Deep learning analysis method for impact echo detection data

Abstract

The invention relates to the field of nondestructive testing of concrete structures, in particular to a deep learning analysis method for impact echo detection data. Comprising the following steps: training a deep learning model by using an existing impact echo mark data set, and inputting mark data into a deep learning network for training by using a transfer learning method to obtain the deep learning model; formatting to-be-analyzed data into a two-dimensional image suitable for deep learning network input, then inputting image data into a deep learning model for classification, and obtaining a probability value of each group of impact echo data corresponding to a defect area; and generating a defect probability contour map based on probability to realize hidden defect evaluation. The impact echo detection data analysis method based on deep learning has important significance in reducing manual intervention, lowering the professional threshold of impact echo data analysis and realizing rapid and accurate hidden defect detection of the concrete structure.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing of concrete structures, and in particular to a deep learning analysis method for impact echo detection data. Background Technology

[0002] Impact-echo nondestructive testing (PET) 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, and then the impact-echo data is received using a contact accelerometer or an air-coupled microphone. Traditional PTT uses the peak frequency method to convert the collected stress wave signal into the frequency domain via Fourier transform for analysis. The peak frequency corresponding to the impact-echo mode is used to analyze the structural thickness or determine the location of hidden defects. However, the data collected by PTT is an acceleration signal caused by structural or air vibration. Researchers have recognized that the collected impact-echo signals are difficult to analyze directly using the peak frequency method. While peak frequency acquisition can be automated with simple methods, manual acquisition is still necessary for some signals with poor data quality. This requires human intervention, and those involved in peak frequency analysis must be professionals with a strong theoretical foundation in stress wave propagation. As PTT evolves from single-point detection using handheld instruments to multi-sensor scanning detection, it faces even greater data processing challenges. With the increasing amount of non-destructive testing data generated by automated equipment, automated data analysis based on artificial intelligence is becoming increasingly important in the field of non-destructive testing because it can reduce human intervention and lower the time and labor costs of data analysis.

[0003] Deep learning methods have demonstrated excellent performance in recent years when dealing with massive amounts of data analysis. Deep learning networks use layer-by-layer transformations of the signal feature space to convert the feature space into a new one. Defect feature representation is automatically implemented. Deep learning networks extract target features from the input data, significantly improving the performance of the recognition and retrieval process. A key reason for this is that the feature extraction method is based on learning from large datasets. This invention aims to develop a data analysis method that uses deep learning to automatically classify collected shock echo data, mapping shock echo signal features to the distribution of hidden defects in concrete structures, and quickly obtaining the distribution of hidden defects in concrete structures. Summary of the Invention

[0004] The purpose of this application is to provide a deep learning analysis method for shock echo detection data, which aims to solve the problems in the prior art.

[0005] This application provides a deep learning analysis method for shock echo detection data, including the following steps:

[0006] S1. Train a deep learning model using an existing shock echo labeled dataset. Use transfer learning to input labeled data into the deep learning network for training. After training, a deep learning model that can classify shock echo data is obtained.

[0007] S2. Format the data to be analyzed 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 the probability value of each group of shock echo data corresponding to the defect area.

[0008] S3. Generate a probability-based contour map of defects to achieve hidden defect assessment.

[0009] In S1, a labeled shock echo dataset is established using shock echo signals of known data types. The shock echo signals are marked as defective or complete according to the location of the detection point, and then a time-frequency domain image is generated using wavelet transform.

[0010] In the time-frequency domain image, the horizontal axis represents time, the vertical axis represents frequency, and the image color represents the energy level. In the process of generating time-frequency features, the output frequency value is limited to the range of 1-20kHz.

[0011] The transfer learning method is used to input labeled data into the ResNet network for training. After training, a deep learning model that can classify shock echo data is obtained.

[0012] In S2, the data to be analyzed is formatted as a time-frequency domain image, and then the image data is input into the trained deep learning model for classification, and the data labels and corresponding probabilities are obtained.

[0013] In S3, the defect probability values ​​obtained using deep learning methods are used to generate defect probability contour maps to achieve accurate detection of hidden defects in concrete structures.

[0014] The beneficial effects of this invention are: the deep learning-based impact echo detection data analysis method of this invention is of great significance for reducing manual intervention, lowering the professional threshold for impact echo data analysis, and achieving rapid and accurate detection of hidden defects in concrete structures. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0016] Figure 2 A time-frequency domain image is generated for the impulse echo signal using wavelet transform.

[0017] Figure 3 This shows the distribution of hidden defects on the defective board.

[0018] Figure 4 This is a contour plot showing the probability of defects. Detailed Implementation

[0019] 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.

[0020] Example 1

[0021] like Figure 1 The following is a deep learning analysis method for shock echo detection data, with the following steps:

[0022] S1. Obtain the shock echo data classification model.

[0023] A labeled shock echo dataset is established based on shock echo signals of known data types. The labeled data is then input into a ResNet network for training using the transfer learning method. After training, a deep learning model capable of classifying shock echo data is obtained.

[0024] The impact echo signal is marked as defective or intact according to the location of the detection point, and then a time-frequency domain image is generated using wavelet transform, such as... Figure 2 As shown, in the time-frequency domain image, the horizontal axis represents time, the vertical axis represents frequency, and the image color represents the energy level. In the process of generating time-frequency features, the output frequency value is limited to 1-20kHz, and unnecessary low-frequency and high-frequency signals are removed.

[0025] S2. Obtain formatted data to be analyzed.

[0026] The shock echo data to be analyzed is formatted into a two-dimensional image suitable for input to a deep learning network. The image data is then input into the trained deep learning model for classification, and data labels and corresponding probabilities are obtained.

[0027] Visualizing the acceleration sequence data of the shock echo as follows: Figure 2 The image shown is in 224×224×3 format.

[0028] S3. Use the defect probability values ​​obtained by deep learning methods to generate defect probability contour maps to achieve accurate detection of hidden defects in concrete structures.

[0029] Example 2

[0030] The effectiveness of this method was verified by using impact echo data from eight actual concrete artificial defect slabs, deep learning analysis results, and artificial labeling results.

[0031] The specific implementation method is as follows:

[0032] (1) Impact echo detection was performed on the defective plates. Each defective plate contained 8 hidden defects. The distribution of the hidden defects in the defective plates was as follows: Figure 3 As shown, based on the relationship between the defect distribution and the collection location of the shock echo detection data, the shock echo data were labeled as complete and defective. Then, the shock echo data of 7 plates were used to train a deep learning network. After training, a deep learning model capable of classifying shock echo detection data was obtained.

[0033] (2) The impact echo data of the concrete slab that was not trained by the deep learning model were visualized and then input into the trained deep learning model. The deep learning model classified the input data and obtained the corresponding defect probability. A defect probability contour map was then used. Figure 4 Lateral distance (lateral distance), Longitudinal distance (longitudinal distance), and probability of defect presence (probability of defect presence) can be used to obtain the distribution of hidden defects in the plate and realize the analysis of impact echo detection data.

[0034] (3) Comparing the present invention with the actual defect distribution; the results show that the present invention can realize the data analysis of impact echo detection data, automatically identify the distribution of hidden defects in concrete slabs, and has a more prominent identification effect on shallow layered defects. The present invention can reduce the theoretical threshold of impact echo detection data analysis.

[0035] 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 deep learning analysis method for impact-echo detection data, comprising the following steps: S1. Training a deep learning model using an existing impact-echo labeled data set, inputting the labeled data into a deep learning network for training using a transfer learning method, and obtaining a deep learning model capable of classifying impact-echo data after training is completed; S2. Formatting the data to be analyzed into a two-dimensional image suitable for input into the deep learning network, then inputting the image data into the trained deep learning model for classification, and obtaining a probability value corresponding to the defect area for each set of impact-echo data; S3. Generating a defect probability contour map based on the probability, and achieving hidden defect evaluation. 2.The method of claim 1, wherein, In S1, a labeled impact-echo data set is established using known data types of impact-echo signals, the impact-echo signals are labeled as defects or complete according to the position of the detection point, and then a time-frequency domain image is generated using wavelet transform. 3.The deep learning analysis method of impact-echo data according to claim 2, characterized in that, In the time-frequency domain image, the horizontal axis represents time, the vertical axis represents frequency, and the image color represents energy size. In the process of generating time-frequency features, the output value of frequency is limited to the range of 1-20 kHz. 4.The method of claim 3, wherein, The labeled data is input into the ResNet network for training using a transfer learning method, and a deep learning model capable of classifying impact-echo data is obtained after training is completed. 5.The method of claim 1, wherein, In S2, the data to be analyzed is formatted into a time-frequency domain image, then the image data is input into the trained deep learning model for classification, and data labels and corresponding probabilities are obtained. 6.The method of claim 1, wherein, In S3, a defect probability contour map is generated using the defect probability values obtained by the deep learning method to achieve correct hidden defect detection of concrete structures.

Images