Impact echo nondestructive testing data cleaning method based on deep learning
By establishing a classification model for shock echo data through deep learning, the problem of automatically cleaning invalid data in nondestructive testing was solved, thereby improving testing accuracy and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-19
- Publication Date
- 2026-04-07
AI Technical Summary
In existing technologies, the methods for collecting and removing invalid data in shock echo nondestructive testing are time-consuming, labor-intensive, and inefficient, and are subject to human intervention, making it difficult to achieve automated data cleaning.
A deep learning method was used to establish a classification model for shock echo data. A ResNet network was trained through transfer learning, and shock echo data in two-dimensional image format was used for classification, eliminating invalid data and retaining valid data.
It enables rapid removal of invalid data, reduces human intervention, and improves the accuracy and efficiency of detecting hidden defects in concrete structures.
Smart Images
Figure CN121808463A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nondestructive testing technology for concrete structures, and in particular to a data cleaning method for impact echo nondestructive testing based on deep learning. Background Technology
[0002] Impact echo nondestructive testing (IDT) is a widely used technique for detecting delamination and voids in concrete structures. Stress waves propagate within the concrete by striking the surface with a steel ball or other excitation source. Contact accelerometers or air-coupled microphones are then used to receive the impact echo data. Traditional IDT methods analyze the collected stress wave signals in the frequency domain using Fourier transform, analyzing structural thickness or defect location based on the peak frequencies corresponding to the impact echo modes. The data collected by impact echo testing is an acceleration signal caused by structural or aerodynamic vibrations. The quality of the signal data collected by the sensor affects the accuracy of the analysis results. Detection is invalid when the acceleration signal does not contain any structural information. It is essential to ensure that the impact echo signal used for the peak frequency method effectively reflects delamination defects or boundary information in bridge decks. The use of automated equipment and multi-sensor devices makes it impossible for inspectors to accurately monitor the coupling status and other operational status of sensors, providing conditions for the collection of invalid impact echo data. At the same time, the increased density of inspection points brought about by scanning coverage has resulted in tens of thousands of impact echo signals being collected. However, there is still a lack of convenient methods to remove this invalid data. In actual bridge inspection, manual selection is often used, which is time-consuming, labor-intensive, and inefficient. In addition, there is the influence of human intervention, which has become the reason why inspection departments are reluctant to implement impact echo testing.
[0003] In recent years, deep learning methods have demonstrated excellent performance in analyzing large amounts of data. Deep learning models require fewer defined parameters, and well-trained deep learning models can automate data classification. This invention aims to address the problem of developing a data cleaning method that uses deep learning to assess the quality and automatically select data from collected impact echo data, thereby eliminating invalid data and achieving accurate concrete defect detection. Summary of the Invention
[0004] The purpose of this application is to provide a deep learning-based method for cleaning shock echo nondestructive testing data, which can quickly remove invalid shock echo data and reduce human intervention.
[0005] This application provides a deep learning-based method for cleaning shock echo nondestructive testing data, including the following steps:
[0006] Step 1: Obtain the shock echo data classification model;
[0007] Step 2: Format the data to be cleaned into a two-dimensional image suitable for input to a deep learning network, then input the image data into the trained deep learning model for classification and obtain data labels;
[0008] Step 3: Remove impact echo data with invalid labels to achieve data cleaning. The cleaned data is then used to detect structural defects.
[0009] Furthermore, the specific operation of establishing the classification model in step 1 is to establish a labeled shock echo dataset based on shock echo signals of known data types, and then use the transfer learning method to input the labeled data into the ResNet network for training. After training, a deep learning model capable of classifying shock echo data can be obtained.
[0010] Furthermore, the shock echo dataset is obtained by determining the features corresponding to normal shock echo data through the time-domain image and spectrum of the shock echo signal, summarizing the two-dimensional features corresponding to different state data, assigning corresponding labels to data that meet a certain feature, and finally integrating and collecting shock echo data with different labels to obtain shock echo datasets with different labels.
[0011] Furthermore, the two-dimensional features include five categories, including...
[0012] a, corresponding to the impact echo thickness mode of the complete region or deep defect;
[0013] b corresponds to a bending mode of a shallow defect;
[0014] c, high-frequency noise;
[0015] d, low-frequency noise;
[0016] e, missing signal;
[0017] Among them, a and b are valid impact echo signals, while c and d are invalid acceleration signals collected by the sensor.
[0018] Furthermore, the two-dimensional image in step 2 is a visualization image of the acceleration sequence data of the shock echo, in a format of 224×224×3.
[0019] Furthermore, the data cleaned in step 3 can be used to obtain a defect assessment map using deep learning methods or frequency analysis methods to achieve accurate detection of hidden defects in concrete structures.
[0020] The beneficial effects of this invention are: the deep learning-based shock echo detection data cleaning method can perform quality judgment and automatic selection on the collected shock echo data, realize the rapid elimination of invalid shock echo data and reduce expert intervention; it is of great significance for ensuring the validity of shock echo data and improving the detection accuracy of hidden defects in concrete structures. Attached Figure Description
[0021] Figure 1 This is a flowchart of the data cleaning method of the present invention.
[0022] Figure 2 A visualization of the shock echo data.
[0023] Figure 3 This is a diagram showing the bridge deck inspection path for an impact echo detection machine.
[0024] Figure 4 A visualization of the raw shock echo data.
[0025] Figure 5 This is a contour plot of the peak frequency after frequency analysis of the original data.
[0026] Figure 6 This is a peak frequency contour plot obtained according to the present invention.
[0027] Figure 7 A contour map of peak frequencies obtained by manual labeling. Detailed Implementation
[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0029] In any embodiment, such as Figure 1 The method for cleaning shock echo nondestructive testing data based on deep learning, as shown, includes the following steps:
[0030] Step 1: Obtain the shock echo data classification model;
[0031] Step 2: Format the data to be cleaned into a two-dimensional image suitable for input to a deep learning network, then input the image data into the trained deep learning model for classification and obtain data labels;
[0032] Step 3: Remove impact echo data with invalid labels to achieve data cleaning. The cleaned data is then used to detect structural defects.
[0033] The specific operation of establishing the classification model in step 1 is to create a labeled shock echo dataset based on shock echo signals of known data types, and then use the transfer learning method to input the labeled data into the ResNet network for training. After training, a deep learning model that can classify shock echo data can be obtained.
[0034] The shock echo dataset is obtained by determining the characteristics corresponding to normal shock echo data through the time-domain image and spectrum of the shock echo signal, summarizing the two-dimensional features corresponding to different state data, assigning corresponding labels to data that meet a certain feature, and finally integrating and collecting shock echo data with different labels to obtain shock echo datasets with different labels.
[0035] like Figure 2 As shown, the two-dimensional features include five categories, including...
[0036] a, corresponding to the impact echo thickness mode of the complete region or deep defect;
[0037] b corresponds to a bending mode of a shallow defect;
[0038] c, high-frequency noise;
[0039] d, low-frequency noise;
[0040] e, missing signal;
[0041] Among them, a and b are valid impact echo signals, while c and d are invalid acceleration signals collected by the sensor.
[0042] like Figure 2 As shown, the two-dimensional image in step 2 is a visualization of the acceleration sequence data of the shock echo, in a format of 224×224×3.
[0043] The data cleaned in step 3 can be used to obtain a defect assessment map using deep learning methods or frequency analysis methods, so as to achieve accurate detection of hidden defects in concrete structures.
[0044] In one embodiment using impact echo data and deep learning analysis results from actual concrete bridge decks, the specific implementation method is as follows:
[0045] First, obtain the classification model for the shock echo data;
[0046] Automated impact echo detection of bridge decks is performed. The automatically operating impact echo detection machine follows... Figure 3The path for impact echo detection was along the bridge deck. The detection machine had two impact echo arrays, performing impact echo detection at regular intervals along each path. Each time, the two parallel arrays were subjected to mechanical impact at eight locations along the transverse direction of the bridge, collecting 16 sets of impact echo signals. A total of 114 tests were performed along the detection path, collecting 1816 sets of impact echo signals. The average interval between each test was 0.6m. The collected raw impact echo data is as follows: Figure 4 As shown. Frequency analysis was performed on the raw shock echo data to extract the distribution of peak frequencies, and the resulting peak frequency contour map is shown below. Figure 5 As shown.
[0047] Secondly, the collected shock echo data is analyzed using the novel method proposed in this invention. The data is visualized as a time-domain signal image and then input into a trained deep learning network for classification and labeling. Based on the labeling results, the original shock echo data is classified as valid or invalid.
[0048] Finally, invalid data was removed, and valid data was retained. Peak frequency contour maps were regenerated using the peak frequency information from the valid data, as shown below. Figure 6 As shown; finally, peak frequency contour maps are used to determine the distribution of hidden defects inside the bridge deck.
[0049] It also provides manually labeled peak frequency contour maps, such as... Figure 7 As shown, the peak frequency contour maps obtained by comparing the present invention, the original data, and the manually labeled method are compared. The manually labeled result is considered to be the true distribution of hidden defects. The results show that the present invention can achieve data cleaning of shock echo detection data, reduce the influence of at least 90% of invalid data, and the data cleaning efficiency of the present invention is much higher than that of the manual labeling method.
[0050] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for cleaning shock echo nondestructive testing data based on deep learning, characterized in that, Includes the following steps: Step 1: Obtain the shock echo data classification model; Step 2: Format the data to be cleaned into a two-dimensional image suitable for input to a deep learning network, then input the image data into the trained deep learning model for classification and obtain data labels; Step 3: Remove impact echo data with invalid labels to achieve data cleaning. The cleaned data is then used to detect structural defects.
2. The method for cleaning shock echo nondestructive testing data based on deep learning according to claim 1, characterized in that, The specific operation of establishing the classification model in step 1 is to create a labeled shock echo dataset based on shock echo signals of known data types, and then use the transfer learning method to input the labeled data into the ResNet network for training. After training, a deep learning model capable of classifying shock echo data can be obtained.
3. The method for cleaning shock echo nondestructive testing data based on deep learning according to claim 2, characterized in that, The shock echo dataset is obtained by determining the characteristics corresponding to normal shock echo data through the time-domain image and spectrum of the shock echo signal, summarizing the two-dimensional features corresponding to different state data, assigning corresponding labels to data that meet a certain feature, and finally integrating and collecting shock echo data with different labels to obtain shock echo datasets with different labels.
4. The method for cleaning shock echo nondestructive testing data based on deep learning according to claim 3, characterized in that, The two-dimensional features include five categories, including a, corresponding to the impact echo thickness mode of the complete region or deep defect; b corresponds to a bending mode of a shallow defect; c, high-frequency noise; d, low-frequency noise; e, missing signal; Among them, a and b are valid impact echo signals, while c and d are invalid acceleration signals collected by the sensor.
5. The method for cleaning shock echo nondestructive testing data based on deep learning according to claim 1, characterized in that, The two-dimensional image in step 2 is a visualization of the acceleration sequence data of the shock echo, in a format of 224×224×3.
6. The method for cleaning shock echo nondestructive testing data based on deep learning according to claim 1, characterized in that, The data cleaned in step 3 can be used to obtain a defect assessment map using deep learning methods or frequency analysis methods, so as to achieve accurate detection of hidden defects in concrete structures.