Height searching method based on improved VGG16 height finding radar

By using VGG16 neural network and WOA-CNN-LSTM-Attention model in the altimeter radar, the time and frequency domain characteristics of the echo signal were extracted respectively, and the problem of poor high-precision measurement in complex terrain environments was solved and high-precision was achieved Highly high information extraction and anti-interference ability improvement.

CN119087379APending Publication Date: 2024-12-06NORTHWESTERN POLYTECHNICAL UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411171316.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing altimeter radar has poor anti-interference capability in complex terrain environments, low altitude measurement accuracy, and it is difficult to accurately extract echo information.

Method used

The improved height search method of altitude measurement radar based on VGG16 neural network is adopted, combined with the WOA-CNN-LSTM-Attention model, the time domain characteristics and frequency domain characteristics of the echo signal are extracted and integrated to improve the accuracy of the measurement.

Benefits of technology

Effectively overcome the altitude measurement error caused by complex terrain, improve the altitude information extraction capability, and meet the high-precision altitude measurement needs of unmanned aerial vehicles in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119087379A_ABST
    Figure CN119087379A_ABST
Patent Text Reader

Abstract

The invention provides a VGG16 height finding radar height searching method based on improvement, a WOA-CNN-LSTM-Attention model and a VGG16 network model are built, time domain features and frequency domain features in echo signals are respectively extracted through two different neural networks, and then data extracted by the two networks are fused together for training and testing. The network model is equivalent to a double-input network model, actually-measured data are converted into sequences and images to be input into the network, features are extracted on the two levels of the time domain and the frequency domain, the output accuracy is improved, clutters are effectively restrained, and therefore the height measurement precision of the height measurement radar of the unmanned aerial vehicle in different complex environments is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of aviation altimetry, and in particular relates to an altitude search method based on an improved VGG16 altimetry radar, which is suitable for altitude search of an unmanned aerial vehicle altimetry radar in a complex environment. Background Art

[0002] Altimeter radar is an indispensable core equipment for aerospace vehicles. It is used to measure the relative height between the aircraft and the ground or water surface and is widely used in various fields of aerospace. There are three main methods for radar altimeter altimeter measurement: phase method, pulse method and frequency method. Phase method altimeter is to obtain the altitude by measuring the phase difference between the reflected echo and the transmitted wave. It is mainly used in laser radar and is widely used in high-precision instruments; but there is a problem of height ambiguity. Pulse method altimeter is to obtain the altitude by directly measuring the delay time of the echo pulse relative to the transmitted pulse. The pulse method receiving antenna and the transmitting antenna are shared, the transmission power can be relatively large, the working distance is relatively long, and it is often used for long-distance altitude measurement; but there is a distance blind spot and the measurement resolution is relatively low. Frequency method altimeter is to use the frequency difference between the echo signal and the transmitted signal at the same time, that is, to determine the altitude by the frequency of the beat signal. There is no distance blind spot in frequency method altimeter and it has extremely high distance resolution. Secondly, the transmission power is small and the signal is not easy to be intercepted, but the height measurement accuracy is poor in complex environments.

[0003] The radar altimeter targets the ground, which is a complex surface target. The echo information reflected by the ground is very rich. In addition, during the flight of the aircraft, there may be various abnormal conditions or interference, making it difficult to accurately extract the altitude information during spectrum extraction, thus affecting the accuracy of altitude measurement. In complex terrain scenes, the altitude radar will also be interfered by various undulating terrains, making it more difficult to extract altitude information.

[0004] Therefore, whether the useful part of the echo information can be extracted continuously and accurately and whether it can have strong resistance to interference are issues that must be considered in the design of radar altimeters. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of existing altimetry radar search methods, such as poor anti-interference ability and poor altitude measurement accuracy when measuring altitude in complex terrain, and provide an altitude search method based on the improved VGG16 altimetry radar. On the basis of the VGG16 neural network, a whale group convolutional long short-term memory neural network (WOA-CNN-LSTM-Attention) based on the attention mechanism is added. Two neural networks are used to extract the time domain characteristics and frequency domain characteristics of the echo signal data respectively, which can overcome the altitude measurement error caused by complex terrain, effectively improve the altitude information extraction capability, thereby improving the altitude measurement accuracy, and can meet the high-precision altitude measurement needs of unmanned aerial vehicles in complex environments.

[0006] To achieve the above purpose, the technical solution provided by the present invention is:

[0007] The height search method based on the improved VGG16 altimeter radar is special in that it includes the following steps:

[0008] Step 1, collecting echo signal data of the experimental ground through a height measuring radar, and constructing a training set and a test set of a data sequence type, and a training set and a test set of an image sequence type according to the echo signal data;

[0009] Step 2: construct a WOA-CNN-LSTM-Attention network model, perform training and prediction using the training set and test set of the data sequence type obtained in step 1, and obtain the time domain feature data of the echo signal;

[0010] Step 3, construct a VGG16 network model, perform training and prediction using the training set and test set of the image sequence type obtained in step 1, and obtain frequency domain feature data of the echo signal;

[0011] Step 4, integrating the time domain feature data obtained in step 2 and the frequency domain feature data obtained in step 3, so that each frame of feature data output by the two network models corresponds one-to-one with the actual height data of the experiment, and obtaining integrated data;

[0012] Step 5: Input the integrated data into the hybrid model of the WOA-CNN-LSTM-Attention network model and the VGG16 network model for training and prediction to obtain the actual height information.

[0013] Furthermore, the specific process of constructing a training set and a test set of a data sequence type according to the echo signal data in step 1 is as follows:

[0014] First, the collected echo signal data is re-inserted;

[0015] Then, the re-inserted data is imported into Excel, and the dimensions corresponding to the collected echo signal data are;

[0016] Again, the collected echo signal data is reconstructed in Excel, and a data set with 100,000 data items is obtained after reconstruction. The input data dimension of the data set is and the output data dimension is;

[0017] Finally, the reconstructed data set is divided into training set and test set.

[0018] Furthermore, the specific process of constructing a training set and a test set of an image sequence type according to the echo signal data in step 1 is as follows:

[0019] First, the collected echo signal data is re-inserted;

[0020] Then, the re-interpolated data is transformed by CZT, and the corresponding spectrum is plotted to obtain a CZT spectrum image;

[0021] Thirdly, the CZT spectrum image is divided into ten types of data samples according to the front position of the spectrum, with a total of 100,000 frames of data; each frame of data represents a frame of CZT spectrum image, and 100,000 frames of CZT spectrum images are obtained;

[0022] Finally, the obtained CZT spectrum images are divided into training sets and test sets, and batch label annotations are performed on the training sets and test sets so that the data, spectrum images and classification categories corresponding to each frame correspond one to one.

[0023] Furthermore, the training set and the test set of the image sequence type are both mixed data of ten types of data samples.

[0024] Furthermore, the specific steps of step 2 are:

[0025] Step 2.1, constructing a WOA-CNN-LSTM-Attention network model, wherein the WOA-CNN-LSTM-Attention network model includes a WOA part, a CNN-LSTM-Attention part, and a data part;

[0026] Step 2.2: Input the training set of the data sequence type obtained in step 1 to the WOA-CNN-LSTM-Attention network model. The WOA part transmits the parameter vector of the training set to the CNN-LSTM-Attention part. The CNN-LSTM-Attention part decodes the parameter vector to obtain the corresponding specific parameter information. The data part trains the network with the training set data.

[0027] Step 2.3, predict the trained WOA-CNN-LSTM-Attention network model using the test set data;

[0028] Step 2.4, after prediction, calculate the mean square error between the predicted output value and the actual output value, and pass the mean square error to the WOA part, iteratively calculate the fitness value, and take the minimum fitness value obtained at the end of the iteration as the optimal result to obtain the corresponding optimal network hyperparameters;

[0029] In step 2.5, the WOA-CNN-LSTM-Attention network model is tested with the optimal network hyperparameters and test set data to obtain the time domain feature data of the echo signal.

[0030] Furthermore, step 4 includes the following steps:

[0031] Step 4.1, convert the time domain feature data and frequency domain feature data obtained in step 2 and step 3 into cell arrays respectively through the mat2cell function in Matlab;

[0032] Step 4.2, first convert the experimental ground truth height data into a categorical array using the categorical function, and then convert the categorical array into a cell array using the mat2cell function;

[0033] In step 4.3, the three cell arrays containing the data sequence, image sequence, and classification array are put into one cell array, so that the sequence and image of each frame of data can correspond one-to-one with the experimental ground truth height data.

[0034] The advantages of the present invention are:

[0035] The height search method based on the improved VGG16 altimeter radar proposed in the present invention builds a WOA-CNN-LSTM-Attention model and a VGG16 network model, takes advantage of the ability of neural networks to process large-scale data, and uses two different neural networks to respectively extract the time domain features and frequency domain features in the echo signal, thereby improving the height measurement accuracy of the unmanned aerial vehicle altimeter radar in different complex environments.

[0036] In the feature extraction stage of the height search of the altimeter radar, the WOA-CNN-LSTM-Attention network is added on the basis of the VGG16 neural network. The two networks extract the time domain features and frequency domain features of the data respectively, and then the data extracted by the two networks are fused together for training and testing. It is equivalent to a dual-input network, which converts the measured data into sequences and images and inputs them into the network respectively, extracts features in the time domain and frequency domain, thereby improving the accuracy, effectively suppressing clutter, and improving the height measurement accuracy of the altimeter radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0038] Figure 1 It is a structural diagram of the WOA-CNN-LSTM-Attention model constructed by the present invention;

[0039] Figure 2 It is a structural diagram of the VGG16 model constructed by the present invention;

[0040] Figure 3 It is the overall framework diagram of the network of the present invention;

[0041] Figure 4 It is the VGG16 model training diagram in the present invention, where the upper figure is the curve of accuracy changing with the increase of iteration number, and the lower figure is the curve of loss changing with the increase of iteration number;

[0042] Figure 5 is a schematic diagram of height measurement results in a flat ground environment in the embodiment;

[0043] Figure 6 is a schematic diagram of height error results in a flat environment in the embodiment;

[0044] Figure 7 is a schematic diagram of height measurement results in a grass environment in the embodiment;

[0045] Figure 8 Schematic diagram of height error results in a grass environment in the embodiment. DETAILED DESCRIPTION

[0046] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0047] This embodiment adopts an improved VGG16 altimeter radar height search method proposed by the present invention, and then inputs the neural network to extract the height information, which specifically includes the following steps:

[0048] Step 1: Collect experimental ground target height data through radar transceiver to build test set and training set.

[0049] The experiment uses a radar transceiver with a starting frequency of 4.3GHz and a frequency modulation bandwidth of 150MHz to collect data from outdoor ground targets.

[0050] Since the WOA-CNN-LSTM-Attention network dataset is a data sequence, and the VGG16 network dataset is an image sequence, it is necessary to process the collected data to obtain test sets and training sets of different data types.

[0051] (1) Process the data sequence required to obtain the WOA-CNN-LSTM-Attention network model.

[0052] First, the collected data is re-inserted and then imported into Excel. The dimension of the original data is (100000, 6). After reconstructing the original data, the input data dimension is (100000, 11) and the output data dimension is (100000, 1), which successfully constructs a data set with 100,000 data. The next step is to divide the obtained data into a training set and a test set: 70,000 data are selected as the training set and 30,000 data are selected as the test set.

[0053] (2) Build the image dataset required for the VGG16 network model.

[0054] The same collected data is re-inserted, and then the processed data is CZT transformed, and finally the corresponding spectrum diagram is drawn. The present invention divides the CZT spectrum image into ten categories of data samples according to the position of the spectrum front. A total of 100,000 frames of data are collected for the ten categories of data samples. Each frame of data consists of 1 chirp, and each chirp samples 512 points, that is, each frame of data represents a frame of CZT transformed spectrum image. According to the ratio of 7:3 between the training set and the test set, a total of 100,000 pictures are divided, and both the test set and the training set are mixed data of ten categories of data samples. The data set samples are batch labeled and annotated using Python language. The neural network model belongs to a supervised learning algorithm, so the data needs to clearly identify its information before it can be input into the network for training and learning. Finally, 70,000 training sets of 224×224 and 30,000 test sets of 224×224 are obtained. Finally, the data corresponding to each frame, the corresponding spectrum diagram and the corresponding classification category are matched one by one.

[0055] Step 2, refer to Figure 1 , build a whale group convolutional long short-term memory neural network (WOA-CNN-LSTM-Attention) model based on the attention mechanism, input the training set and test set obtained in step 1 for training and testing, and extract the time domain features of the echo signal;

[0056] like Figure 1As shown in the figure, the WOA-CNN-LSTM-Attention network model is built. The WOA-CNN-LSTM-Attention model is generally divided into three parts: WOA part, CNN-LSTM-Attention part and data part.

[0057] First, the CNN-LSTM-Attention part decodes the parameter vector sent by the WOA part to obtain specific parameters, and then uses the data part to pass the training set to the network for training;

[0058] Then, the trained model is used to predict the test set, and then the mean square error between the actual output value and the predicted output value is calculated, and the mean square error is passed to the WOA part to calculate the fitness value to update the population and the global optimal solution. The actual output value is the actual height, which can be obtained by simulation experiments or measured by instruments in actual tests.

[0059] Repeat the above process until the end condition is met. Using this method, you can get the optimized network hyperparameters, which can make the entire network output the predicted value closest to the actual value, and finally use the optimized network hyperparameters for subsequent network training and testing.

[0060] The whale optimization algorithm imitates the unique bubble net hunting mode of whales. The whale optimization algorithm breaks down the algorithm into three stages: encirclement hunting, spiral update and prey search. First, in the prey search stage, the whale slowly obtains information about the prey, then the whale will approach the prey in a certain way, surround the prey, and finally find the prey, thus finding the best way to solve the problem. This method is the spiral swimming mode.

[0061] When whales are looking for food, if they do not know the exact location of the food, or detect that the food is very close to it, they will communicate with the group and exchange information to learn the location of the whale closest to the prey (the current optimal solution). At this time, the entire group of whales will move towards the prey. At the same time, the whale closest to the prey will try to get closer to the prey in a certain way. In WOA, assuming that the whale population is N and the search space is d-dimensional, the position of the i-th whale in the d-dimensional space can be expressed as X i =(X i1 ,X i2 ,X i3 ,...,X id )(i=1,2,3,...,N). Based on the imitation of whale's predation behavior, the predation model of the encirclement stage can be expressed as follows.

[0062] D=C×X p (txt)

[0063] X(t+1)=X p (t)-A×D

[0064] In the formula, t represents the current iteration number; X(t) represents the individual position vector; X p (t) is the position of the prey; A and C are coefficient vectors, specifically:

[0065] A=2a×r1-a

[0066] C=2×r2

[0067] In the formula, r1 and r2 are random numbers between the interval [0,1], and a is a control parameter that is related to the number of iterations and decreases linearly. The specific expression of a is as follows: max Refers to the maximum number of iterations.

[0068]

[0069] In the spiral update phase, the algorithm simulates that when a whale finds a target, it will keep rotating around the prey and gradually approach the target to capture it. Before it surrounds the prey, it first calculates the distance between itself and the target, and then surrounds the prey with a spiral swimming path. Its mathematical model is as follows:

[0070] X(t+1)=D×e bl ×cos(2πl)+X P (t)

[0071] Where b is a constant used to limit the shape of the logarithmic spiral; l is a random number ranging from -1 to 1.

[0072] In order to coordinate the shrinking and spiraling updates, the probability P is used to determine which stage the whale's position update belongs to. The range of P is the interval [0,1], and its mathematical model is:

[0073]

[0074] The value of |A| is used to determine whether it is in the stage of searching for food or in the stage of encirclement and predation. If |A|>1, it means that the whale cannot get any useful information at this time and will enter the search stage, using the random walk method to try to get information about prey. The search model is as follows:

[0075] D=|C×X rand (t)-X(t)|

[0076] X(t+1)=X rand (t)-A×D

[0077] Where, X rand represents a position randomly selected from the current population position.

[0078] The problem of finding the optimal solution of the whale algorithm is actually the problem of finding the maximum or minimum value of a function. In the present invention, the classification results of the convolutional long short-term memory neural network (CNN-LSTM-Attention) model based on the attention mechanism and the mean square error of the actual output are used as the fitness function of the problem. So far, the whale optimization problem is to find a set of network hyperparameters so that the test results obtained by the CNN-LSTM-Attention network have the highest accuracy. This is the core idea of ​​optimizing the WOA-CNN-LSTM-Attention model. The WOA-CNN-LSTM-Attention process is as follows.

[0079] Step 1: Divide the input variables and output variables in the data into training sets and test sets according to corresponding proportions.

[0080] Step 2: First, initialize the parameters, including the initial learning rate, the number of training rounds, the learning reduction factor, and the learning rate reduction period, and select the upper and lower bounds of each parameter.

[0081] Step 3: Randomly construct n whales, decode the parameters obtained in the previous step, and establish a model in one-to-one correspondence with the network.

[0082] Step 4: Take the mean square error between the classification result value predicted by the model and the true value as the fitness value, calculate the fitness value of each whale, take the minimum fitness value as the best result, and compare it with the global best result. If the fitness value of the best result is smaller than the fitness value corresponding to the global best result, then the best result is better and should be replaced by the global best.

[0083] Step 5: Start iteration and use the whale optimization algorithm to find the optimal solution for the parameters.

[0084] Step 6: Repeat Step 4-5 until the maximum number of iterations is reached.

[0085] Step 7: Decode the final optimization result and output the parameters of initial learning rate, number of training rounds, learning reduction factor and learning rate reduction cycle. The optimal result is achieved by continuously optimizing the parameters initialized in Step 2.

[0086] Step 3: Build a VGG16 network model, input the training set and test set obtained in step 1 for training and testing, and extract the frequency domain features of the echo signal.

[0087] like Figure 2 As shown, a VGG16 network model is built. The overall structure of the VGG16 network consists of five convolutional layers, five pooling layers and three fully connected layers. All convolutional structures in the network model use 3×3 convolution kernels as feature extractors to detect features from more angles. The size of the network input image is 224×224×3. After convolution by the convolutional layer, the image size is reduced to half of the original size, but the number of convolution channels increases exponentially from 64. After five layers of convolution, the image size input to the fully connected layer becomes a 7×7×512 image. The image is expanded using the fully connected layer, and finally 1000 classification detection results are output. Finally, the Softmax classifier is used to output the corresponding classification detection probability. The main features of the VGG16 network model constructed in the present invention are as follows:

[0088] (1) A 3×3 small convolution kernel is used to replace the 5×5 and 7×7 large convolution kernels commonly used in previous CNNs, and the step size of the convolution operation is reduced to reduce the convolution operation parameters and ensure the change of image size;

[0089] (2) The depth of the entire network is deepened by stacking small convolution kernels. Small convolution kernels ensure that the deepening of the network structure will not cause an increase in parameters, and the training effect will not degrade or the gradient will disappear.

[0090] Step 4, integrate the time domain features output from step 2 and the frequency domain feature data output from step 3. The data integration stage is to make each frame feature data of the two networks correspond one-to-one with the experimental real height data. The specific process is that after extracting features, all data sequences and image sequences are first converted into cell arrays in Matlab through the mat2cell function; then the experimental real height data is converted into a categorical array through the categorical function, and then the categorical array is converted into a cell array through the mat2cell function; finally, the three cell arrays containing the data sequence, image sequence and categorical array are put into a cell array, so that each frame of data sequence and image can correspond one-to-one with the real data, and the trained feature data can also correspond one-to-one with the real data, thus completing the integration of the series data and image data that were separately saved in the early stage.

[0091] Step 5: Train and test the integrated data in a hybrid model of the WOA network model and the VGG16 network model to obtain actual height information and realize high-precision height measurement of the unmanned aerial vehicle height measurement radar.

[0092] Reference Figure 4, is the result of training and testing the echo signal data using the method of the present invention. It can be seen from the figure that the network test result can reach 89.31%, which is significantly improved compared with the accuracy of simply using the WOA-CNN-LSTM-Attention or VGG16 network. This shows that processing the data in both the time domain and the frequency domain can improve the extraction of feature information, thereby improving the accuracy of classification.

[0093] The altitude information processed by the improved VGG16 network is converted into actual altitude, and compared with the actual altitudes processed by other different networks and the actual altitudes processed by the altimeter altitude search algorithm based on the CFAR algorithm at different altitudes in different flight stages of the UAV. Figure 5 and Figure 6 The height measurement results and height error results comparison diagrams of five different measurement methods in a flat environment are shown in the figure. It can be seen from the figure that in a flat environment, the height curve obtained by the improved VGG16 is most consistent with the actual height RTK curve, and the height error is the smallest. Compared with the traditional height measurement algorithm, this algorithm also has better accuracy and stability, indicating that this method has high height measurement accuracy in a flat environment and can perform high-precision height measurement in this scenario. However, the height obtained by GoogleNet has large errors at individual heights, low accuracy, and poor stability; the height obtained by AlexNet is relatively stable, but has poor accuracy.

[0094] Figure 7 and Figure 8 The height measurement results and height error comparison diagrams of five different measurement methods in a grassland environment are shown in the figure. It can be seen from the figure that in a grassland environment, the height curve obtained by processing based on the improved VGG16 altimeter radar height search algorithm is most consistent with the actual height RTK curve, with the highest height measurement accuracy and the smallest height error. Compared with the traditional altimeter algorithm, the algorithm also has better accuracy and stability, indicating that the method also has good height measurement accuracy in a grassland environment, and the method of the present invention can be used to achieve high-precision height measurement in a grassland scene. However, the heights obtained by GoogleNet and AlexNet have large errors at individual heights, low accuracy, and poor stability; the height measurement accuracy obtained by processing based on the VGG16 altimeter radar height search algorithm is moderate, and it is impossible to obtain higher-precision height data like the improved VGG16 altimeter radar height search algorithm. This is because the height information obtained by extracting feature information only at the frequency domain level is relatively small, and the height search method still has a large room for improvement.

[0095] From the above analysis, it can be concluded that in both flat and grassland environments, the height search method based on the improved VGG16 altimeter radar can achieve high-precision height measurement. Compared with the traditional method, the height obtained by the improved VGG16 altimeter radar height search algorithm is more accurate and has the best accuracy and stability. This is because the improved VGG16 altimeter radar height search algorithm can extract the time domain characteristics and frequency domain characteristics of the data at the same time, so it can obtain more feature information relative to other networks and algorithms, which can greatly improve the accuracy of network training and testing, so that it can obtain more accurate height information. The height data obtained by the improved VGG16 altimeter radar height search algorithm can reach 5‰ in accuracy, achieving high-precision height measurement, indicating that the use of neural networks for height search of altimeter radars has a good application space.

[0096] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. Based on the improved VGG16 altimeter radar height search method, It is characterized in that The following steps are involved: Step 1, collecting echo signal data of the experimental ground through a height measuring radar, and constructing a training set and a test set of a data sequence type, and a training set and a test set of an image sequence type according to the echo signal data; Step 2: construct a WOA-CNN-LSTM-Attention network model, perform training and prediction using the training set and test set of the data sequence type obtained in step 1, and obtain the time domain feature data of the echo signal; Step 3, construct a VGG16 network model, perform training and prediction using the training set and test set of the image sequence type obtained in step 1, and obtain frequency domain feature data of the echo signal; Step 4, integrating the time domain feature data obtained in step 2 and the frequency domain feature data obtained in step 3, so that each frame of feature data output by the two network models corresponds one-to-one with the actual height data of the experiment, and obtaining integrated data; Step 5: Input the integrated data into the hybrid model of the WOA-CNN-LSTM-Attention network model and the VGG16 network model for training and prediction to obtain the actual height information.

2. According to claim 1, the height search method based on the improved VGG16 altimeter radar, It is characterized in that The specific process of constructing a training set and a test set of data sequence types according to the echo signal data in step 1 is as follows: First, the collected echo signal data is re-inserted; Then, the re-inserted data is imported into Excel, and the dimension corresponding to the collected echo signal data is (100000, 6); Again, the collected echo signal data is reconstructed in Excel, and a data set with 100,000 data items is obtained after reconstruction. The input data dimension of the data set is (100,000, 11), and the output data dimension is (100,000, 1); Finally, the reconstructed data set is divided into training set and test set.

3. According to claim 2, the height search method based on the improved VGG16 altimeter radar, It is characterized in that The specific process of constructing the training set and the test set of the image sequence type according to the echo signal data in step 1 is as follows: First, the collected echo signal data is re-inserted; Then, the re-interpolated data is transformed by CZT, and the corresponding spectrum is plotted to obtain a CZT spectrum image; Thirdly, the CZT spectrum image is divided into ten types of data samples according to the front position of the spectrum, with a total of 100,000 frames of data; each frame of data represents a frame of CZT spectrum image, and 100,000 frames of CZT spectrum images are obtained; Finally, the obtained CZT spectrum images are divided into training sets and test sets, and batch label annotations are performed on the training sets and test sets so that the data, spectrum images and classification categories corresponding to each frame correspond one to one.

4. According to claim 3, the height search method based on the improved VGG16 altimeter radar, It is characterized in that The training set and the test set of the image sequence type are both mixed data of ten types of data samples.

5. According to claim 1, the height search method based on the improved VGG16 altimeter radar, It is characterized in that The specific steps of step 2 are: Step 2.1, constructing a WOA-CNN-LSTM-Attention network model, wherein the WOA-CNN-LSTM-Attention network model includes a WOA part, a CNN-LSTM-Attention part, and a data part; Step 2.2: Input the training set of the data sequence type obtained in step 1 to the WOA-CNN-LSTM-Attention network model. The WOA part transmits the parameter vector of the training set to the CNN-LSTM-Attention part. The CNN-LSTM-Attention part decodes the parameter vector to obtain the corresponding specific parameter information. The data part trains the network with the training set data. Step 2.3, predict the trained WOA-CNN-LSTM-Attention network model using the test set data; Step 2.4, after prediction, calculate the mean square error between the predicted output value and the actual output value, and pass the mean square error to the WOA part, iteratively calculate the fitness value, and take the minimum fitness value obtained at the end of the iteration as the optimal result to obtain the corresponding optimal network hyperparameters; In step 2.5, the WOA-CNN-LSTM-Attention network model is tested with the optimal network hyperparameters and test set data to obtain the time domain feature data of the echo signal.

6. According to claim 1, the height search method based on the improved VGG16 altimeter radar, It is characterized in that The step 4 comprises the following steps: Step 4.1, convert the time domain feature data and frequency domain feature data obtained in step 2 and step 3 into cell arrays respectively through the mat2cell function in Matlab; Step 4.2, first convert the experimental ground truth height data into a categorical array using the categorical function, and then convert the categorical array into a cell array using the mat2cell function; In step 4.3, the three cell arrays containing the data sequence, image sequence, and classification array are put into one cell array, so that the sequence and image of each frame of data can correspond one-to-one with the experimental ground truth height data.