An invalid data identification method for airborne ground penetrating radar

Abstract

The application discloses an aviation ground penetrating radar invalid data identification method, and comprises the following steps: S1, image data acquisition is carried out through the aviation ground penetrating radar, geological radar profile data is obtained, and a sample data set is formed; S2, the aviation ground penetrating radar data profile image is labeled, and a training set, a verification set and a test set are randomly divided; S3, the training set is trained by using a convolutional neural network model, the model is checked by using the verification set during training, and a weight model is initially obtained; S4, the model is tested by using the test set sample weight model, and a trained aviation ground penetrating radar invalid data identification convolutional neural network model is obtained; and S5, invalid data target detection is carried out on the aviation ground penetrating radar data profile image test set, and a calibrated invalid data ground penetrating radar data profile image is obtained. The method can effectively improve the identification efficiency and accuracy, and lays a foundation for effective ground penetrating radar data interpretation.

Description

TECHNICAL FIELD

[0001] The application relates to a ground penetrating radar data image processing method, in particular to an aerial ground penetrating radar invalid data identification method. BACKGROUND

[0002] The ground penetrating radar is a kind of shallow geophysical exploration means, which is commonly applied to nondestructive testing of engineering quality, hydrogeological survey, mineral resource exploration, underground pipeline survey and geological disaster hazard investigation, etc. In recent years, with the development of exploration demand, the handheld dragging exploration has gradually developed into the exploration in the form of carrying devices, such as vehicle-mounted and airborne. The airborne ground penetrating radar often causes a series of invalid data due to the flight of the unmanned aerial vehicle during data acquisition. The traditional invalid data identification is in the form of manual identification, which requires professional technicians to manually distinguish the invalid data caused by flight. The manual identification is low in efficiency and lacks unified standard, and problems such as omission or misjudgment often occur. With the development of artificial intelligence, various neural network models have been applied to the geological industry, for example, the ground penetrating radar based on the neural network algorithm is applied to the identification of goaf and karst. However, the unmanned aerial vehicle-mounted ground penetrating radar often obtains invalid data generated in the flight process, and at present, the manual annotation is still used. SUMMARY

[0003] The purpose of the present application is to provide an aerial ground penetrating radar invalid data identification method, which solves the problems of low manual calibration efficiency and high error rate of the prior art aerial ground penetrating radar data, realizes automatic annotation of invalid data generated in the flight process of the unmanned aerial vehicle, and effectively improves the identification efficiency and accuracy.

[0004] The purpose of the present application is achieved by the following technical scheme:

[0005] An aerial ground penetrating radar invalid data identification method comprises the following steps:

[0006] Step S1: image data acquisition is performed by the aerial ground penetrating radar to obtain geological radar profile data and form a sample data set, wherein the geological radar profile data contains four types of invalid data: flight turning period data, unmanned aerial vehicle ascending period data, unmanned aerial vehicle descending period data and unmanned aerial vehicle sudden acceleration or deceleration brake period data, and also contains stable flight effective data, and the invalid data and the effective data form a group of acquisition data;

[0007] Step S2: using the sample data set of step S1, the aerial ground penetrating radar data profile image is labeled using the VOC format, and the training set, the verification set and the test set are randomly divided;

[0008] Step S3: the training set of step S2 is trained by using a convolutional neural network model, and the verification set is used to check the model during training to preliminarily obtain a weight model.

[0009] Step S4: using the test set samples to test the weight model trained in step S3, if the test model precision does not meet the requirement, continue to train by increasing the samples, if the test model precision meets the requirement, obtain the trained aviation ground penetrating radar invalid data identification convolutional neural network model;

[0010] Step S5: using the trained aviation ground penetrating radar invalid data identification convolutional neural network model obtained in step S4 to test the invalid data target of the aviation ground penetrating radar data profile image test set, and obtain the ground penetrating radar data profile image of the calibrated invalid data.

[0011] Compared with the prior art, the present application has the following advantages:

[0012] The present application uses the manually annotated aviation ground penetrating radar invalid data image as the training set, trains the invalid data identification model based on the convolutional neural network using the training set, can automatically identify the invalid data of the ground penetrating radar profile image collected during the flight turning period, the unmanned aerial vehicle ascending period, the unmanned aerial vehicle descending period, and the sudden acceleration or deceleration brake period of the unmanned aerial vehicle, and effectively improves the identification efficiency and accuracy, and lays a foundation for effective ground penetrating radar data interpretation. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The flowchart of the aviation ground penetrating radar invalid data identification method;

[0014] Figure 2 The structure diagram of the convolutional neural network model. DETAILED DESCRIPTION

[0015] The technical solutions of the present application will be further described below in conjunction with the drawings, but are not limited thereto, and any modifications or equivalent replacements to the technical solutions of the present application without departing from the spirit and scope of the present application shall be covered in the protection scope of the present application.

[0016] The present application provides an aviation ground penetrating radar invalid data identification method, as shown in Figure 1 The method comprises the following steps:

[0017] Step S1: image data is collected by the aviation ground penetrating radar to obtain the geological radar profile data, and a sample data set is formed, wherein the geological radar profile data contains stable flight effective data and the following four types of invalid data: flight turning period data, unmanned aerial vehicle ascending period data, unmanned aerial vehicle descending period data, and unmanned aerial vehicle sudden acceleration or deceleration brake period data, and the invalid data and the effective data form a group of collected data.

[0018] In this step, the flight route planning of the airborne ground penetrating radar data collection is planned, and the ground penetrating radar profile data is collected using the ground penetrating radar in a specific part of the unmanned aerial vehicle flight trajectory.

[0019] In this step, the data set collects at least 200 groups of data, each group of data should contain valid data and 4 types of invalid data, and the 200 groups of data collected should be selected in different collection areas.

[0020] Step S2: Using the sample data set of step S1, the airborne ground penetrating radar data profile image is labeled using the VOC format, and the training set, validation set and test set are randomly divided.

[0021] In this step, the proportion of the training set, the validation set and the test set is 8:1:1.

[0022] In this step, when the data profile image is symmetrical or approximately symmetrical, this section can be labeled as flight turning period data; when the data profile image is downward at a certain angle (symmetrical with the upward flight angle of the unmanned aerial vehicle), this section can be labeled as unmanned aerial vehicle ascending period data; when the data profile image is suddenly blurred and the resolution is obviously different from the overall situation, this section can be labeled as unmanned aerial vehicle sudden acceleration or deceleration brake period data; when the data profile image is upward at a certain angle (symmetrical with the downward landing angle of the unmanned aerial vehicle), this section can be labeled as unmanned aerial vehicle descending period data.

[0023] Step S3: using a convolutional neural network model to train the training set of step S2, automatically using the model trained in this round to evaluate and calculate the accuracy of the validation set during training, and preliminarily obtaining the weight model.

[0024] In this step, the convolutional neural network model is a YOLOX detection model, which specifically includes an input layer (an image tensor of 640*640*3), a backbone feature extraction network, a strengthened feature extraction part, and a network detection output head.

[0025] The Backbone is a CSPDarknet composed of Focus, CBS, Csplayer and SPP modules. First, the Focus layer is used to slice the input data, which can effectively improve the quality of image feature extraction. Then, multiple CBS and Csplayer layers are used to convolve the image to extract effective features in the image. Finally, the SPP layer is used to fuse the channel information extracted at multiple scales to increase the receptive field of the model.

[0026] The Neck is composed of FPN+PAN. The FPN structure first downsamples the image, and then the PAN is used to splice the downsampled branches through upsampling.

[0027] The Head is different from the Head of the previous YOLO version, and classification and regression are split and implemented in 1*1 convolution respectively, and finally integrated together when predicting;

[0028] The training set image is transmitted from the input layer to the backbone feature extraction network, first extracts the features of invalid data in the aerial ground penetrating radar data profile image through the Backbone; then the features extracted by the Backbone are multi-scale fused through the Neck to improve the generalization ability of the model; finally the extracted features are decoded through the Head to output the detection result.

[0029] Step S4: use the test set sample to test the weight model trained in step S3, if the test model precision does not meet the requirements, increase the sample to continue training, if the test model precision meets the requirements, get the trained aerial ground penetrating radar invalid data recognition convolutional neural network model.

[0030] In this step, the performance and classification ability of the preliminary model are measured by model testing, the model parameters are adjusted continuously, the model accuracy is recorded, the best model parameters are selected, and the trained aerial ground penetrating radar invalid data recognition convolutional neural network model is obtained.

[0031] Step S5: use the trained aerial ground penetrating radar invalid data recognition convolutional neural network model obtained in step S4 to detect invalid data targets in the aerial ground penetrating radar data profile image test set, and get the calibrated invalid data ground penetrating radar data profile. The specific steps are as follows:

[0032] Call the trained aerial ground penetrating radar invalid data recognition convolutional neural network model in step S4, input the aerial ground penetrating radar data profile image which does not participate in training and testing to detect invalid data targets, get invalid data automatic frame taking and invalid data type probability value size, and automatically select the maximum probability result as the recognition result through the non-maximum suppression algorithm.

Claims

1. An invalid data identification method for airborne ground penetrating radar, characterized in that The method comprises the following steps: Step S1: image data acquisition is performed by an airborne ground penetrating radar to obtain geological radar profile data to form a sample data set, wherein the geological radar profile data contains stable flight effective data and four types of invalid data: flight turning period data, unmanned aerial vehicle ascending period data, unmanned aerial vehicle descending period data, and unmanned aerial vehicle sudden acceleration or deceleration brake period data, and the invalid data and the effective data form a group of acquisition data; Step S2: using the sample data set of step S1, the airborne ground penetrating radar data profile image is labeled using a VOC format, and a training set, a verification set and a test set are randomly divided, wherein when the data profile image is symmetrical or approximately symmetrical, this section is labeled as flight turning period data; when the data profile image is symmetrical downward at a certain angle and the unmanned aerial vehicle flies upward at a certain angle, this section is labeled as unmanned aerial vehicle ascending period data; when the data profile image is suddenly blurred and the resolution is obviously different from the overall situation, this section is labeled as unmanned aerial vehicle sudden acceleration or deceleration brake period data; when the data profile image is symmetrical upward at a certain angle and the unmanned aerial vehicle descends downward at a certain angle, this section is labeled as unmanned aerial vehicle descending period data; Step S3: the training set of step S2 is trained using a convolutional neural network model, the verification set is used to view the model during training, and a weight model is preliminarily obtained; Step S4: the weight model obtained by training in step S3 is tested using a test set sample, if the test model accuracy does not meet the requirements, the sample is increased for continuous training, if the test model accuracy meets the requirements, a trained airborne ground penetrating radar invalid data identification convolutional neural network model is obtained; Step S5: the trained airborne ground penetrating radar invalid data identification convolutional neural network model obtained in step S4 is used to detect invalid data targets in the test set of the airborne ground penetrating radar data profile image, and a ground penetrating radar data profile image with labeled invalid data is obtained.

2. The method of claim 1, wherein In step S1, at least 200 groups of data are collected in the data set, each group of data should contain effective data and four types of invalid data, and the 200 groups of collected data select different collection areas.

3. The method of claim 1, wherein In step S2, the ratio of the training set, the verification set and the test set is 8:1:

1.

4. The method of claim 1, wherein In step S3, the convolutional neural network model is a YOLOX detection model, which includes an input layer, a backbone feature extraction network, a strengthened feature extraction part, and a network detection output head. The training set image is transmitted from the input layer to the backbone feature extraction network. First, the features of the invalid data in the airborne ground penetrating radar data profile image are extracted by the backbone. Then, the features extracted by the backbone are fused by the neck to improve the generalization ability of the model. Finally, the extracted features are decoded by the head to output the detection results.

5. The method of claim 4, wherein The Backbone is CSPDarknet, which is composed of Focus, CBS, Csplayer and SPP modules, first, the input data is sliced by the Focus layer, then the image is convolved by multiple CBS and Csplayer layers to extract effective features in the image, finally, the channel information extracted by multiple scales is fused by the SPP layer to increase the receptive field of the model; the Neck is composed of FPN+PAN, FPN first down-samples the picture, and then the down-sampled branches are spliced by up-sampling through PAN; the classification and regression of the Head are split, and are realized in 1 1 convolution respectively, and are integrated together when finally predicting.

6. The method of claim 1, wherein In step S4, the performance and classification ability of the preliminary model are measured by model testing, the model accuracy is recorded by continuously adjusting the model parameters, the best model parameters are selected, and the trained airborne ground penetrating radar invalid data identification convolutional neural network model is obtained.

7. The method of claim 1, wherein In the step S5, the trained aviation ground penetrating radar invalid data recognition convolutional neural network model in the step S4 is called, the aviation ground penetrating radar data profile images not participating in the training and the test are input, invalid data target detection is performed, invalid data automatic framing and invalid data type probability value size are obtained, and the maximum probability result is automatically selected as the recognition result through a non-maximum suppression algorithm.

Images