Machine learning-oriented electromagnetic spectrum data labeling method
By generating simplified electromagnetic spectrum data through short-time Fourier transform and deep learning models, the problems of large data volume and laborious annotation in spectrum monitoring systems are solved, achieving efficient electromagnetic spectrum data storage and high-precision annotation, thus improving the efficiency and accuracy of reconnaissance equipment.
Patent Information
- Application Number
- CN202511523711.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-23
- Publication Date
- 2025-11-21
AI Technical Summary
In existing spectrum monitoring or communication reconnaissance systems, the collection, storage, and analysis of massive electromagnetic spectrum data face challenges such as huge data volume, high storage and processing costs, and time-consuming and labor-intensive manual annotation with difficulty in guaranteeing accuracy.
Using short-time Fourier transform and deep learning models, simplified electromagnetic spectrum data is generated through signal detection and parameter calculation. Electromagnetic spectrum IQ data is collected using electromagnetic signal acquisition equipment, short-time Fourier transform and energy detection are performed, signal parameters are recorded, a simplified electromagnetic spectrum data storage format is generated, and annotation is performed using a convolutional neural network.
It significantly reduced data storage requirements, improved annotation accuracy and deep learning model training efficiency, and enhanced detection accuracy and reconnaissance efficiency.
Smart Images
Figure CN120993045A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of electronic reconnaissance and electromagnetic spectrum monitoring, in particular to an electromagnetic spectrum data labeling method for machine learning. BACKGROUND
[0002] In the field of spectrum monitoring or communication reconnaissance, the collection, storage and analysis of massive electromagnetic spectrum data face many challenges. Existing spectrum monitoring or communication reconnaissance systems generate conventional data for fast Fourier transform frequency domain data and wide / narrow band time domain IQ data generated by direct wide / narrow band down conversion when collecting and analyzing wide electromagnetic spectrum signals. These data are huge, which brings difficulties to storage and deep learning model training. For example, the existing method shows exponential growth in data volume when processing high spectral resolution data, resulting in a sharp rise in storage and processing costs. In addition, existing labeling methods rely heavily on manual labeling, which is not only time-consuming and labor-intensive, but also difficult to ensure labeling accuracy, which seriously affects the training effect and generalization ability of deep learning models. Therefore, the present application provides an electromagnetic spectrum data labeling method for machine learning, which aims to reduce data storage, improve labeling accuracy and provide high-quality labeled data for deep learning models. SUMMARY
[0003] To solve the above technical problems, according to one aspect of the present application, the present application provides the following technical scheme:
[0004] An electromagnetic spectrum data labeling method for machine learning, comprising the following specific steps:
[0005] S1, data collection and short-time Fourier transform: using an electromagnetic signal collection device to collect electromagnetic spectrum IQ data, and performing short-time Fourier transform on each cycle according to time T;
[0006] S2, signal presence detection: performing energy detection on each short-time Fourier transform signal, and if the energy exceeds a set relative threshold or absolute threshold , it is determined that a signal is present; at the same time, the signal center frequency , signal bandwidth , signal energy , and occurrence time are recorded;
[0007] S3, continuous signal detection and recording: continuously performing short-time Fourier transform and recording signal samples according to S2; if the signal energy in M consecutive short-time Fourier transform cycles exceeds the set relative threshold or absolute threshold , it is considered that the signal persists, and the duration of the signal is recorded ;
[0008] S4, short-time signal detection and parameter calculation: detect whether the corresponding center frequency, signal bandwidth signal disappears, if not, add the time identifier to the signal ; at the same time, calculate the signal bandwidth fluctuation parameter, center frequency deviation fluctuation parameter and signal power fluctuation parameter;
[0009] S5, short-time signal sequence generation: after the above calculation, the short-time signal sequence arranged according to the appearance time and disappearance time can be obtained, and contains the signal bandwidth fluctuation parameter, center frequency fluctuation parameter and signal energy fluctuation parameter;
[0010] S6, time domain duty cycle calculation: statistical analysis of the time domain duty cycle of the short-time signal , as another dimension attribute of the short-time signal;
[0011] S7, data storage and labeling: store the short-time signal sequence containing the center frequency , signal bandwidth , signal bandwidth fluctuation parameter, center frequency deviation fluctuation parameter and signal power fluctuation parameter in the above steps as the basic data of the simplified electromagnetic spectrum data short-time signal dimension;
[0012] S8, deep learning model matching: use deep learning model to match the frequency deviation, bandwidth deviation and signal energy deviation in different samples, and match with a certain specific target device.
[0013] As a preferred scheme of the electromagnetic spectrum data labeling method for machine learning, the window function of the short-time Fourier transform in S1 is selected as the Hamming window, the window length is N sampling points, and the overlap rate is 50%, so as to ensure the frequency-time resolution of the signal.
[0014] As a preferred scheme of the electromagnetic spectrum data labeling method for machine learning, the energy of each short-time Fourier transform signal in S1 is detected, and the energy :
[0015]
[0016] wherein, is the frequency domain signal after short-time Fourier transform.
[0017] As a preferred scheme of the electromagnetic spectrum data labeling method for machine learning, the relative threshold in S2 is set to 3 times the standard deviation of the background noise energy, and the absolute threshold is set to 10 times the system noise level.
[0018] As a preferred scheme of the machine learning-oriented electromagnetic spectrum data labeling method, in the S4, the calculation process of the signal bandwidth fluctuation parameter is as follows: the signal bandwidth deviation calculated by different short-time Fourier transforms is statistically calculated to obtain the average bandwidth , the maximum bandwidth , and the minimum bandwidth , and the average bandwidth, the maximum bandwidth, and the minimum bandwidth are recorded as the signal bandwidth fluctuation parameter.
[0019] As a preferred scheme of the machine learning-oriented electromagnetic spectrum data labeling method, in the S4, the calculation process of the center frequency deviation fluctuation parameter is as follows: the center frequency deviation calculated by different short-time Fourier transforms is statistically calculated to obtain the average center frequency , the maximum frequency , and the minimum frequency ; if the maximum frequency is less than 10% of the average signal bandwidth, it is determined as the same signal, and the average center frequency, the maximum frequency, and the minimum frequency are recorded as the center frequency fluctuation parameter; otherwise, it is processed as different signals.
[0020] As a preferred scheme of the machine learning-oriented electromagnetic spectrum data labeling method, in the S4, the calculation process of the signal power fluctuation parameter is as follows: the short-time signal detected by the short-time Fourier transform is averaged to calculate the average signal power ; if the average signal power does not exceed a specific value , it is determined as the same signal, and the signal power deviation is recorded as the signal power fluctuation parameter; otherwise, it is processed as different signals.
[0021] As a preferred scheme of the machine learning-oriented electromagnetic spectrum data labeling method, in the S6, the time domain duty cycle calculation formula is as follows:
[0022]
[0023] wherein, is the signal duration, is the signal monitoring period.
[0024] As a preferred scheme of the machine learning-oriented electromagnetic spectrum data labeling method, in the S7, the storage format is as follows:
[0025] { , , , BW, E, , , , , }。
[0026] As a preferred scheme of the machine learning-oriented electromagnetic spectrum data labeling method, in the S8, the deep learning model adopts a convolutional neural network (CNN), the input is a feature vector of a short-time signal sequence, and the output is a category label of the target device.
[0027] Compared with the prior art, the present application has the following advantages:
[0028] 1. Data storage capacity and computing efficiency are improved
[0029] The conventional deep learning model based on electromagnetic spectrum data uses electromagnetic spectrum baseband IQ data as training data, and the training data capacity is huge, and the operation amount is also very large. The deep learning model of the present application is based on the basic parameters of the short-time signal detection result, the frequency domain center frequency fluctuation parameter, the signal bandwidth fluctuation parameter, and the time domain duty cycle, which can significantly reduce the electromagnetic spectrum data capacity, reduce the data storage pressure; at the same time, the training time of the deep learning model is greatly shortened, and the model training efficiency is improved.
[0030] 2. Labeling accuracy and detection accuracy are improved
[0031] The short-time signal detection result with high confidence obtained after the above detection can ignore more useless details of the baseband IQ data for the deep learning model, avoid overfitting of the IQ data stream for the deep learning model, and significantly improve the detection efficiency and accuracy. According to the simulation results, the signal labeling accuracy and the detection accuracy of the deep learning model are significantly improved.
[0032] 3. Application scenario advantage
[0033] In the field of communication reconnaissance, the present application can effectively reduce the data storage and processing burden of the reconnaissance equipment, and improve the reconnaissance efficiency and accuracy. In the field of spectrum monitoring, the present application can quickly and accurately identify and classify electromagnetic signals, and provide strong support for spectrum resource management. DETAILED DESCRIPTION
[0034] In order to make the purpose, technical scheme and advantages of the present application more clear, the embodiments of the present application will be further described in detail below.
[0035] The present application provides a machine learning-oriented electromagnetic spectrum data labeling method, which includes the following specific steps:
[0036] S1, data acquisition and short-time Fourier transform: using an electromagnetic signal acquisition device (such as a software-defined radio (SDR) or a spectrum analyzer) to acquire electromagnetic spectrum IQ data, and performing short-time Fourier transform every period according to time T;
[0037] wherein:
[0038] The window function of short-time Fourier transform is selected as Hamming window (suppressing spectrum leakage), the window length is N sampling points (specifically set according to the signal bandwidth (such as monitoring narrowband signals in the frequency band of 1-3 GHz, the sampling rate is 10 MHz, N is 1024 sampling points, corresponding to the window length of 1024 / 10 MHz=102.4 μs)), and the overlap rate is 50% to ensure the frequency-time resolution of the signal;
[0039] Energy detection is performed on each short-time Fourier transform signal, and the energy :
[0040]
[0041] wherein, is the frequency domain signal after short-time Fourier transform.
[0042] S2, signal appearance detection: energy detection is performed on each short-time Fourier transform signal, and if it exceeds the set relative threshold or absolute threshold , it is determined that a signal appears; at the same time, the signal center frequency , signal bandwidth , signal energy , and appearance time are recorded;
[0043] wherein:
[0044] The relative threshold is typically set as 3 times the standard deviation of the background noise energy (such as the average background noise power of-95 dBm, and the standard deviation of 5 dBm, then the relative threshold is-80 dBm), and the absolute threshold is typically set as 10 times the system noise level (such as the average background noise power of-100 dBm, then the absolute threshold is-90 dBm).
[0045] S3, continuous signal detection and recording: short-time Fourier transform is continuously performed, and signal sample recording is performed according to S2; if the signal energy exceeds the set relative threshold or absolute threshold in M (M is set according to the signal stability) consecutive short-time Fourier transform periods, it is considered that the signal continuously exists, and the duration (duration=M×T) is recorded.
[0046] S4, short-time signal detection and parameter calculation: whether the corresponding center frequency and signal bandwidth signal disappear is detected, and if not, the time identifier of the signal is added ; and the signal bandwidth fluctuation parameter, the center frequency deviation fluctuation parameter and the signal power fluctuation parameter are calculated simultaneously;
[0047] wherein:
[0048] The calculation process of the signal bandwidth fluctuation parameter is as follows: the signal bandwidth deviation calculated by different short-time Fourier transforms is statistically calculated to obtain the average bandwidth , the maximum bandwidth , and the minimum bandwidth , which are recorded as the signal bandwidth fluctuation parameter;
[0049] The calculation process of the center frequency deviation fluctuation parameter is as follows: the center frequency deviation calculated by different short-time Fourier transforms is statistically calculated to obtain the average center frequency , the maximum frequency shift ; if the maximum frequency shift is less than 10% of the average signal bandwidth, it is determined to be the same signal, and the average center frequency , the maximum frequency shift are recorded as the fluctuation parameters; otherwise, it is treated as different signals;
[0050] The calculation process of the signal power fluctuation parameter is as follows: the power values within the signal bandwidth of the short-time signal detected by the short-time Fourier transform are averaged to calculate the average signal power ; if the average signal power does not exceed a specific value , it is determined to be the same signal, and the signal power deviation is recorded as the signal power fluctuation parameter; otherwise, it is treated as different signals.
[0051] S5, short-time signal sequence generation: after the above calculation, a short-time signal sequence arranged according to the occurrence time and disappearance time is obtained, and the signal bandwidth fluctuation parameter, the center frequency fluctuation parameter and the signal energy fluctuation parameter are included.
[0052] S6, time domain duty cycle calculation: the time domain duty cycle of the short-time signal is statistically analyzed as another dimensional attribute of the short-time signal;
[0053] wherein:
[0054] The time domain duty cycle calculation formula is:
[0055]
[0056] wherein, is the signal duration, is the signal monitoring period.
[0057] S7, data storage and labeling: the center frequency , signal bandwidth , short-time signal sequence of parameters such as signal bandwidth fluctuation parameter, center frequency deviation fluctuation parameter and signal power fluctuation parameter, as the basic data of simplified electromagnetic spectrum data short-time signal dimension are stored;
[0058] Wherein:
[0059] The storage format is as follows:
[0060] { , , , BW, E, , , , , }.
[0061] S8, deep learning model matching: use deep learning model to match the frequency deviation, bandwidth deviation and signal energy deviation in different samples, and match with a certain specific target device;
[0062] Wherein:
[0063] The deep learning model uses convolutional neural network (CNN), the input is the feature vector of short-time signal sequence (such as center frequency, bandwidth, fluctuation parameter, etc.), and the output is the category label of the target device (such as “4G base station”, “frequency hopping radio station”, “illegal jammer”).
[0064] Although the present application has been described with reference to the embodiments above in the foregoing description, it is to be understood that various modifications can be made without departing from the scope of the present application. In particular, the features of the disclosed embodiments can be used in any combination without structural conflict, and the combinations of these features are not exhaustively described in the specification only for the purpose of omitting the length and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A machine learning-oriented electromagnetic spectrum data labeling method, characterized in that, The specific steps include the following: S1, data acquisition and short-time Fourier transform: using electromagnetic signal acquisition equipment to collect electromagnetic spectrum IQ data, and performing short-time Fourier transform on each cycle according to time T; S2, signal appearance detection: energy detection is performed on each short-time Fourier transform signal, if it exceeds a set relative threshold or absolute threshold , then it is determined that a signal appears; at the same time, the signal center frequency , signal bandwidth , signal energy , appearance time are recorded; S3, continuous signal detection and recording: continuously perform short-time Fourier transform and record signal samples according to S2; if signal energy exceeds a set relative threshold or absolute threshold in M continuous short-time Fourier transform periods , then consider that the signal is continuously present, and record the duration thereof ; S4, short-time signal detection and parameter calculation: detecting the corresponding center frequency , signal bandwidth whether the signal disappears, if not, adding a time identifier to the signal segment ; and simultaneously calculating the signal bandwidth fluctuation parameter, the center frequency deviation fluctuation parameter, and the signal power fluctuation parameter; S5, short-time signal sequence generation: after the above calculation, the short-time signal sequence arranged according to the appearance time and disappearance time can be obtained, and contains the signal bandwidth fluctuation parameter, the center frequency fluctuation parameter and the signal energy fluctuation parameter; S6, time-domain duty cycle calculation: statistically analyze the time-domain duty cycle of the short-time signal as another dimensional property of the short-time signal , as another dimensional property of the short-time signal S7, data storage and labeling: store the short-time signal sequence containing the parameters of the center frequency , signal bandwidth , signal bandwidth fluctuation parameters, center frequency deviation fluctuation parameters, and signal power fluctuation parameters, as the basis data of the short-time signal dimension of the simplified electromagnetic spectrum data; S8, deep learning model matching: using a deep learning model to match the frequency deviation, bandwidth deviation and signal energy deviation in different samples, and matching with a certain specific target device.
2. The electromagnetic spectrum data labeling method for machine learning according to claim 1, wherein, The window function of the short-time Fourier transform in S1 is selected as a Hamming window, the window length is N sampling points, and the overlap rate is 50%, so as to ensure the frequency-time resolution of the signal.
3. The electromagnetic spectrum data labeling method for machine learning according to claim 1, wherein, The S1 is detected by energy detection for each short-time Fourier transform signal segment, and its energy is calculated : ； wherein is the frequency domain signal after short-time Fourier transform.
4. The electromagnetic spectrum data labeling method for machine learning of claim 1, wherein, The relative threshold in S2 A typical value is set to 3 times the standard deviation of the background noise energy, and the absolute threshold A typical value is set to 10 times the system noise level.
5. The electromagnetic spectrum data labeling method for machine learning according to claim 1, wherein, The calculation process of the signal bandwidth fluctuation parameter in S4 is: the signal bandwidth deviation calculated by different short-time Fourier transform is statistically calculated to obtain the average bandwidth , the maximum bandwidth , the minimum bandwidth field, which is recorded as the signal bandwidth fluctuation parameter.
6. The machine learning oriented electromagnetic spectrum data labeling method of claim 1, wherein, The calculation process of the center frequency deviation fluctuation parameter in S4 is as follows: the center frequency deviation calculated by different short-time Fourier transforms is counted, and the average value of the center frequency is counted , the maximum frequency ; if the maximum frequency is less than 10% of the average signal bandwidth, it is determined as the same signal, and the average value of the center frequency is recorded The maximum frequency deviation is the fluctuation parameter; otherwise, it is processed as different signals.
7. The electromagnetic spectrum data labeling method for machine learning according to claim 1, wherein, The calculation process of the signal power fluctuation parameter in S4 is: the short-time signal detected by the short-time Fourier transform is subjected to average processing on the power value in the signal bandwidth to calculate the signal power average value If the signal power average value does not exceed a specific value , it is determined as the same signal, and the signal power deviation is recorded as the signal power fluctuation parameter; otherwise, it is processed as different signals.
8. The machine learning oriented electromagnetic spectrum data labeling method of claim 1, wherein, The time domain duty cycle calculation formula in S6 is as follows: ； wherein is the signal duration, is the signal monitoring period.
9. The machine learning oriented electromagnetic spectrum data labeling method of claim 1, wherein, The storage format in S7 is as follows: { , , ,BW, E, ,, , , }。 10. The electromagnetic spectrum data labeling method for machine learning according to claim 1, wherein, The deep learning model in S8 adopts a convolutional neural network, the input is a feature vector of the short-time signal sequence, and the output is a category label of the target device.