corresponding to the acquisition time period of the signal enhancement samples, perform similarity analysis on the category and quantity distribution with the marked data. If the analysis result meets the preset similarity condition, construct a new background signal set or a new second background signal set according to the marked data; S205. For the new background signal set or the new second background signal set, perform the operations of steps two to five to construct a dynamic threshold of the background signal applicable to any time period; The working principle and beneficial effects of the above technical solution are as follows: In the actual application scenarios outside the laboratory, except that it is easy to determine whether the monitored interface of the FRP-concrete structure is in a normal state at the initial stage of the FRP-concrete adhesion, generally, it is still necessary to use visual inspection or the tapping method to determine whether the monitored interface of the FRP-concrete structure is in a normal state. However, this method seriously depends on the inspection experience of the inspectors. If the inspectors lack experience and give incorrect state judgment results, it will lead to errors in the background signal dynamic threshold constructed in the above technical solution, resulting in a decrease in the prediction accuracy of the peeling state in the subsequent process; Based on this, in order to improve the applicability of this application in different service life of the FRP-concrete interface, this application provides another construction scheme for the background signal dynamic threshold to solve the above technical problems; Specifically, based on the multi-source separability of the acoustic emission signal, this technical solution completes the separation of different sound source signals, and peels off reliable background acoustic emission signals for constructing the background signal dynamic threshold; First, when the user cannot determine whether the monitored interface of the FRP-concrete structure is in a normal state, multiple background signal sets and a second background signal set are constructed by collecting acoustic emission signals through the deployed acoustic emission sensors. Here, the background signal sets and the second background signal set are the same as those in the foregoing technical solution. That is, each background signal set corresponds to a peak working period, and each second background signal set corresponds to a non-peak working period. Then, the background signal sets and the second background signal set are input into the initial enhancement system for processing. Among them, the construction of the initial enhancement system includes functions such as a signal enhancement module, a shared encoder, and an identification module. Taking any background signal set or second background signal set as an example, the corresponding input background signal set or second background signal set is filtered and enhanced by the signal enhancement module in the initial enhancement system to obtain a signal enhancement sample. Among them, the signal enhancement module includes multiple amplification filter units, and each amplification filter unit corresponds to a sensitive feature frequency band of a background signal type. The background signal types include environmental signals and inherent material vibration signals, etc. Since there are significant differences in the frequency domain distribution and time-frequency characteristics of different environmental vibrations and inherent material vibrations, after obtaining the background signal set or the second background signal set, the main frequency bands of different signals are identified by calculating the spectral energy distribution and amplified. Finally, all the amplification results are integrated to obtain a signal enhancement sample. It should be noted that the signals described here are regular background signals with periodic fluctuations, such as the sound of rain or mechanical vibrations caused by cars, as well as the inherent vibrations of materials, rather than irregular, non-periodic pulse noises and random white noises, etc. After obtaining the signal enhancement sample, it is input into the shared encoder of the initial enhancement system to extract the sensitive features of different sensitive feature frequency bands. The sensitive features are the specific parts of this type of signal in terms of frequency domain distribution and time-frequency characteristics relative to other signals, which are used to be input into the identification module for identification and classification. Finally, the background signal classification result is obtained to mark the corresponding sensitive feature frequency band, and marked data is obtained. After obtaining the labeled data, in order to improve the accuracy of the extraction results, this technical solution also designs a dual judgment mechanism, that is, to obtain the background signal type corresponding to the acquisition time period of the signal enhancement sample. The acquisition of this background signal type can be input in advance by the user, or can be determined by multiple detections and similarity discrimination screening for each acquisition time period corresponding background signal type. The acquisition time period includes the peak working period and the non-peak working period. Based on the background signal type, a similarity analysis of the category and quantity distribution is performed with the labeled data, that is, a similarity result analysis is performed on the categories included in the background signal type and the corresponding categories existing in the labeled data, and at the same time, a similarity analysis is performed on the category quantity in the background signal type and the corresponding category quantity in the labeled data to ensure that there is no misdetection or missed detection. When the analysis result meets the preset similarity condition, a new background signal set or a new second background signal set is constructed according to the labeled data; usually, the preset similarity condition is preferably determined that the corresponding categories existing in the labeled data can include the categories included in the background signal type, and the similarity between the category quantity in the background signal type and the corresponding category quantity in the labeled data reaches more than 85%; finally, for the new background signal set or the new second background signal set, the operations of steps two to five are performed to construct a background signal dynamic threshold applicable to any time period.

[0022] In one embodiment, the signal anomaly recognition for the data in the first dataset includes: Obtain the periodic acquisition mark of the first dataset; wherein, the periodic acquisition mark includes the peak working period mark and the non-peak working period mark; Based on the periodic acquisition mark, determine the target value of the background signal dynamic threshold, and screen the data in the first dataset according to the target value to obtain the data with a signal frequency greater than the target value, and mark the anomaly recognition result; The working principle of the above technical solution is as follows: Multiple acoustic emission sensors arranged on the interface area to be monitored of the FRP-concrete structure periodically collect dynamic acoustic emission signals to construct a first data set, and perform periodic acquisition marking on the first data. Here, periodic acquisition includes performing an acquisition every preset time period, such as 30 days, and performing multiple acquisitions at different time periods on the day of acquisition to ensure its accuracy. The periodic acquisition marking includes marking the specific working peak period and non-working peak period of multiple acquisitions on the day of acquisition. Since, under normal conditions, when collecting acoustic emission signals from the FRP-concrete interface of the concrete structure, the signals usually show a stable state of low intensity, low frequency, and low activity. At the same time, due to the relatively long distance between the sound source of the environmental signal and the acoustic emission sensor, the collected part of the signal is usually also in a state of low frequency and low intensity. However, the FRP-concrete interface in the peeling state usually generates high-frequency acoustic emission signals during the peeling process. Based on this, the present application determines the target value of the background signal dynamic threshold as the critical value through periodic acquisition marking, and screens the data in the first data set according to the critical value to obtain the data with a signal frequency greater than the critical value as the acoustic emission signal reflecting the peeling state, and performs abnormal recognition result marking for subsequent peeling state prediction; The beneficial effect of the above technical solution is as follows: Through the above technical solution, before predicting the peeling state, the data participating in the prediction is identified in advance, and the acoustic emission signals that are abnormal and can reflect the specific peeling state are accurately screened out, which is beneficial to improving the prediction accuracy of the subsequent peeling state.

[0023] In one embodiment, the preprocessing includes removing high-frequency interference signals from the data in the second data set, and performing a signal enhancement operation on the data in the second data set after the removal operation is completed; The working principle and beneficial effect of the above technical solution are as follows: Under normal circumstances, the preprocessing of signals generally includes operations such as filtering, noise reduction, and signal enhancement. However, in the previous solution, we have already screened out the normal background signals and low-frequency noises through the background signal dynamic threshold. Therefore, when preprocessing the second data set, only the high-frequency interference signals such as electromagnetic signals need to be removed, and a signal enhancement operation is performed on the data in the second data set after the removal operation is completed. It should be noted that in the specific preprocessing process, other processing methods such as noise reduction can also be added according to actual needs, which will not be elaborated here.

[0024] In one embodiment, the time-frequency feature parameters include: time domain features, frequency domain features, and time-frequency domain features; The time domain features include peak value, rise time, duration, and ring count; The frequency domain features include main frequency and energy spectral density; The time-frequency domain features include the central frequency of the time-frequency energy concentration region, the energy proportion, the time span, and the entropy value of the time-frequency energy distribution; The working principle and beneficial effects of the above technical solution are as follows: To improve the prediction accuracy of the subsequent peeling state prediction model, when obtaining training data and data to be monitored, this application selects time-frequency feature parameters (time domain features, frequency domain features, and time-frequency domain features) as the feature vectors of the model input. The peeling state prediction model generated by fusing time-frequency feature parameters can more comprehensively and accurately describe the mapping relationship between the signal and the interface peeling state; for example, in the initial stage of the FRP-concrete interface peeling, there are situations such as the shortening of the signal rise time (more severe mutation) and the extension of the duration, and the time domain features can quickly capture such abnormalities; in the severe peeling stage of the FRP-concrete interface, the main frequency of the signal may shift to a lower frequency (structural damage causes the vibration frequency to decrease), and the frequency domain features can quantify this trend; in the development stage of the FRP-concrete interface peeling, the time-frequency domain features may show that specific frequency components continuously appear within a certain time period, reflecting the gradualness and spatial expansion of the peeling. Through multi-dimensional feature complementarity, the prediction accuracy and reliability of the prediction model can be significantly improved; in addition, the peak value, rise time, duration, and ringing count, main frequency and energy spectral density, central frequency of the time-frequency energy concentration region, energy proportion, time span, and entropy value of the time-frequency energy distribution specifically included in the time domain features, frequency domain features, and time-frequency domain features proposed in this application are the main parameter types used in this application for training and monitoring. Further, to improve the prediction accuracy, other relevant time domain feature parameters, frequency domain feature parameters, and time-frequency domain feature parameters can also be included.

[0025] In one embodiment, the construction process of the peeling state prediction model includes: Obtain training data; where the training data includes several feature data groups with peeling state marks, and each feature data group includes several time-frequency feature parameters; Construct an initial convolutional neural network model, use the time-frequency feature parameters as the input, and the peeling state mark as the output to train the initial convolutional neural network model until the preset training conditions are completed to obtain the peeling state prediction model; The working principle and beneficial effects of the above technical solution are as follows: Acoustic emission technology is a method for dynamically monitoring internal damage of materials. It is used to judge the occurrence and development of damage by detecting elastic wave signals released by materials during the damage process. However, due to the complexity and multi-source nature of acoustic emission signals, it is difficult for traditional signal analysis methods to accurately extract signal parameters reflecting interface peeling characteristics. The method of simply analyzing acoustic emission signals will result in low early warning accuracy. Based on this, this application combines a peeling state prediction model based on deep learning technology to predict the peeling state of the monitored signal parameters, automatically extracting the deep features of the signals, which can effectively improve the recognition and early warning capabilities for interface peeling. The training process of this model specifically includes: obtaining training data, which includes several groups of feature data with peeling state labels. Each group of feature data includes several 10-dimensional feature vectors. The 10 dimensions include peak value, rise time, duration, and ring count, main frequency and energy spectral density, center frequency of the time-frequency energy concentration region, energy ratio, time span, and entropy value of the time-frequency energy distribution included in time domain features, frequency domain features, and time-frequency domain features; it should be noted that the source of the training data can be real monitored data or simulated data obtained by gradually increasing the load on the interface of a normal FRP-concrete structure; after obtaining sufficient training data, a convolutional neural network model is constructed, including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Among them, the input layer is used to receive 10-dimensional feature vectors. The convolutional layer uses 3 convolutional kernels with sizes of 3, 5, and 7 respectively, for feature extraction of the input feature vectors. The pooling layer uses the maximum pooling method to facilitate reducing the data dimension and computational complexity. The fully connected layer preferably uses 2 hidden layers with the number of neurons being 18 and 64 respectively. The output layer uses the softmax activation function to output the probability distribution of the interface peeling state corresponding to the peeling state label. Usually, there are four peeling states: normal state, initial stage of peeling occurrence, development stage of peeling, and severe peeling stage; during the training process, preferably 70% of the data in the training data is used as training data, and 30% of the data is used as the test set, and the cross-entropy loss function and the stochastic gradient descent algorithm are used for optimization.

[0026] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology, characterized in that: include: A first data set is constructed by periodically collecting dynamic acoustic emission signals through a plurality of acoustic emission sensors arranged on the interface area to be monitored of the FRP-concrete structure; Constructing a dynamic threshold value of background signals for performing signal anomaly identification on the data in the first data set, and extracting data marked as anomaly identification results to construct a second data set; Preprocessing the data in the second data set to obtain a number of target signals for time domain feature extraction to obtain time-frequency feature parameters; The time-frequency characteristic parameters are input into the debonding state prediction model based on deep learning, and the debonding state of the current FRP-concrete interface is output; a corresponding early warning signal is issued according to the debonding state.

2. The FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology according to claim 1 is characterized in that: Also includes: There are no less than three acoustic emission sensors arranged on the interface area to be monitored of the FRP-concrete structure, and there is at least one acoustic emission sensor at both ends and in the middle of the interface area to be monitored.

3. The FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology according to claim 1 is characterized in that: Constructing a dynamic threshold for background signals includes: Step 1: When the interface to be monitored of the FRP-concrete structure is in a normal state, acoustic emission signals in each peak working period and non-peak working period are collected by the deployed acoustic emission sensors to construct multiple background signal sets and second background signal sets; wherein each background signal set corresponds to a peak working period, and each second background signal set corresponds to a non-peak working period; Step 2: Perform an outlier analysis on the signal frequencies in any background signal set to obtain a concentrated record value of the signal frequencies, extract the maximum value of the concentrated record value, and set it as the initial background signal threshold; at the same time, perform an outlier analysis on the signal frequencies in any second background signal set to obtain a second concentrated record value of the signal frequencies, extract the maximum value of the second concentrated record value, and set it as the second background signal threshold; Step 3: Repeat step 2 until the thresholds are acquired for all background signal sets and all second background signal sets, and a number of initial background signal thresholds and second background signal thresholds are obtained; Step 4: Mark each initial background signal threshold with a peak working period to obtain a number of peak background signal thresholds; and extract the maximum value of all second background signal thresholds to obtain a mean background signal threshold; Step 5: Obtain the peak background signal threshold and the mean background signal threshold, and construct a dynamic background signal threshold that is applicable to any time period.

4. The FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology according to claim 3 is characterized in that: Constructing a dynamic threshold for background signals also includes: S201, when it is impossible to determine whether the interface to be monitored of the FRP-concrete structure is in a normal state, collecting acoustic emission signals by the deployed acoustic emission sensors to construct multiple background signal sets and a second background signal set; S202: Inputting any one background signal set or the second background signal set into the initial enhancement system, and performing filtering and enhancement processing through the signal enhancement module in the initial enhancement system to obtain a signal-enhanced sample; wherein the signal enhancement module includes a plurality of amplification filter units, each amplification filter unit corresponding to a sensitive characteristic frequency band of a background signal type; S203: Input the signal enhancement sample into the shared encoder of the initial enhancement system to extract sensitive features of different sensitive feature frequency bands, and input them into the recognition module of the initial enhancement system for classification. The background signal classification result is obtained, and the corresponding sensitive feature frequency bands are marked to obtain marked data; S204: Obtain the background signal type of the signal-enhanced sample corresponding to the acquisition time period, perform similarity analysis on the type and quantity distribution with the labeled data, and if the analysis result meets the preset similarity condition, construct a new background signal set or a new second background signal set based on the labeled data; S205 : Perform the operations of step 2 to step 5 on the new background signal set or the new second background signal set to construct a dynamic background signal threshold applicable to any time period.

5. The FRP-concrete interface peeling early warning method based on acoustic emission and deep learning technology according to claim 1 is characterized in that: Identifying signal anomalies on the data in the first data set includes: Obtain a periodic collection mark of the first data set; wherein the periodic collection mark includes a work peak time period mark and a non-work peak time period mark; The target value of the background signal dynamic threshold is determined based on the periodic acquisition mark, and the data in the first data set is screened according to the target value to obtain data with a signal frequency greater than the target value, and an abnormal recognition result is marked.

6. The FRP-concrete interface debonding early warning method based on acoustic emission and deep learning technology according to claim 1 is characterized in that: The preprocessing includes removing high-frequency interference signals from the data in the second data set, and performing a signal enhancement operation on the data in the second data set after the removal operation is completed.

7. The FRP-concrete interface debonding early warning method based on acoustic emission and deep learning technology according to claim 1 is characterized in that: Time-frequency feature parameters include: time domain features, frequency domain features and time-frequency domain features; Time domain characteristics include peak value, rise time, duration, and ring count; Frequency domain features include dominant frequency and energy spectrum density; The time-frequency domain features include the center frequency, energy proportion, time span and entropy value of the time-frequency energy concentration area.

8. The FRP-concrete interface debonding early warning method based on acoustic emission and deep learning technology according to claim 1 is characterized in that: The process of building the peeling state prediction model includes: Acquire training data; wherein the training data includes a plurality of feature data groups with stripping state marks, and each feature data group includes a plurality of time-frequency feature parameters; An initial convolutional neural network model is constructed, and the time-frequency feature parameters are used as input and the stripping state mark is used as output to train the initial convolutional neural network model until the preset training conditions are completed to obtain a stripping state prediction model.

Images