Machine learning data labeling method and system for ground penetrating radar signals

Abstract

This invention relates to a machine learning data annotation method and system for ground-penetrating radar (GPR) signals. The method converts raw GPR data into an XYZ format file. The program then converts continuous single-channel scan data within the XYZ file into an S×T×E matrix, where S represents the number of single-point scans in each annotation interval, T represents the time window depth of a single-channel scan, and E represents the electric field intensity of the radar echo within the S×T range. The program then uses the maximum normalization method to normalize these electric field intensities E to the range of 0-255, resulting in a grayscale image of the relative electric field intensity of size S×T, which is displayed in real-time in a B-scan window to aid annotation. The "three-line positioning method" is then used to annotate the specific time window interval of a single-channel scan. A dataset construction program saves the annotation results as the original data and their corresponding labels. This invention integrates raw data interpretation, visual-assisted annotation, annotation result storage, and dataset creation, deeply combining GPR with machine learning.

Description

Technical Field

[0001] This invention relates to a machine learning data labeling method and system for ground-penetrating radar signals, belonging to the field of machine learning data labeling methods. Background Technology

[0002] With the continuous updates and development of neural network classifiers, deep learning has gradually been applied to the automated recognition of radar maps. The technology of using deep learning to recognize ground-penetrating radar B-scan profiles is relatively mature. Table 1 lists the relevant research on the combination of deep learning and B-scan profile recognition.

[0003] Table 1. Road surface distress detection method based on deep neural network and ground penetrating radar

[0004]

[0005] Image recognition based on B-scan profiles is only one approach to interpreting ground-penetrating radar signals, and it has significant drawbacks, with information loss being the most prominent. During image recognition, the original electric field intensity echo data is limited to the range of 0-255, but the amplitude of radar echo data can actually exceed 30,000. Clearly, a larger data range can carry more information, but image recognition technology cannot effectively extract this information, becoming a major bottleneck restricting the development of image recognition technology.

[0006] Temporal neural networks (RNNs) have significant advantages in interpreting a set of logically related column vectors. Radar echo data is precisely a set of time-related column vectors, making RNN-based identification a natural approach. However, this technology faces challenges from its early stages, including a lack of corresponding annotation methods and datasets. Regarding datasets, most radar data is presented directly as echo maps and must be acquired using specialized software. The encapsulation functions used in this process can lead to data loss. Furthermore, even if the raw data can be interpreted, current research lacks a comprehensive set of annotation and dataset creation software and methods.

[0007] Therefore, to address the above issues, it is necessary to construct a labeling method for ground-penetrating radar signals to form a foundational technology for further deep integration with machine learning. Summary of the Invention

[0008] This invention provides a machine learning data annotation method and system for ground penetrating radar signals, which integrates raw data interpretation, visual-assisted annotation, annotation result storage, and dataset creation, deeply combining ground penetrating radar with machine learning.

[0009] The technical solution adopted by this invention to solve its technical problem is:

[0010] A machine learning data annotation method for ground-penetrating radar signals includes the following steps:

[0011] Step S1: Collect raw data from ground penetrating radar, preprocess the raw data to obtain an XYZ format file;

[0012] Step S2: The XYZ format file contains multiple consecutive single-channel scan data. Each single-channel scan data is in a cyclic format. The cyclic format single-channel scan data is divided into a list of single-channel scan electric field intensities. The single-channel scan data is saved in a temporary file as a unit.

[0013] Step S3: Define the number of single-channel scans within each labeled interval, and obtain the S×T×E matrix from the temporary file, where S represents the number of single-point scans in each labeled interval, T represents the time window depth of a single-channel scan, and E represents the electric field intensity of the radar echo at the corresponding point within the S×T range.

[0014] Step S4: Use the maximum normalization method to normalize the S×T×E matrix to the range of 0-255 on the electric field intensity scale, obtain the grayscale image of the relative electric field intensity of size S×T, and display the grayscale image of the relative electric field intensity in the program window;

[0015] Step S5: Use the three-line positioning method to mark the specific time window interval of a single-channel scan within each marked interval. The red line marks the location of the single-channel scan, and the blue and green lines mark the two boundaries of the time window interval of the single-channel scan. That is, the three lines represent a segment of the red line on the grayscale map of the relative value of the electric field intensity, and in the marking results, they represent the radar echo signal anomaly within a certain time window interval of the single-channel scan.

[0016] Step S6: Following step S5, compile the annotation results of a single S×T×E matrix into a dictionary, with the format {(S i ,T iu ,T il ):'N i TYPE i '}, where S i T represents the single-channel scan number where the abnormal signal is located. iu T represents the upper limit of the time window interval for the abnormal signal in a single-channel scan. il This indicates the lower limit of the time window interval for the abnormal signal in a single-channel scan, TYPE. i Indicates the type of disease, N i This indicates the index assigned within the dictionary; continue saving the individual S×T×E matrices and the resulting dictionary in the program's text file;

[0017] Step S7: Following step S6, write a dataset generation function, set the input path of the text file of the annotation results and the output path of the dataset. The function automatically recognizes the shape of the annotation result matrix and generates an empty matrix with the same shape as the annotation results. Read the dictionary formed by the labeling results, assign values ​​to the signals corresponding to the n disease types in the range of 0-255 and write them to the corresponding positions in the empty matrix, record the background signal as 0, match it with the labeling results and save it in the specified path to complete the generation of the dataset.

[0018] As a further preferred embodiment of the present invention, in step S1, the preprocessing of the raw data includes static correction and removal, signal gain and horizontal interference removal of the raw data to obtain an interpretable XYZ format file.

[0019] As a further preferred embodiment of the present invention, in step S2, each single-channel scan data saves all echo information in a cyclic format of sequence number-time window-field strength;

[0020] The data annotation system built based on the machine learning data annotation method for ground penetrating radar signals integrates the following steps: obtaining XYZ format files in step S1, interpreting data according to the training requirements of time-series neural networks in steps S2-S3, annotating ground penetrating radar signals in steps S4-S5, saving annotation results in step S6, and generating datasets in step S7. This process forms a machine learning data annotation system for ground penetrating radar signals.

[0021] By employing the above technical solutions, the present invention has the following beneficial effects compared to the prior art:

[0022] 1. The machine learning data annotation method for ground penetrating radar signals provided by this invention has corresponding annotation methods and datasets, and can form a complete set of annotation and dataset creation methods and systems;

[0023] 2. The machine learning data annotation method for ground-penetrating radar signals provided by this invention directly interprets the raw data of ground-penetrating radar, avoiding data loss during function encapsulation. It also uncovers effective information over a wider data range, improving the scientific rigor of the annotation method and contributing to the accuracy of machine learning training results. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is a flowchart of the machine learning data annotation method for ground-penetrating radar signals provided by the present invention;

[0026] Figures 2-3 This is a schematic diagram of the interface after opening and splitting an XYZ format file in an embodiment provided by the present invention;

[0027] Figures 4-5 This is a schematic diagram of the mode interface before entering the annotation mode using the three-line positioning method in the embodiment provided by the present invention;

[0028] Figure 6 This is a schematic diagram of the interface that needs to be labeled in the embodiments provided by the present invention;

[0029] Figure 7 This is a schematic diagram of the interface in the embodiment provided by the present invention, in which the red line marks the selected disease type;

[0030] Figures 8-12 This is a schematic diagram of the interface for saving the tag discrimination process in the embodiments provided by the present invention;

[0031] Figure 13 This is a schematic diagram of the interface after inputting the dataset and annotations into the convolutional neural network in an embodiment provided by the present invention;

[0032] Figure 14 This is a schematic diagram of the final identification result of the embodiment provided by the present invention. Detailed Implementation

[0033] The invention will now be described in further detail with reference to the accompanying drawings.

[0034] In the background section, the applicant elaborates on the shortcomings of current ground-penetrating radar signal interpretation based on B-scan profile images, which suffers from information loss. Therefore, this application attempts to combine temporal neural networks (RNNs) into the aforementioned image recognition technology. However, there is no established system for RNN recognition that can both avoid data loss and label the data.

[0035] Based on the aforementioned difficulties, this application fills this gap. Figure 1 As shown, a machine learning data annotation method for ground-penetrating radar signals is constructed, which specifically includes the following steps:

[0036] Step S1: Collect raw data from ground penetrating radar. Since the raw data obtained from ground penetrating radar is encrypted, it is necessary to preprocess the raw data. After steps such as static correction and removal, signal gain and horizontal interference removal, an interpretable XYZ format file is finally obtained.

[0037] Step S2: Interpret the data according to the training requirements of the time-series neural network. The XYZ format file contains multiple continuous single-channel scan data. Each single-channel scan data saves all echo information in a cyclic format of "serial number-time window-field strength". The program divides the cyclic format single-channel scan data in memory to form a list of single-channel scan electric field strengths, and saves them in temporary files in units of single-channel scans.

[0038] Step S3: In the program, define the number of single-channel scans within each labeled interval, and obtain the S×T×E matrix from the temporary file into memory, where S represents the number of single-point scans in each labeled interval, T represents the time window depth of a single-channel scan, and E represents the electric field intensity of the radar echo at the corresponding point within the S×T range.

[0039] Step S4: The program uses the maximum normalization method to normalize the S×T×E matrix to the range of 0-255 on the electric field intensity scale, and obtains a grayscale image of the relative electric field intensity of size S×T. The grayscale image of the relative electric field intensity is then displayed in the program's B-scan window as a visual aid for annotation.

[0040] Step S5: Use the "three-line positioning method" to mark the specific time window interval of a single-channel scan within each marked interval. The red line marks the location of the single-channel scan, and the blue and green lines mark the two boundaries of the time window interval of the single-channel scan. Thus, the three lines represent a segment of the red line on the grayscale map of the relative value of the electric field intensity, and in the marking results, they represent the abnormal radar echo signal within a certain time window interval of the single-channel scan.

[0041] Step S6: Following step S5, the program generates a dictionary from the annotation results of a single S×T×E matrix, with the format {(S i ,T iu ,T il ):'N i TYPE i '}, where S i T represents the single-channel scan number where the abnormal signal is located. iu T represents the upper limit of the time window interval for the abnormal signal in a single-channel scan. il This indicates the lower limit of the time window interval for the abnormal signal in a single-channel scan, TYPE. i Indicates the type of disease, N i This indicates the index assigned internally to the dictionary (this index is static and the same in all labeled items); the program continues to save the individual S×T×E matrix and the resulting dictionary in the program's text file.

[0042] Step S7: Following step S6, write the dataset generation function label2dataset.py, set the input path of the text file of the annotation results and the output path of the dataset. The function automatically recognizes the shape of the annotation result matrix and generates an empty matrix with the same shape as the annotation results. Read the dictionary formed by the labeling results, assign values ​​to the signals corresponding to the n disease types in the range of 0-255 and write them to the corresponding positions in the empty matrix, record the background signal as 0, match it with the labeling results and save it in the specified path to complete the generation of the dataset.

[0043] One of the highlights of this application is to directly use the raw data of the ground penetrating radar to improve the accuracy of the original benchmark of the annotation method. At the same time, as can be seen from the description of the above method, this application integrates the program of obtaining the XYZ format file in step S1, interpreting the data according to the requirements of the sequential neural network training in steps S2 - S3, annotating the ground penetrating radar signal in steps S4 - S5, saving the annotation results in step S6, and generating the data set in step S7, and constructs a machine learning data annotation system for the ground penetrating radar signal, that is, integrates the complete set of technologies of "raw data interpretation - visual assisted annotation - annotation result storage - data set production", making the ground penetrating radar and machine learning further deeply combined.

[0044] Next, this application provides an embodiment to elaborate in detail the implementation process of the machine learning data annotation method for the ground penetrating radar signal provided by this application.

[0045] First, obtain the XYZ format file with the help of the REFLEXW software or the RGPR open source program, and obtain the interpretable.XYZ format file "Tiankun Second Road_001_A01.XYZ".

[0046] Open the.XYZ in the program: Run the program, click "File" - "Open", select the.XYZ file to be annotated in the pop-up dialog box, and "The loaded file is: Tiankun Second Road_001_A01.XYZ" will be displayed in the prompt box of the software. Click the "Split" button, and the program will sequentially extract each trace of A-Scan data of this file and write it into a temporary file. The prompt box will display "The A_scan splitting of the file Tiankun Second Road_001_A01 has been completed", as shown in Figure 2 to Figure 3 shown.

[0047] Split the data according to the requirements of the sequential neural network training: Input the total length l represented by the opened file, that is, the distance at which this XYZ format file is collected, in meters; select the sentence length / number of A-Scans n in the drop-down selection box, that is, how many A-Scan data are taken at one time for marking; then click the "Preview" button, and the radar echo map (also called a sentence) composed of the first n A-Scans will be displayed in the B-Scan display box on the left. After comprehensively considering the computing power of the computer and the expected effect, select a suitable n, and click the "Adopt" button to enter the annotation mode, as shown in Figure 4 to Figure 5 shown.

[0048] Radar echo data signal scale annotation: In annotation mode, the file name and mileage segment are displayed at the top of the B-Scan window. Left-clicking the location to be annotated in the B-Scan image will display the corresponding A-Scan image in the A-Scan window. The image sequence and mileage position are displayed at the top of the A-Scan window, as shown in the attached figure. Figure 6 As shown in the attached image; then drag the blue handle (upper control handle) and green handle (lower control handle) next to the A-Scan window to select the A-Scan data range to be labeled. In the drop-down selection box at the bottom of the A-Scan window, select the disease type, and click the "Write" button. LabelGPR will mark the data between the two control handles in the A-Scan data at the location of the red line in the B-Scan window as the selected disease type, as shown in the attached image. Figure 7 As shown.

[0049] Saving Tagging Results: If "None" is not displayed in the "Table of Tags Written in This Sentence" section, it indicates that all content in the table has been written as tags into the data and saved. Each sentence can be tagged with several tags, as shown in the appendix. Figure 8 As shown; after determining that there is no content in the sentence that needs to be marked, click the "Next Sentence" button to repeat the radar echo data signal scale annotation steps. Each time the "Write" or "Next Sentence" button is clicked, the program will automatically save the marking results of the current sentence. If there are no markings, only the original data will be saved, and "None" will be displayed in the label table, as shown in the attached document. Figure 9 As shown; continue marking until the entire .XYZ file is traversed. If the length of the last sentence is less than the initially selected n value, the insufficient A-Scan data is padded with 0 and saved, as shown in the attached figure. Figure 10 To be continued Figure 12 As shown.

[0050] Generate dataset: Call the function label2dataset.py, set the input and output paths, generate an empty matrix with the same shape as the labeling results, read the labeling result dictionary, uniformly assign values ​​to the signals corresponding to the n disease types in the range of 0-255 and write them to the corresponding positions in the empty matrix, and record the background signal as 0; finally, output this matrix and save it in the specified path, corresponding one-to-one with the labeling results.

[0051] A neural network was developed using the dataset: A total of 1200 labeled radar signals were used, each with a classification vector of 478×1. Among these, 600 signals contained depth information, with depth vectors of 2×1. The training set consisted of 840 signals, the validation set of 240 signals, and the test set of 120 signals. The ratio of the training, validation, and test sets was set to 7:2:1, which meets the requirements for the training and testing dataset ratio of convolutional neural networks. The dataset and its annotations are shown in the attached figure. Figure 13In the convolutional neural network shown, the overall accuracy of the final signal-level cascaded convolutional neural network is 97.75%. The accuracy rates for identifying undamaged signals, rebar signals, signals with voids without rebar interference, and signals with voids due to rebar interference are 99%, 94%, 99%, and 99%, respectively. The model can accurately identify abnormal signals caused by voids at the bottom of the slab, as shown in the attached figure. Figure 14 As shown.

[0052] Clearly, the machine learning data labeling method and the resulting system for ground penetrating radar signals provided in this application can accurately identify abnormal signals, becoming a foundational technology for the further deep integration of ground penetrating radar and machine learning.

[0053] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0054] The meaning of "and / or" as used in this application includes situations where each exists alone or both exist simultaneously.

[0055] The term "connection" as used in this application can mean a direct connection between components or an indirect connection between components through other components.

[0056] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A machine learning data annotation method for ground-penetrating radar signals, characterized by: Specifically, the following steps are included: Step S1: Collect raw data from ground penetrating radar, preprocess the raw data to obtain an XYZ format file; Preprocessing of the raw data includes static correction and removal, signal gain adjustment, and horizontal interference removal to obtain an interpretable XYZ format file; Step S2: The XYZ format file contains multiple consecutive single-channel scan data. Each single-channel scan data is in a cyclic format. The cyclic format single-channel scan data is divided into a list of single-channel scan electric field intensities. The single-channel scan data is saved in a temporary file as a unit. Each single-channel scan data is stored in a cyclic format of sequence number-time window-field strength, saving all echo information; Step S3: Define the number of single-channel scans within each labeled interval, and obtain the S×T×E matrix from the temporary file, where S represents the number of single-point scans in each labeled interval, T represents the time window depth of a single-channel scan, and E represents the electric field intensity of the radar echo at the corresponding point within the S×T range. Step S4: Use the maximum normalization method to normalize the S×T×E matrix to the range of 0-255 on the electric field intensity scale, obtain the grayscale image of the relative electric field intensity of size S×T, and display the grayscale image of the relative electric field intensity in the program window; Step S5: Use the three-line positioning method to mark the specific time window interval of a single-channel scan within each marked interval. The red line marks the location of the single-channel scan, and the blue and green lines mark the two boundaries of the time window interval of the single-channel scan. That is, the three lines represent a segment of the red line on the grayscale map of the relative value of the electric field intensity, and in the marking results, they represent the radar echo signal anomaly within a certain time window interval of the single-channel scan. Step S6: Following step S5, compile the annotation results of a single S×T×E matrix into a dictionary, with the format {(S i T iu T il ): 'N i TYPE i '}, where S i T represents the single-channel scan number where the abnormal signal is located. iu T represents the upper limit of the time window interval for the abnormal signal in a single-channel scan. il This indicates the lower limit of the time window interval for the abnormal signal in a single-channel scan, TYPE. i Indicates the type of disease, N i This indicates the index assigned within the dictionary; continue saving the individual S×T×E matrices and the resulting dictionary in the program's text file; Step S7: Following step S6, write a dataset generation function, set the input path of the text file of the annotation results and the output path of the dataset. The function automatically recognizes the shape of the annotation result matrix and generates an empty matrix with the same shape as the annotation results. Read the dictionary formed by the labeling results, assign values ​​to the signals corresponding to the n disease types in the range of 0-255 and write them to the corresponding positions in the empty matrix, and record the background signal as 0. Output the matrix with the background signal as 0, match it with the labeling results and save it in the specified path to complete the generation of the dataset.

2. A data annotation system based on the machine learning data annotation method for ground-penetrating radar signals as described in claim 1, characterized in that: The program integrates steps S1 (obtaining XYZ format files), S2-S3 (interpreting data according to the training requirements of a time-series neural network), S4-S5 (annotating ground-penetrating radar signals), S6 (saving the annotation results), and S7 (generating the dataset) to construct a machine learning data annotation system for ground-penetrating radar signals.

Images