Small sample positioning algorithm based on electromagnetic flying time-space inseparable characteristics

Abstract

The application provides a small sample positioning algorithm based on electromagnetic flying ring space-time inseparable characteristics, and is used for solving the technical problems that the existing technology needs more training amount to have higher accuracy and that too many classification numbers result in low sample fitting effect in the background art.The application discloses a small sample positioning algorithm based on electromagnetic flying ring space-time inseparable characteristics, and comprises the following steps: step one: selecting a line segment on a half space plane along a propagation axis as a boundary to measure, and obtaining space spectrum data on the line segment; step two: for the data detected by the electromagnetic flying ring, after normalization, the data is decomposed according to the detection sequence, r-axis and z-axis, so that the number of categories is reduced and the number of samples is increased; step three: the data is subjected to noise addition and expansion, the sample amount is increased, and the generalization ability is enhanced, training noise and test noise with different variances are set to carry out combination experiments, and relatively accurate noise matching results are obtained.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic flying ring positioning algorithm technology, and in particular to a small sample positioning algorithm based on the spatiotemporal inseparability of electromagnetic flying rings. Background Technology

[0002] This invention relates to a method for measuring the frequency domain characteristics of an electromagnetic flying ring, which possesses spatiotemporal inseparability and rotational symmetry. In electromagnetic flying ring localization, the localization algorithm faces challenges related to a limited number of samples and a large number of classifications. For neural networks, only sufficient training data can achieve high accuracy; simultaneously, an excessive number of classifications leads to a decrease in the sample fitting effect. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a small-sample localization algorithm based on the spatiotemporal inseparability of electromagnetic flying rings. This algorithm solves the technical problems mentioned in the background section, such as the need for a large amount of training to achieve high accuracy and the resulting low sample fitting effect due to an excessive number of classifications.

[0004] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0005] The small-sample localization algorithm based on the spatiotemporal inseparability of electromagnetic flying rings includes the following steps:

[0006] Step 1: Select a line segment on half of the spatial plane bounded by the propagation axis and measure it to obtain the spatial spectrum data of the line segment.

[0007] Step 2: After normalizing the data detected by the electromagnetic flying ring, decompose it along the r-axis and z-axis according to the detection order to reduce the number of categories and increase the number of samples.

[0008] Step 3: Add noise to the data to increase the sample size and enhance generalization ability. Set up training noise and test noise with different variances to conduct combined experiments to obtain more accurate noise matching results.

[0009] Step 4: For the real part, imaginary part, or amplitude and phase, fit the horizontal and vertical coordinates respectively, and use a combination of random forest and neural network as machine learning methods for training and testing;

[0010] Step 5: Based on the evaluation results and actual application needs, select the machine learning method that performs best and meets the actual conditions for optimization, and adjust the model parameters according to the actual situation.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] 1. This patent provides a small-sample localization algorithm based on electromagnetic flying rings, which can effectively overcome the difficulties of small sample size and large number of classifications.

[0013] 2. By expanding the data and selecting machine learning methods, the accuracy and efficiency of the algorithm were improved.

[0014] 3. The algorithm provided in this patent can be widely used for electromagnetic wave positioning, as well as for situations requiring small sample positioning, such as drones, robots, and autonomous driving. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the electromagnetic flying ring positioning implementation design method in an embodiment of the present invention, wherein the horizontal axis is the z-axis and the vertical axis is the r-axis;

[0016] Figure 2 This is a schematic diagram of the electromagnetic flying ring detection sequence in an embodiment of the present invention. The detection space is gridded, with every 81 detection points having the same z coordinate and every 161 detection points having the same r coordinate.

[0017] Figure 3 This is a schematic diagram of spectrum data classification and labeling in an embodiment of the present invention (taking the real part as an example).

[0018] Figure 4 This is a schematic diagram of the data processing and training detection algorithm structure in an embodiment of the present invention;

[0019] Figure 5 This is a table showing the learning results of different algorithms and different noise combinations in the embodiments of the present invention;

[0020] Figure 6 This is an operation flowchart in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Example:

[0023] 1. Data Measurement

[0024] To measure its frequency domain characteristics, we select a line segment on half of the spatial plane along the propagation axis, with the propagation axis as the boundary. In this embodiment of the invention, the r-axis is divided into 81 equal parts, and the z-axis is divided into 161 equal parts, resulting in 81*161=13041 spatial points. By changing the r-coordinate and z-coordinate, different distances are selected on this line segment for measurement. By feeding the time-domain waveform of the electromagnetic flying ring into the antenna, the spatial spectrum distribution on this line segment can be obtained.

[0025] 2. Data Processing

[0026] After normalizing the 13,041 data points, they were decomposed into 81 and 161 classes according to the detection order of the electromagnetic flying rings and their r-coordinates (column C of the table) and z-coordinates (column B of the table), respectively. The x and y coordinates were fitted separately to reduce the number of classes, increase the number of samples per class, and improve data utilization. Data augmentation was performed by adding noise to each data point. However, too little noise could lead to overfitting, while too much could interfere with the neural network. Experiments using different combinations of training and test set variances were conducted to determine a suitable noise range.

[0027] If real and imaginary data are required for positioning in practical applications, the amplitude and phase should first be converted into real and imaginary parts, then the real and imaginary parts should be merged to obtain a total table of 13041*162, and then further processed.

[0028] 3. Selection of Machine Learning Methods

[0029] In terms of amplitude and phase angles, neural networks are selected as the primary machine learning method. However, for the real and imaginary parts, only one table is needed to analyze their x and y coordinates each time, making real and imaginary part angle analysis more efficient in utilizing the data. Therefore, decision trees, support vector machines, random forests, and neural networks are chosen as machine learning methods. The following is a comparison of the advantages and disadvantages of the four learning methods:

[0030] The advantage of Support Vector Machines (SVMs) is that they have kernel functions, making them suitable for handling multi-feature data. The disadvantages are that they are not suitable for large-scale datasets; and they are prone to overfitting with noisy data.

[0031] Decision trees have the advantage of handling both discrete and continuous data well, and they also handle missing values ​​effectively. Their disadvantages include sensitivity to outliers, making them less tolerant of experimental randomness, and they are not suitable for datasets with significant class imbalance.

[0032] Random forests can handle high-dimensional data and large-scale datasets, and they can reduce the risk of overfitting through randomization. Compared to trees, they handle class imbalance problems well. The downside is that they are prone to overfitting to noisy data.

[0033] Neural networks can adapt well to various types of data.

[0034] It has good scalability and is suitable for large-scale datasets. However, compared to other learning methods, it requires a large amount of training data because it is prone to overfitting on small datasets.

[0035] 4. Positioning Effectiveness Evaluation

[0036] The first factor is accuracy, which is the proportion of correctly identified location points out of the total number of identified points. The second is recall, which is the proportion of correctly identified location points out of the total number of target location points. The third is resolution, defined as the number of predicted points divided by the total number of points. In the error matrix, the range where more than 95% of the predicted points are distributed is denoted as 'a', and the distance between the two predicted points with the largest difference is denoted as 'b'. Taking the z-coordinate of the real part data as an example, the primary resolution is a / 81, and the minimum resolution is b / 81.

[0037] 5. Implementation Steps

[0038] Step 1: Select a line segment on half of the spatial plane along the propagation axis as the boundary and measure it to obtain spatial spectrum data of 81*161=13041 spatial points.

[0039] Step 2: After normalizing the data detected by the electromagnetic flying ring, decompose it into 81 categories and 161 categories according to the detection order along the r-axis and z-axis, thereby reducing the number of categories and increasing the number of samples;

[0040] Step 3: Add noise to the data to increase the sample size and enhance generalization ability. Set different training noise and test noise with variances of 0.01, 0.04, 0.09, 0.16 and 0.25 to conduct combined experiments to obtain more accurate noise matching results.

[0041] Step 4: For the real part, imaginary part, or amplitude and phase, fit the horizontal and vertical coordinates respectively, and use support vector machines, trees, random forests, and neural networks as machine learning methods for training and testing respectively;

[0042] Step 5: Final Evaluation and Recommendation of the Combined Random Forest and Neural Network Algorithm. Based on the actual application requirements, select the machine learning method that performs best and meets the actual conditions for optimization, and adjust the model parameters according to the actual situation to improve the accuracy and stability of the localization algorithm.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A small sample positioning algorithm based on electromagnetic flying time-space inseparable characteristics, characterized in that, Comprising the following steps: Step one: select a line segment on the plane along the propagation axis, and measure the spatial frequency spectrum data on the line segment; Step two: for the data detected by the electromagnetic fly ring, after normalization, according to the detection order, it is decomposed according to the r axis and the z axis, so as to reduce the category number and increase the sample number; Step three: add noise to the data to increase the sample size and enhance the generalization ability, set up training noise and test noise with different variances for combination experiment, and get noise matching result; Step four: for real part and imaginary part or amplitude and phase, respectively fit the horizontal and vertical coordinates, and use support vector machine, tree, random forest and neural network as machine learning method to train and test respectively; Step five: according to the evaluation result and the actual application demand, select the machine learning method with the best effect and meeting the actual condition for optimization, and adjust the model parameters according to the actual situation.

2. The small sample positioning algorithm based on electromagnetic flying ring space-time inseparable characteristics according to claim 1, characterized in that: Get 81*161=13041 spatial point spatial frequency spectrum data.

3. The small sample positioning algorithm based on electromagnetic flying ring space-time inseparable characteristics according to claim 1, characterized in that: Divide the r axis in step two into 81 parts and the z axis into 161 parts to get 81*161=13041 spatial points.

4. The small sample positioning algorithm based on electromagnetic flying ring space-time inseparable characteristics according to claim 1, characterized in that: The variance is 0.01, 0.04, 0.09, 0.16 and 0.25.

Images