A satellite transponder identification model and its training method and use method

By building a satellite forwarder recognition model based on artificial neural networks, using data preprocessing and cleaning technology, the problem of low efficiency and experience dependence on traditional recognition methods is solved, and efficient and automatic satellite forwarder recognition is achieved, reducing the Missing detection rates and provide all-weather recognition capabilities.

CN113269135BActive Publication Date: 2024-12-03CHINA ACADEMY OF SPACE TECHNOLOGY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202110670593.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-06-17
Publication Date
2024-12-03
Estimated Expiration
2041-06-17

AI Technical Summary

Technical Problem

The traditional satellite forwarder recognition method relies on public configuration and historical monitoring spectrum, the manual recognition efficiency is low, the accuracy depends on experience, and the signal spectrum lacking prior knowledge cannot be identified, and there is a high possibility of missed detection, especially in many cases. Satellites are particularly serious.

Method used

The satellite forwarder identification model is constructed using artificial neural networks. By constructing a training data set, including data classification and preprocessing steps, the time domain waveform curve is extracted, and a three-layer fully connected neural network is constructed using the PyTorch framework to perform data cleaning and spectrum waveform calibration. Convert to image classification problem and automatically identify the satellite transponder to which the signal belongs.

Benefits of technology

It improves the efficiency and accuracy of satellite forwarder identification, reduces missed detection rate, provides automatic identification auxiliary information, alleviates the pressure on monitoring personnel, and achieves uninterrupted identification capabilities 24 hours a day.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113269135B_ABST
    Figure CN113269135B_ABST
Patent Text Reader

Abstract

The present application provides a training method for a satellite transponder recognition model, comprising: constructing a satellite transponder recognition model using an artificial neural network; constructing a training data set; and training the satellite transponder recognition model using the constructed training data set. The step of constructing the training data set includes a data classification step and a data preprocessing step. The data classification step includes: for a plurality of signal spectra acquired from a receiver, establishing a mapping relationship between each signal spectrogram and a corresponding satellite transponder name label, wherein the signal spectrogram includes a background and a time domain waveform curve. The data preprocessing step includes: extracting a time domain waveform curve from the signal spectrogram, and processing the two-dimensional data of the extracted time domain waveform curve.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a human-in-the-loop satellite transponder automatic identification and classification auxiliary technology, and in particular to a satellite transponder identification model and a training method thereof suitable for application scenarios with multiple satellite transponders, a large amount of received signal spectrum data, and low manual classification and identification efficiency. Background Art

[0002] At present, most traditional satellite transponder identification methods are based on existing expert information such as public transponder configuration (satellite beacon, transponder starting frequency band, polarization mode, satellite orbit position), historical monitoring spectra in previous years, etc., and complete the identification operation by manual comparison and matching features to identify the satellite transponder from which the signal received by the ground receiver comes. Therefore, the traditional satellite transponder identification method is heavily dependent on existing data such as public transponder configuration and historical monitoring spectra in previous years, and cannot identify signal spectra that lack prior knowledge. In addition, the recognition accuracy depends on the experience and level of the staff, which is highly subjective, lacks timeliness, and has a high possibility of missed detection. At the same time, it consumes a lot of manpower and material resources. This problem is particularly serious in the case of multiple satellites. Since multiple satellites may change orbits and co-orbit during operation, in this scenario, it is even impossible to manually identify which satellite the signal received by the ground receiver comes from based on prior knowledge such as satellite transponder configuration and existing satellite reference signals. Summary of the invention

[0003] At least one purpose of the present application is to provide a satellite transponder identification model and a training method and a use method thereof, through which the satellite transponder identification model and the use method provided by the present application can automatically identify the satellite transponder from which the received signal comes. The identification operation can be performed 24 hours a day, 7 days a week, so as to provide auxiliary identification information for monitoring personnel, help monitoring personnel quickly find and identify targets, and reduce the workload of monitoring personnel in identifying and classifying satellite transponders from massive signal spectra.

[0004] The present application provides a method for training a satellite transponder recognition model, comprising:

[0005] S01: Building a satellite transponder recognition model using artificial neural networks;

[0006] S02: Build training data set;

[0007] S03: Use the constructed training data set to train the satellite transponder recognition model.

[0008] Wherein, the step S02 includes a data classification step and a data preprocessing step.

[0009] The data classification step includes: for a plurality of signal spectrograms acquired from a receiver, establishing a mapping relationship between each signal spectrogram and a corresponding satellite transponder name tag, wherein the signal spectrogram includes a background and a time domain waveform curve;

[0010] The data preprocessing step includes: extracting a time domain waveform curve from the signal spectrum, and processing the two-dimensional data (x 1 ,y 1 ),(x 2 ,y 2 ),…(x n ,y n ))Perform the following processing:

[0011] x i =(x i -x min ) / (x max -x min )

[0012] y i =(y i -y min ) / (y max -y min );

[0013] Among them, x i is the x-axis coordinate of the i-th point of the time domain waveform curve, x min is the minimum x-axis coordinate of the time domain waveform curve, x max is the maximum x-axis coordinate of the time domain waveform curve,

[0014] Among them, y i is the y-axis coordinate of the i-th point of the time domain waveform curve, y min is the minimum y-axis coordinate of the time domain waveform curve, y max is the maximum y-axis coordinate of the time domain waveform curve,

[0015] The two-dimensional data of each signal spectrogram processed by the data preprocessing step and the satellite transponder name label corresponding to the signal spectrogram constitute a training data, and a training data set is constructed using multiple training data.

[0016] In at least one embodiment of the present application, step S02 further includes a data cleaning step, in which the original data received from the receiver is cleaned according to prior information to remove data containing erroneous information.

[0017] In at least one embodiment according to the present application, in the data preprocessing step, the background and the time domain waveform curve are distinguished by color to extract the time domain waveform curve from the signal spectrogram.

[0018] In at least one embodiment of the present application, in step S01, a satellite transponder recognition model is constructed based on a PyTorch framework and an artificial neural network is used to construct the model.

[0019] In at least one embodiment of the present application, the satellite transponder identification model is composed of three layers of fully connected neural networks, the activation function between each layer of the fully connected neural network is leaky_relu, the last layer is a softmax layer, and the loss function is MSELoss.

[0020] In at least one embodiment according to the present application, the signal spectrum diagram is a time domain waveform diagram.

[0021] In at least one embodiment according to the present application, the input of the satellite transponder identification model is a training data set, and the output of the satellite transponder identification model is a name tag of the satellite transponder to which the signal belongs.

[0022] The present application also provides a satellite transponder identification model, which is trained by the above method.

[0023] The present application also provides a method for using a satellite transponder identification model, which uses the above-mentioned satellite transponder identification model for identification.

[0024] In at least one embodiment according to the present application, the above-mentioned method of use includes:

[0025] S11: Acquire a signal from a receiver and draw a signal spectrum according to the original data of the signal;

[0026] S12: Extract the time domain waveform curve in the signal spectrum and save it as two-dimensional data ((x 1 ,y 1 ),(x 2 ,y 2 ),…(x n ,y n ));

[0027] S13: extracting the two-dimensional data of the time domain waveform curve ((x 1 ,y 1 ),(x 2 ,y 2 ),…(x n ,y n ))Perform the following processing:

[0028] x i =(x i -x min ) / (x max -xmin )

[0029] y i =(y i -y min ) / (y max -y min )

[0030] Among them, x i is the x-axis coordinate of the i-th point of the time domain waveform curve, x min is the minimum x-axis coordinate of the time domain waveform curve, x max It is the maximum x-axis coordinate of the time domain waveform curve.

[0031] Among them, y i is the y-axis coordinate of the i-th point of the time domain waveform curve, y min is the minimum y-axis coordinate of the time domain waveform curve, y max is the maximum y-axis coordinate of the time domain waveform curve;

[0032] S14: Input the two-dimensional data obtained in step S13 into the satellite transponder identification model, and output the name label of the satellite transponder to which the signal spectrum belongs.

[0033] The satellite transponder recognition model provided by the present application can automatically identify the satellite transponder from which the received signal comes. In the technical solution provided by the present application, the satellite transponder recognition problem is converted into an image classification problem, thereby making the complex multi-parameter problem intuitive and simple.

[0034] According to the training method of the satellite transponder recognition model provided in the present application, a high-quality training data set is obtained by adopting a specific data preprocessing method, so that a high-quality model can be trained. The model needs to process a small amount of data and has a fast model convergence speed, which can improve the recognition accuracy of the satellite transponder recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The following will further explain the above characteristics, technical features, advantages and implementation methods of the present application in a clear and understandable manner through the description of the preferred embodiments and in combination with the accompanying drawings. The following drawings are only intended to illustrate and explain the present application, and do not limit the scope of the present application. Among them:

[0036] Figure 1 A satellite transponder identification model training method according to an embodiment of the present application is shown;

[0037] Figure 2 A method for using a satellite transponder identification model according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0038] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described with reference to the accompanying drawings.

[0039] In order to identify which satellite transponder the received signal comes from, the existing technology mainly uses manual comparison and matching of features to identify the signal spectrum. This identification method is heavily dependent on the experience and level of the monitoring personnel, and has low identification efficiency and accuracy.

[0040] In order to improve the recognition efficiency and accuracy of the satellite transponder to which the signal belongs, the present application provides a satellite transponder recognition model based on a neural network and its training method and use method. In the technical solution of the present application, based on the time domain waveform of the received signal, an artificial neural network is used to construct (for example, based on the PyTorch framework) a satellite transponder recognition model, and the satellite transponder recognition model is trained using a training data set constructed from a satellite transponder data set. Using the trained satellite transponder recognition model, the time domain waveform of the received signal can be classified according to which satellite it belongs to, thereby identifying which satellite the signal comes from.

[0041] It can be seen that in the technical solution provided by the present application, the satellite transponder identification problem is converted into an image classification problem, thereby making the complex multi-parameter problem intuitive and simple. In addition to significantly improving the efficiency of satellite transponder identification and reducing the missed detection rate, the technical solution provided by the present application can also provide soft-decision pre-classification auxiliary information for the manual and precise identification of the satellite transponder to which the ground receiving signal belongs. The satellite transponder automatic identification auxiliary technology can relieve the pressure of observers, reduce the misjudgment caused by observers' fatigue due to long-term observation, and can provide auxiliary effects for subsequent manual and precise identification and judgment.

[0042] In theory, the signal emitted by each device (for example, each different satellite transponder) has fine features that are uniquely bound to the manufacturing defects of the device and the characteristics of each component of the device. By extracting and analyzing the features of the received signal, the transmitting device corresponding to the signal can be inferred, thereby obtaining the mapping relationship between the signal and the signal transmitting source device. The prior art mainly uses manual comparison and matching features to identify signal spectra. The deep artificial neural network used in this application does not require manual design and feature extraction. The neural network can extract high-dimensional abstract features that are sufficient to describe a specific feature pattern based on the original input, thereby identifying the satellite transponder of which satellite the received signal comes from.

[0043] A satellite transponder recognition model training method according to an embodiment of the present invention comprises the following steps:

[0044] S01: Building a satellite transponder recognition model using artificial neural networks

[0045] A satellite transponder recognition model based on artificial neural network is constructed based on PyTorch framework. The backbone of the satellite transponder recognition model consists of three layers of fully connected neural network. The activation function between each layer of the network is leaky_relu, and the last layer is a softmax layer. The loss function is MSELoss. The input of the satellite transponder recognition model is the training data set, and the output of the satellite transponder recognition model is the name label of the satellite transponder to which the signal belongs.

[0046] S02: Build training dataset

[0047] This step S02 includes data cleaning, data classification and data preprocessing.

[0048] Deep learning is a data-driven representation learning method. The performance of the neural network model depends on the quality of the training data set, so it is crucial to build a high-quality training data set. A high-quality training data set can be obtained through data cleaning, data classification, and data preprocessing steps.

[0049] Among them, the data cleaning step includes: cleaning the original data received from the receiver (i.e., multiple signal spectra, with the horizontal axis being frequency and the vertical axis being amplitude) according to prior information (existing satellite transponder configuration, reference signals of existing satellites, etc.), cleaning out data containing erroneous information, and obtaining clean data.

[0050] The data classification step includes: effectively classifying the clean data according to the prior information, and the classification is based on the typical signal spectrum pattern corresponding to each satellite transponder. After the data classification step, the mapping relationship between each signal spectrum and its corresponding satellite transponder is established.

[0051] The steps of data preprocessing include: 1) removing redundant information in the signal spectrum that is useless for improving the model performance; 2) calibrating the spectrum waveform. These two steps will be described in detail below with reference to specific embodiments.

[0052] 1) Remove redundant information in the signal spectrum that is useless for improving model performance

[0053] As described above, in the technical solution provided by the present application, the satellite transponder identification problem is converted into an image classification problem. In the prior art, the processing methods for image classification problems or image recognition problems all analyze and calculate the entire area of ​​the image, which results in large consumption of computing resources and long processing time.

[0054] However, for the satellite transponder identification problem to be solved by the present application, the important information in the signal spectrogram (i.e., the time domain waveform diagram) is not distributed over the entire area of ​​the image. The signal spectrogram of the original data received from the receiver is an image that includes a background (e.g., white) and a time domain waveform curve (e.g., red) that is different from the background color. The background is redundant information that is useless for improving the performance of the satellite transponder identification model. If the background data is introduced into the training data in the training data set, it will interfere with and affect the training of the model. In fact, the important information in the signal spectrogram is sparse, and only the time domain waveform curve therein is useful information.

[0055] In order to improve the quality of training data, in the technical solution provided by the present application, in the data preprocessing step, the signal spectrogram is subjected to time domain waveform curve extraction. The background of the signal spectrogram is removed, and only the time domain waveform curve in the signal spectrogram is extracted as training data. Specifically, the useless background and the useful time domain waveform curve can be distinguished according to the color. Then, the extracted time domain waveform curve is saved as two-dimensional data ((x 1 ,y 1 ),(x 2 ,y 2 ),…(x n ,y n )).

[0056] By taking such steps to remove redundant information, the following technical effects can be achieved: 1. The quality of training data can be improved. High-quality training data can train high-quality models. Compared with background information that is useless for identification and classification, the time domain waveform curve (overall trend, number of peaks, relative size of peaks, relative position relationship of peaks, etc.) data is the target to be analyzed and processed. The neural network only learns and models the data of the time domain waveform curve, and it is easier to learn the essential features that are only strongly correlated with the time domain waveform curve itself; 2. After the extraction of the time domain waveform curve data, the amount of data that the model needs to process can be reduced exponentially, and only the high-value data waveform of concern is learned, and the model converges quickly.

[0057] 2) Spectrum waveform calibration

[0058] The jamming of the signal transmission and reception data link network will cause data transmission anomalies, and the waveform distortion caused by human factors will introduce noise, the relative position of the waveform peak and the deviation of the waveform envelope. However, conventional data processing methods do not take these factors into consideration. As a result, for the same category label, different data sample instances may be very different, while data sample instances belonging to different category labels may be relatively similar. This will prevent the neural network classifier training from converging, leading to the failure of the classification task.

[0059] In order to solve this problem, the technical solution of the present application adopts spectrum waveform calibration to process the two-dimensional data of the extracted time domain waveform curve to eliminate the influence of external factors, thereby improving the recognition accuracy of the satellite transponder recognition model.

[0060] Specifically, the extracted two-dimensional data of the time domain waveform curve is processed as follows:

[0061] x i =(x i -x min ) / (x max -x min )

[0062] y i =(y i -y min ) / (y max -y min )

[0063] Among them, x i is the x-axis coordinate of the i-th point of the time domain waveform curve, x min is the minimum x-axis coordinate of the time domain waveform curve, x max It is the maximum x-axis coordinate of the time domain waveform curve.

[0064] Among them, y i is the y-axis coordinate of the i-th point of the time domain waveform curve, y min is the minimum y-axis coordinate of the time domain waveform curve, y max It is the maximum y-axis coordinate of the time domain waveform curve.

[0065] The coordinates of the processed data are restricted to (0, 1), restoring the overall trend of the waveform data and the relationship between relative positions and relative peak sizes. Through this preprocessing method, different sample instances under the same label will have similar waveforms.

[0066] The label information of the data can be encoded using one-hot encoding. This can solve the problem that the classifier is not good at processing attribute data, and can also expand the features to a certain extent.

[0067] S03: Use the training dataset to train the satellite transponder recognition model

[0068] After step S02, a training data set consisting of high-quality training data can be obtained. The satellite transponder recognition model obtained in step S01 is trained using the training data set, thereby obtaining a neural network satellite transponder recognition model with excellent performance.

[0069] According to another embodiment of the present application, a satellite transponder recognition model is provided, and the satellite transponder recognition model is trained by the satellite transponder recognition model training method provided by the present application.

[0070] According to another embodiment of the present application, a method for using a satellite transponder identification model is also provided. The method includes:

[0071] S11: Acquire a signal from a receiver and draw a signal spectrum according to the original data of the signal;

[0072] S12: Extract the time domain waveform curve in the signal spectrum and save it as two-dimensional data ((x 1 ,y 1 ),(x 2 ,y 2 ),…(x n ,y n ));

[0073] S13: Perform spectrum waveform calibration, and perform the following processing on the extracted two-dimensional data of the time domain waveform curve:

[0074] x i =(x i -x min ) / (x max -x min )

[0075] y i =(y i -y min ) / (y max -y min )

[0076] Among them, x i is the x-axis coordinate of the i-th point of the time domain waveform curve, x min is the minimum x-axis coordinate of the time domain waveform curve, x max It is the maximum x-axis coordinate of the time domain waveform curve.

[0077] Among them, y i is the y-axis coordinate of the i-th point of the time domain waveform curve, y min is the minimum y-axis coordinate of the time domain waveform curve, y max is the maximum y-axis coordinate of the time domain waveform curve;

[0078] S14: Input the two-dimensional data after spectrum waveform calibration into the satellite transponder identification model, and output the name tag of the satellite transponder to which the signal spectrum belongs, so as to identify the satellite transponder of which satellite the signal spectrum comes from.

[0079] In order to test the recognition efficiency and accuracy of the satellite transponder recognition model provided by the present application, the inventor tested the satellite transponder recognition model obtained by the training method of the present application. According to another embodiment of the present application, a satellite transponder recognition model is trained according to the above method including steps S01-S03 on a training data set containing 13 satellite transponder data. After testing, the classification accuracy rate reached 96%, and the prediction classification accuracy rate was high, which can provide high-quality auxiliary information for subsequent manual precise judgment and review. It can be seen that in the satellite transponder automatic identification task, the trained model can process the signal spectrogram obtained from the ground receiver in real time and give a predicted classification result. The automatic identification auxiliary technology has the ability to work continuously 24 hours a day, all day long.

[0080] In summary, the present application provides a satellite transponder recognition model based on a neural network and a training method and a use method thereof. Through the satellite transponder recognition model, it is possible to automatically identify the satellite transponder from which the received signal comes. In the technical solution provided by the present application, the satellite transponder recognition problem is converted into an image classification problem, thereby making the complex multi-parameter problem intuitive and simple.

[0081] According to the training method of the satellite transponder recognition model provided in the present application, a high-quality training data set is obtained by adopting a specific data preprocessing method, so that a high-quality model can be trained. The model needs to process a small amount of data and has a fast model convergence speed, which can improve the recognition accuracy of the satellite transponder recognition model.

[0082] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0083] The above description is only an illustrative embodiment of the present application and is not intended to limit the scope of the present application. Any equivalent changes, modifications and combinations made by any technician in the field without departing from the concept and principle of the present application shall fall within the scope of protection of the present application.

Claims

1. A method for training a satellite transponder recognition model, comprising: S01: Building a satellite transponder recognition model using artificial neural networks; S02: Build training data set; S03: Use the constructed training data set to train the satellite transponder recognition model. Wherein, the step S02 includes a data cleaning step, a data classification step, and a data preprocessing step. The data cleaning step includes: performing data cleaning on a plurality of signal spectrograms obtained from a receiver to clean out data containing erroneous information; The data classification step includes: establishing a mapping relationship between each signal spectrum and a corresponding satellite transponder name tag, wherein the signal spectrum includes background and time domain waveform curves; The data preprocessing step includes: extracting a time domain waveform curve from the signal spectrogram, processing the extracted two-dimensional data (x1, y1), (x2, y2), ... (x n ,y n )) Perform the following processing: ; in, is the x-axis coordinate of the i-th point of the time domain waveform curve, is the minimum x-axis coordinate of the time domain waveform curve, is the maximum x-axis coordinate of the time domain waveform curve, in, is the y-axis coordinate of the i-th point of the time domain waveform curve, is the minimum y-axis coordinate of the time domain waveform curve, is the maximum y-axis coordinate of the time domain waveform curve, The two-dimensional data of each signal spectrogram processed by the data preprocessing step and the satellite transponder name label corresponding to the signal spectrogram constitute a training data, and a training data set is constructed using multiple training data.

2. The method according to claim 1, wherein: In the data cleaning step, data cleaning is performed based on prior information.

3. The method according to claim 1, wherein: In the data preprocessing step, the background and the time domain waveform curve are distinguished by color to extract the time domain waveform curve from the signal spectrogram.

4. The method according to claim 1, wherein: In the step S01, a satellite transponder recognition model is constructed based on the PyTorch framework and using an artificial neural network.

5. The method according to claim 4, wherein: The satellite transponder recognition model consists of three layers of fully connected neural networks, the activation function between each layer of the fully connected neural network is leaky_relu, the last layer is a softmax layer, and the loss function is MSELoss.

6. The method according to claim 1, wherein: The signal spectrogram is a time domain waveform diagram.

7. The method according to claim 1, wherein: The input of the satellite transponder identification model is a training data set, and the output of the satellite transponder identification model is a name label of the satellite transponder to which the signal belongs.

8. A method for using a satellite transponder recognition model, using the satellite transponder recognition model trained by the method according to any one of claims 1 to 7 for recognition.

9. The method according to claim 8, comprising: S11: Acquire a signal from a receiver and draw a signal spectrum according to the original data of the signal; S12: Extract the time domain waveform curve in the signal spectrum and save it as two-dimensional data ((x1,y1),(x2,y2),…(x n ,y n )); S13: The extracted two-dimensional data of the time domain waveform curve ((x1, y1), (x2, y2), … (x n ,y n )) Perform the following processing: ; in, is the x-axis coordinate of the i-th point of the time domain waveform curve, is the minimum x-axis coordinate of the time domain waveform curve, is the maximum x-axis coordinate of the time domain waveform curve, in, is the y-axis coordinate of the i-th point of the time domain waveform curve, is the minimum y-axis coordinate of the time domain waveform curve, is the maximum y-axis coordinate of the time domain waveform curve; S14: Input the two-dimensional data obtained in step S13 into the satellite transponder identification model, and output the name label of the satellite transponder to which the signal spectrum belongs.

Citation Information

Patent Citations

  • Satellite identification method of ISAR image

    CN110705508A

  • Beidou satellite navigation interference source identification method based on BP neural network

    CN111562597A