Planetary subsurface layer trafficability discrimination method based on planetary vehicle ground penetrating radar data

Abstract

The invention provides a planetary subsurface layer trafficability discrimination method based on planetary vehicle ground penetrating radar data, and relates to the field of ground penetrating radar data interpretation, and the method comprises the steps: obtaining the planetary vehicle ground penetrating radar data; according to the planet vehicle ground penetrating radar data, a planet vehicle number surface trafficability data set is constructed; constructing a non-local attention classification network based on the B-scan data of the ground penetrating radar; iteratively training the non-local attention network through the planetary train number surface trafficability data set to obtain a trained deep learning model; acquiring B-scan data of the ground penetrating radar to be classified; and classifying the to-be-classified ground penetrating radar B-scan data through the trained deep learning model to obtain a star catalogue trafficability identification result, and completing planetary subsurface trafficability discrimination of a region corresponding to the ground penetrating radar data. According to the technical scheme, subsurface planetary vehicle trafficability prediction of single-environment star soil is realized.

Description

Technical Field

[0001] This application relates to the field of interpretation of ground penetrating radar data, and particularly to a method for discriminating the passability of the subsurface of a planet based on ground penetrating radar data of a planetary rover. Background Art

[0002] For the passability analysis of a planetary rover, it mainly focuses on identifying obstacle targets such as meteorite craters and stones on the planet surface to analyze the passability of the planetary rover. However, the actual situation is that there may be obstacles such as cavities and fractures in the subsurface of the moon, which affect the passability of the planetary rover. Therefore, it is imperative to analyze the passability of the subsurface of the planet. For the detection of the subsurface, a ground penetrating radar carried by a planetary rover is usually used. Through the ground penetrating radar technology, the subsurface structure of the planet is detected in a non-contact manner to discriminate the passability of the planetary rover.

[0003] Currently, for the passability prediction of ground penetrating radar data of the subsurface of a planet using an intelligent recognition method based on a deep learning network model, no relevant research results have been reported. Summary of the Invention

[0004] The purpose of the present invention is to fill the technical gap in the subsurface passability prediction of a planetary rover using ground penetrating radar data, and to provide a method for discriminating the passability of the subsurface of a planet based on ground penetrating radar data of a planetary rover.

[0005] The above object of this application is achieved through the following technical solutions: S1: Obtain the ground penetrating radar data of the planetary rover; construct a subsurface passability dataset of the planetary rover according to the ground penetrating radar data of the planetary rover; S2: Construct a non-local attention classification network based on the B-scan data of the ground penetrating radar; S3: Iteratively train the non-local attention network through the subsurface passability dataset of the planetary rover to obtain a trained deep learning model; S4: Obtain the B-scan data of the ground penetrating radar to be classified; classify the B-scan data of the ground penetrating radar to be classified through the trained deep learning model to obtain the recognition result of the surface passability, and complete the discrimination of the passability of the subsurface of the planet corresponding to the area of the ground penetrating radar data.

[0006] Optionally, step S2 includes: The non-local attention classification network includes: an input convolution unit, a non-local attention module, a global average pooling layer, and a fully connected classification output layer; The input convolution unit is connected to the non-local attention module; The non-local attention module is connected to the global average pooling layer; The global average pooling layer is connected to the fully connected classification output layer; The input convolutional unit includes: a 3×3 convolutional layer, a normalization module layer, and an h-swish non-linear activation function layer; the 3×3 convolutional layer, the normalization module layer, and the h-swish non-linear activation function layer are connected in sequence. The global average pooling layer includes: a fully connected layer, a normalization layer, and an h-swish non-linear activation function layer; the fully connected layer, the normalization layer, and the h-swish non-linear activation function layer in the global average pooling layer are connected in sequence.

[0007] Optionally, step S2 further includes: The non-local attention module includes: a channel attention branch and a context attention branch; The channel attention branch and the context attention branch are connected in parallel; The channel attention branch includes: a global pooling layer, a first 1×1 convolutional layer, an h-swish non-linear activation function layer, a second 1×1 convolutional layer, and a Sigmoid non-linear activation function layer; the global pooling layer, the first 1×1 convolutional layer, the h-swish non-linear activation function layer, the second 1×1 convolutional layer, and the Sigmoid non-linear activation function layer in the channel attention branch are connected in sequence. The context attention branch includes: a Flatten flattening module, a third 1×1 convolutional layer, a Softmax unit, a matrix multiplication unit, and a fourth 1×1 convolutional layer; The third 1×1 convolutional layer, the Softmax unit, and the matrix multiplication unit are connected in sequence. The Flatten flattening module is connected to the matrix multiplication unit; The matrix multiplication unit is connected to the fourth 1×1 convolutional layer.

[0008] Optionally, step S3 further includes: S31: Select the cross-entropy loss function as the loss function of the non-local attention classification network; S32: Train the non-local attention classification network through the planetary train surface passability dataset and the cross-entropy loss function.

[0009] Optionally, step S4 includes: The steps of classifying the ground penetrating radar B-scan data to be classified through the trained deep learning model include: S41: Adjust the size of the ground penetrating radar B-scan data to be classified to H×W×C, where H, W, and C represent the height, width, and number of channels of the data cube respectively; S42: Extract the features of the ground penetrating radar B-scan data to be classified through the input convolutional unit of the trained deep learning model, and obtain feature F1; S43: Perform downsampling on feature F1 through the non-local attention module stacked in the deep learning model to obtain feature F2; S44: Input feature F2 into the global average pooling layer of the deep learning model to perform feature pooling on feature F2, and obtain the non-local attention feature map F3; S45: Classify and identify the non-local attention feature map F3 through the fully connected classification output layer of the deep learning model, and output the recognition result of the star catalog passability.

[0010] Optionally, step S43 includes: S431: Receive the input feature F1, extract the planetary subsurface anomaly information contained in the feature in the channel dimension through the channel attention branch, and output the first branch feature FCA; S432: Multiply the feature FCA obtained from the first branch by the feature F1 in matrix form, and output the channel attention feature FCA1; S433: Extract the non-local feature difference of feature F1 through the context attention branch, and output the second branch feature FGA2; S434: Add the feature FGA2 obtained from the second branch to the output FCA1 of the first branch in matrix form to obtain the final output feature F2 of the non-local attention module.

[0011] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory so that the electronic device executes a method for judging the passability of the planetary subsurface based on the ground penetrating radar data of the planetary rover.

[0012] A computer-readable storage medium stores instructions. When the instructions are executed, a method for judging the passability of the planetary subsurface based on the ground penetrating radar data of the planetary rover is executed.

[0013] The beneficial effects brought by the technical solution provided in this application are: Taking the speculation of the passability of the planetary subsurface of the planetary rover as the research goal, and to ensure the large-scale mobile detection task of the planetary rover, this application constructs a non-local attention classification network based on the ground penetrating radar B-scan data. The proposed non-local attention model extracts the non-local classification features of the input ground penetrating radar B-scan data through the staggered stacking of its non-local attention modules, and improves the discrimination effect of the planetary subsurface passability with a relatively low model parameter quantity. Description of the Drawings

[0014] The following will further illustrate the present application in conjunction with the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 is the flowchart of the method according to the embodiment of the present invention; Figure 2 shows the CH-2B radar data profile diagram after preprocessing such as data stitching, redundant data removal, background subtraction, gain, and zero-time correction for the lunar rover ground-penetrating radar data provided by the embodiment of the present invention; Figure 3 shows an example diagram of the lunar subsurface ground-penetrating radar B-scan data provided by the embodiment of the present invention; Figure 4 shows the lunar subsurface simulation model diagram provided by the embodiment of the present invention; Figure 5 shows the simulation result diagram of the lunar subsurface ground-penetrating radar data provided by the embodiment of the present invention; Figure 6 shows the fusion result diagram of the real ground-penetrating radar A-Scan data of the lunar rover and the simulated A-Scan data of the ground-penetrating radar provided by the embodiment of the present invention; Figure 7 shows the fusion result diagram of the simulated ground-penetrating radar B-scan data with added noise; Figure 8 shows the enhancement result diagram of the simulated ground-penetrating radar B-scan data provided by the embodiment of the present invention; Figure 9 shows the network structure diagram of the non-local attention ground-penetrating radar data passability discrimination network provided by the embodiment of the present invention; Figure 10 shows the specific structure diagram of the non-local attention module provided by the embodiment of the present invention; Figure 11 shows the network model hierarchical heat map provided by the embodiment of the present invention; Figure 12 is the schematic diagram of the electronic device structure in the embodiment of the present application. Specific Embodiments

[0015] In order to have a clearer understanding of the technical features, objectives, and effects of the present application, the specific embodiments of the present application will now be described in detail with reference to the accompanying drawings.

[0016] The embodiment of the present application provides a method for discriminating the passability of a planetary subsurface based on planetary rover ground-penetrating radar data.

[0017] Please refer to Figure 1 , Figure 1It is a step diagram of a method for discriminating the passability of the planetary subsurface based on the ground penetrating radar data in an embodiment of the present application, including: S1: Obtain the ground penetrating radar data of the planetary rover; construct a dataset for the passability of the planetary rover's subsurface according to the ground penetrating radar data of the planetary rover; S2: Construct a non-local attention classification network based on the ground penetrating radar B-scan data; S3: Iteratively train the non-local attention network with the dataset for the passability of the planetary rover's subsurface to obtain a trained deep learning model; S4: Obtain the ground penetrating radar B-scan data to be classified; classify the ground penetrating radar B-scan data to be classified through the trained deep learning model to obtain the recognition result of the surface passability, and complete the discrimination of the passability of the planetary subsurface corresponding to the ground penetrating radar data.

[0018] As an embodiment, input the ground penetrating radar B-scan data to be classified in the test set into the trained non-local attention passability discrimination network model. The specific network layers are shown in Table 2. Call the non-local attention passability discrimination network model proposed by the present invention for inference to obtain the recognition result of the ground penetrating radar passability, and realize the efficient and accurate recognition of the passability of the planetary subsurface of the ground penetrating radar data.

[0019] Table 2 Specifications and parameters of the non-local attention model

[0020] Regarding the specific structure of the ground penetrating radar data, the non-local attention module proposed by the present invention helps the network extract classification features in the ground penetrating radar data. Therefore, the present invention reduces the number of channels in each layer of the baseline model and removes the redundant settings for feature extraction in the network, and finally obtains the model structure shown in Table 1. Compared with the MobileNetV3 network, its number of parameters is reduced from the original 1.678M to 1.494M, and the experimental results also show that this improvement can reduce the risk of model overfitting to a certain extent.

[0021] The ground penetrating radar data is input into the network with a dimension of 224×224×3. The basic features of the data are extracted by the first-layer conventional convolution operation, and the semantic feature map with a size of 112×112 is obtained after dimensionality reduction. And in order to improve the running speed of the network and keep the model parameters from increasing too much, the input convolution operation only raises the feature dimension to 16. After the feature extraction and dimensionality reduction operations of each layer of the network, the high-level semantic feature information with a size of 7×7×576 is obtained. The subsequent classification operation performs global pooling on these high-level semantic features and uses 1×1 convolution to calculate the internal relationship of different features, compresses the number of features to the required number of classification k, and obtains the passability probability of the input ground penetrating radar data.

[0022] Table 3 Comparison test results of NLANet

[0023] Table 3 shows the results of the comparison test. It can be seen that the ground penetrating radar data non-local attention network model NLANet proposed by the present invention has certain performance improvement compared with MobileNetV3. Its classification accuracy on the ground penetrating radar passability data set has increased from 86.01 to 87.52, an increase of 1.51 points, and there is a breakthrough improvement in the number of model parameters. This indicates that the ground penetrating radar data non-local attention network model proposed by the present invention can achieve a level beyond that of similar models under the condition of low computing resources.

[0024] Meanwhile, in order to reflect the improvement of the non-local attention network (NLAN) in focusing on the non-local features of its input images, the "Gradient-Weighted Class Activation Mapping" (GradCAM) method is used in this section to visualize the feature activation performance before the output of each model. The results are as Figure 11 shown, where Figure 11 -a is an example of a test set sample, Figure 11 -b is the heatmap of the activation of the MobileNetV3 network. The highlighted heat part in the figure coincides with the hyperbola part of the ground penetrating radar defect, indicating that the model has the ability to focus on the semantic features of the defective star soil. However, compared with Figure 11 -c, the heatmap of the activation of VGG shown, the highlighted range of this feature focus is smaller and cannot accurately cover the entire defective hyperbola area. The heatmap of the activation of the model (NLAN) of the present invention Figure 11 -c shows that the non-local attention model can improve the focus range of key features, that is, the model extends its attention beyond local key features. This result clearly proves the effectiveness of the non-local attention proposed by the present invention, enabling the model to achieve a higher classification accuracy (ACC).

[0025] Step S2 includes: The non-local attention classification network includes: an input convolutional unit, a non-local attention module, a global average pooling layer, and a fully connected classification output layer; The input convolutional unit is connected to the non-local attention module; The non-local attention module is connected to the global average pooling layer; The global average pooling layer is connected to the fully connected classification output layer; The input convolutional unit includes: a 3×3 convolutional layer, a normalization module layer, and an h-swish non-linear activation function layer; the 3×3 convolutional layer, the normalization module layer, and the h-swish non-linear activation function layer are connected in sequence; As an embodiment, the specific input convolution unit is as follows: The ground penetrating radar B-scan data input into the model passes through an initial feature extraction stage composed of a 3×3 convolution layer, a normalization module layer, and an h-swish non-linear activation function layer, and a semantic feature map with a size of H / 2×W / 2×16 is obtained.

[0026] The global average pooling layer includes: a fully connected layer, a normalization layer, and an h-swish non-linear activation function layer; the global average pooling layer includes: the fully connected layer, the normalization layer, and the h-swish non-linear activation function layer are connected in sequence.

[0027] As an embodiment, the global average pooling layer performs average pooling on the results of the non-local attention module, reduces the redundant feature information, and then obtains the stage feature output. The global average pooling layer includes: a fully connected layer, a normalization layer, and an h-swish non-linear activation function layer. The data passing through the global average pooling layer is normalized by the normalization layer and undergoes non-linear transformation based on the h-swish activation function.

[0028] As an embodiment, the fully connected classification output layer: Based on the fully connected layer, finally, the analysis result of the passability of the planetary subsurface based on the ground penetrating radar data is classified and output.

[0029] As an embodiment, the input convolution unit is used to receive the input ground penetrating radar data; a preset number of non-local attention modules are stacked and connected to obtain the features of the input convolution unit; the global average pooling layer is connected to the low-dimensional features of the non-local attention module; the fully connected classification output layer receives the pooling result of the global average pooling layer and performs passability discrimination and output.

[0030] Step S2 further includes: The non-local attention module includes: a channel attention branch and a context attention branch; The channel attention branch and the context attention branch are connected in parallel; The channel attention branch includes: a global pooling layer, a first 1×1 convolution layer, an h-swish non-linear activation function layer, a second 1×1 convolution layer, and a Sigmoid non-linear activation function layer; the global pooling layer, the first 1×1 convolution layer, the h-swish non-linear activation function layer, the second 1×1 convolution layer, and the Sigmoid non-linear activation function layer in the channel attention branch are connected in sequence; The context attention branch includes: a Flatten flattening module, a third 1×1 convolution layer, a Softmax unit, a matrix multiplication unit, and a fourth 1×1 convolution layer; The third 1×1 convolution layer, the Softmax unit, and the matrix multiplication unit are connected in sequence; The Flatten module is connected to the matrix multiplication unit; The matrix multiplication unit is connected to the fourth 1×1 convolutional layer.

[0031] As an embodiment, the non-local attention module structure includes: a channel attention branch and a context attention branch. Among them, the channel attention branch includes a global pooling layer, a 1×1 convolutional layer, an h-swish non-linear activation function layer, a 1×1 convolutional layer, and a Sigmoid non-linear activation function layer.

[0032] As an embodiment, the context attention branch first takes the input and calculates it into a single-channel fusion feature of size H×W×1 using pointwise convolution. After that, the feature dimension is reconstructed to H*W×1×1 so that it can participate in the operation in vector form, and the feature output is obtained after activation with Softmax. In addition, the original input X is flattened channel by channel to obtain a set of channel vectors of size C×H*W. The outputs of the two branches are regarded as keys and values and are multiplied by matrix. After the multiplication result is convolved through a 1×1 convolutional layer, the importance degree of each feature position is obtained. Adding this result to the input X point by point can attach the context attention to the original feature map.

[0033] Step S3 also includes: S31: Select the cross-entropy loss function as the loss function of the non-local attention classification network; S32: Train the non-local attention classification network through the planetary train surface passability dataset and the cross-entropy loss function.

[0034] In one example, the non-local attention passability discrimination network is iteratively trained to optimize the parameter values of the model so that the loss function value continuously approaches zero, thereby obtaining a model with good ground-penetrating radar data recognition effect; the present invention selects the cross-entropy loss function as the direction of network optimization, and the calculation formula of this loss function is

[0035] where N is the number of samples; is the probability distribution of the prediction result, and the value of each element is between [0,1]; is the one-hot representation of the sample label; First, initialize the parameters of the model. Use the training dataset to iteratively train the non-local attention passability discrimination network for ground penetrating radar. The training uses the AdamW optimizer to adjust the learning rate. The initial learning rate for training is 0.0001. Calculate the loss function value and perform gradient backpropagation every time a forward propagation is completed. Continuously optimize the loss using the gradient descent method. Finally, obtain a deep learning model that fits the data distribution of the training set. The output result will obtain the predicted probability value of the passability of the ground penetrating radar data through the Softmax activation function, and take the category with the highest predicted probability as the final category.

[0036] Step S4 includes: The steps of classifying the ground penetrating radar B-scan data to be classified through the trained deep learning model include: S41: Resize the size of the ground penetrating radar B-scan data to be classified to H×W×C, where H, W, and C represent the height, width, and number of channels of the data cube respectively; S42: Extract the features of the ground penetrating radar B-scan data to be classified through the input convolutional unit of the trained deep learning model to obtain feature F1; S43: Perform downsampling on feature F1 through the stacked non-local attention modules of the deep learning model to obtain feature F2; S44: Input feature F2 into the global average pooling layer of the deep learning model to perform feature pooling on feature F2 to obtain the non-local attention feature map F3; As an embodiment, perform global average pooling operation on feature F2 with dimensions of H×W×C to obtain the non-local attention feature map F3 with dimensions of 1×C.

[0037] S45: Classify and recognize the non-local attention feature map F3 through the fully connected classification output layer of the deep learning model, and output the recognition result of the star catalog passability.

[0038] As an embodiment, the working process of the ground penetrating radar data passability discrimination network based on non-local attention is as Figure 9 shown.

[0039] Step S43 includes: S431: Receive the input feature F1, extract the planetary subsurface anomaly information contained in the feature in the channel dimension through the channel attention branch, and output the first branch feature FCA; As an example, the input feature F1 with dimensions H×W×C is input into the global pooling module to obtain the feature F1G of 1×1×C / r, where r is the channel stride. Then, the feature F1G is input into a 1×1 convolutional layer to obtain the 1×1×C / r feature F1GC. Next, a non-linear transformation is performed using the h-swish activation function to obtain the 1×1×C / r feature F1GCh. Then, a 1×1 convolution is performed to obtain the 1×1×C / r feature F1GChC. Further, a non-linear transformation is performed using the Sigmoid activation function to output the first branch 1×1×C feature FCA (as Figure 10 )

[0040] S432: Matrix multiply the feature FCA obtained from the first branch with the feature F1 to output the channel attention feature FCA1; In one example, the 1×1×C feature FCA obtained from the first branch is matrix multiplied with the size H×W×C input F1 to output the size H×W×C channel attention feature FCA1; S433: Extract the non-local feature differences of the feature F1 through the context attention branch and output the second branch feature FGA2; In one example, the input feature F1 with dimensions H×W×C is input into a 1×1 pointwise convolutional layer to obtain the single-channel fusion feature F2C of size H×W×1. Then, the single-channel fusion feature F2C is reconstructed through a flattening operation into a feature vector F2CF with a feature dimension of H*W×1×1. Further, the input feature F1 is flattened to obtain the feature F1F with a dimension of H*W×C. The F1F feature and the feature vector F2CF are regarded as the key Q∈R^(H*W×1×1) and the value KV∈R^(C×H*W) for matrix multiplication. After the multiplication result is convolved through a 1×1 convolutional layer, the importance degree of each feature position is obtained, and the global context feature FGA2 with a dimension of feature H×W×C is output; S434: Matrix add the feature FGA2 obtained from the second branch with the output FCA1 of the first branch to obtain the final output feature F2 of the non-local attention module.

[0041] In one example, the dimension of the global context feature FGA2 obtained from the second branch is H×W×C, the dimension of the channel attention feature obtained from the first branch is H×W×C, and the dimension of the output feature F2 is H×W×C.

[0042] As an example, the process of downsampling the feature F1 through the stacked non-local attention module to obtain the feature F2 is as Figure 10 shown.

[0043] As an example, a dataset of the passability of the planetary vehicle's subsurface is constructed based on the data of the planetary vehicle's ground-penetrating radar as follows: (1)Interpretation and analysis of the terrain and lunar penetrating radar data in the landing area of Chang'e-4 The Chang'e-4 probe successfully landed at the bottom of the Von Kármán crater in the oldest and largest South Pole–Aitken basin on the far side of the Moon at 10:26:00 on January 3, 2019 Beijing Time. According to the in-situ detection data of the lunar penetrating radar carried on the Yutu-2 lunar rover of Chang'e-4, within a depth of 40 meters underground in the landing area on the far side of the Moon, the geological stratification structure is such that no large-scale abnormal areas (such as cavities, faults, etc.) affecting the passability of the planetary vehicle are found in the subsurface. Therefore, it is necessary to add negative sample lunar penetrating radar signal data that the planetary vehicle cannot pass through through lunar penetrating radar simulation experiments. That is, it is necessary to conduct a forward simulation test of the ground-penetrating radar on the terrain that causes changes in the magnetoelectric properties (dielectric constant ε_r, conductivity σ) of the lunar regolith. Finally, it is determined that the simulated lunar regolith geological structure includes three terrains: cavities in the shallow lunar regolith, large cavities in the deep lunar regolith, and sudden changes in lunar regolith density. And to ensure the balance of the simulated samples, three passable terrains of lunar regolith with a slope, lunar regolith with small cavities in the deep layer, and geologically uniform lunar regolith are also simulated.

[0044] (2)Preprocessing of lunar penetrating ground-penetrating radar data The CH-2 data of the lunar penetrating radar of Chang'e-4 is used. Before the analysis and interpretation of the data, a series of preprocessing is required to obtain the original lunar penetrating radar profile. Specifically, it includes data splicing, redundant data removal, background subtraction, gain, and zero-time correction. The preprocessing results of the lunar penetrating ground-penetrating radar data are as Figure 2 shown. (3)Visualization of ground-penetrating radar data The high-frequency lunar penetrating radar (LPR) of Yutu-2 consists of a set of files of continuous A-Scan measurement point waveforms. To visualize the 2C-level ground-penetrating radar data of Yutu-2, steps such as self-check trace removal, ringing noise removal, repeated trace data deletion, signal horizontal energy equalization, signal gain, band-pass filtering, and time interval adjustment are performed on the radar waveform data to remove the noise in the original data and amplify the effective detection features of the lunar regolith terrain, obtaining the organized and processed A-Scan set data, which is stacked and arranged as B-Scan data and visualized as shown in Figure 3For the lunar subsurface shown, the abscissa of each piece of data shows the information of the ground penetrating radar in the subsurface with a detection path length of 5 meters and an effective detection depth of 4.5 meters for this data. The geological state of the detected area shown by the data is uniform, and there is no abnormal mutation area of low-density lunar regolith. Moreover, the lunar rover travels smoothly during the detection without subsidence caused by subsurface obstacles. Therefore, the common mutations of the ground penetrating radar signals in these data cannot be used as non-passable counterexamples in the study of planetary subsurface passability.

[0045] (4)Simulation of Lunar Surface Ground Penetrating Radar Data Table 1 Parameters of Yutu-2 Lunar Rover

[0046] According to the parameters of Yutu-2 in Table 1 and the installation position of the ground penetrating radar, the horizontal length and vertical depth of the simulated terrain are both set to 5 meters, which can completely contain the body length of Yutu-2 and adapt to the detection depth of the lunar radar on Yutu-2 (0 - 10m, with an effective detection depth of about 5m). At the same time, common lunar surface gravels are added to various terrains to increase the relative authenticity of the subsurface terrain. Specific examples are as Figure 4 shown. The yellow area in the image represents the lunar subsurface regolith with a uniform texture, and there are pits of various shapes and sizes on its shallow subsurface (as shown by the white circles in Figure 4 ), red polygon lunar gravels are distributed in the deeper regolith, and the height of the radar antenna of the simulation model is set to 30 cm, which is the installation height of the lunar radar on Yutu-2.

[0047] Then, ground penetrating radar data simulation is carried out on the lunar subsurface simulation model. The source parameters of the simulated data are determined according to the hardware parameters of the ground penetrating radar of the Yutu-2 rover, that is, a Ricker wavelet with a center frequency of 500 MHZ is used, and 75 points of ground penetrating radar echo signals are sampled during a 5-meter moving distance, and then they are arranged in order to form a ground penetrating radar B-scan simulation image. Some subsurface simulation models and the corresponding GPR simulation results are as Figure 5 shown. For example, in Figure 5 -a and 5-b, there are low-density lunar regolith cavities of different shapes and sizes in the shallower subsurface from 0 to 0.5 meters, and gravels are randomly distributed in the deeper regolith from 2 to 5 meters; Figure 5 -c and 5-d, there are lunar regolith low-density cavities with a diameter of about 30 cm in the shallow lunar regolith; Figure 5 -e has low-density cavities less than 30 cm in the deeper lunar regolith from 3 to 5 meters. This situation will not cause subsidence to the movement of the lunar rover. This example is a simulation example of a passable situation. Figure 5 -f sets the lunar surface to be rough but the lunar regolith density is uniform, which is an example of a passable subsurface simulation model.

[0048] (5)Enhancement of Ground Penetrating Radar A-Scan Data The data of the ground penetrating radar (LPR) of Yutu-2 has revealed the geological structure characteristics of the lunar subsurface to a certain extent. To provide real lunar ground penetrating radar lunar regolith information for the passability discrimination network model proposed by the present invention and ensure that the final samples have the geological characteristics of the defective subsurface environment, the present invention innovates the ground penetrating radar data processing method and uses the fast Fourier transform (FFT) and its inverse transform (IFFT) to fuse the ground penetrating radar simulation data with the real lunar data.

[0049] The present invention uses the ground penetrating radar A-Scan data as the operation object, and performs Fourier transform on the simulated ground penetrating radar data (GPR) and the real subsurface ground penetrating radar data (LPR) respectively, converting the ground penetrating radar A-scan signal from the time domain to the frequency domain to obtain Figure 6 the data frequency domain representations shown in Figures 6-b and 6-d. Then, after removing the extremely low frequency noise in the data in the frequency domain space, the two columns of data are weighted and summed to obtain the synthesized A-scan frequency domain signal. Using the inverse fast Fourier transform (IFFT), the fused frequency domain data is restored to the time domain space, and the fusion result of a single A-scan signal as shown in Figure 6 Figures 6-a and 6-c can be obtained. The above operations are looped, and the simulation signals are fused with the real lunar GPR signals in sequence according to the arrangement order of the simulation signals and stitched into the ground penetrating radar B-Scan data.

[0050] This process can be expressed by the following formula:

[0051] where FFT and IFFT respectively refer to the fast Fourier transform and the inverse fast Fourier transform; m is the A-scan set of the simulated GPR data, and n is the A-scan set of the real lunar regolith; and represent the fusion ratio of the simulated data and the real data; out represents the ground penetrating radar B-scan data composed of all fused A-scan combinations.

[0052] Therefore, in order to provide real lunar GPR lunar regolith information for the deep learning model and ensure that the final samples have the geological characteristics of the impassable subsurface such as cavities and pores, the present invention innovates the ground penetrating radar data processing method and uses the fast Fourier transform (FFT) and its inverse transform (IFFT) to fuse the GPR simulation data with the real lunar data. Among them, Figure 7 Figures 7-a and 7-e are the real B-scan data of the LPR of Chang'e-4, Figure 7 Figures 7-b and 7-f are the ground penetrating radar data simulated by the present invention according to the geological conditions of the real lunar subsurface, Figure 7-c and 7-g are the fusion results obtained by using the GPR A-scan data superposition enhancement method of the present invention, while Figure 7 -d, Figure 7 -h is the result directly obtained by superposing GPR B-scan data. It can be seen that the GPR data enhancement method adopted by the present invention adds more real lunar subsurface data to the simulated data, and this real data can reflect the semantic features of the GPR in the lunar regolith subsurface to a certain extent. Moreover, adopting the GPR A-scan data enhancement method of the present invention can also balance the problem of huge differences in semantic features between positive and negative samples in this research, and improve the generalization ability of the deep learning network model for passability discrimination.

[0053] (6)Simulated GPR B-scan data enhancement Data augmentation involves operating on the original samples to expand the dataset size and data diversity. In the case where it is difficult to obtain more dataset samples, as a data resource augmentation means for deep learning algorithms, it is helpful to improve the model generalization ability and prevent overfitting during training. The present invention performs 8 data augmentation operations on the GPR B-scan sample data, including random horizontal stretching, random vertical stretching, random cropping, horizontal flipping, random erasing, brightness jitter, and contrast enhancement, while trying to keep the real lunar surface information contained in the dataset undamaged during the operation ( Figure 8 ). And the model training data preparation is completed according to the division method of 70% training set and 30% test set.

[0054] This application also discloses an electronic device. Referring to Figure 12 , Figure 12 is a schematic structural diagram of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0055] Among them, the communication bus 502 is used to realize the connection and communication between these components.

[0056] Among them, the user interface 503 may include a display screen. Optionally, the user interface 503 may further include a standard wired interface and a wireless interface.

[0057] Among them, the network interface 504 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0058] This application also discloses a computer-readable storage medium, which stores multiple instructions suitable for being loaded by a processor to execute the above-mentioned method for discriminating the passability of a planetary subsurface based on planetary rover GPR data.

[0059] The above are only exemplary embodiments of the present disclosure and should not be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure still fall within the scope covered by the present disclosure.

[0060] This application is intended to cover any variations, uses, or adaptations of the present disclosure, which follow the general principles of the present disclosure and include known common general knowledge or conventional technical means in the technical field not recorded in the present disclosure. The description and examples are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for determining the traversability of a planetary subsurface based on ground penetrating radar data from a planetary rover, characterized in that: The method comprises the following steps: S1: Obtain the ground-penetrating radar data of the planetary rover; construct the sub-surface accessibility dataset of the planetary rover based on the ground-penetrating radar data of the planetary rover; S2: Construct a non-local attention classification network based on ground penetrating radar B-scan data; S3: Iteratively train the non-local attention network using the planetary vehicle surface accessibility dataset to obtain a trained deep learning model; S4: Obtain the GPR B-scan data to be classified; classify the GPR B-scan data to be classified through the trained deep learning model, obtain the recognition result of the star catalog passability, and complete the judgment of the planetary subsurface passability of the area corresponding to the GPR data.

2. A method for determining the traversability of a planetary subsurface based on ground penetrating radar data of a planetary rover according to claim 1, characterized in that: Step S2 includes: The non-local attention classification network includes: an input convolution unit, a non-local attention module, a global average pooling layer, and a fully connected classification output layer; The input convolution unit is connected to the non-local attention module; The non-local attention module is connected to the global average pooling layer; The global average pooling layer is connected to the fully connected classification output layer; The input convolution unit includes: a 3×3 convolution layer, a normalization module layer, and an h-swish nonlinear activation function layer; the 3×3 convolution layer, the normalization module layer, and the h-swish nonlinear activation function layer are connected in sequence; The global average pooling layer includes: a fully connected layer, a normalization layer, and an h-swish non-linear activation function layer; the global average pooling layer includes: a fully connected layer, a normalization layer, and an h-swish non-linear activation function layer connected in sequence.

3. A method for determining the traversability of a planetary subsurface based on ground penetrating radar data of a planetary rover as claimed in claim 2, characterized in that: Step S2 also includes: The non-local attention module includes: channel attention branch and context attention branch; The channel attention branch is connected in parallel with the context attention branch; The channel attention branch includes: a global pooling layer, a first 1×1 convolution layer, an h-swish nonlinear activation function layer, a second 1×1 convolution layer, and a Sigmoid nonlinear activation function layer; the global pooling layer, the first 1×1 convolution layer, the h-swish nonlinear activation function layer, the second 1×1 convolution layer, and the Sigmoid nonlinear activation function layer in the channel attention branch are connected in sequence; The context attention branch includes: Flatten module, the third 1×1 convolutional layer, Softmax unit, matrix multiplication unit, and the fourth 1×1 convolutional layer; The third 1×1 convolutional layer, the Softmax unit, and the matrix multiplication unit are connected sequentially; The Flatten module connects the matrix multiplication unit; The matrix multiplication unit is connected to the fourth 1×1 convolutional layer.

4. A method for determining the traversability of a planetary subsurface based on ground penetrating radar data of a planetary rover according to claim 1, characterized in that: Step S3 also includes: S31: Select the cross entropy loss function as the loss function of the non-local attention classification network; S32: Training a non-local attention classification network on the planetary subsurface accessibility dataset using the cross-entropy loss function.

5. The method for determining the traversability of a planetary subsurface based on ground penetrating radar data of a planetary rover according to claim 1, characterized in that: Step S4 includes: The steps of classifying the unclassified GPR B-scan data using the trained deep learning model include: S41: adjusting the size of the ground penetrating radar B-scan data to be classified to H×W×C, where H, W, and C represent the height, width, and number of channels of the data cube, respectively; S42: extracting the features of the ground penetrating radar B-scan data to be classified through the input convolution unit of the trained deep learning model to obtain feature F1; S43: Down-sampling the feature F1 through the non-local attention module stacked by the deep learning model to obtain the feature F2; S44: Input the feature F2 into the global average pooling layer of the deep learning model, perform feature pooling on the feature F2, and obtain a non-local attention feature map F3; S45: The non-local attention feature map F3 is classified and identified through the fully connected classification output layer of the deep learning model, and the recognition result of the star catalog accessibility is output.

6. A method for determining the traversability of a planetary subsurface based on ground penetrating radar data of a planetary rover as claimed in claim 5, characterized in that: Step S43 includes: S431: receiving input feature F1, extracting the planetary subsurface anomaly information contained in the feature in the channel dimension through the channel attention branch, and outputting the first branch feature FCA; S432: Perform matrix multiplication on the feature FCA obtained from the first branch and the feature F1, and output the channel attention feature FCA1; S433: extracting the non-local feature difference of feature F1 through the context attention branch, and outputting the second branch feature FGA2; S434: Perform matrix addition on the feature FGA2 obtained by the second branch and the output FCA1 of the first branch to obtain the feature F2 of the final output of the non-local attention module.

7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images