CN114584227B - Automatic burst signal detection method - Google Patents
Automatic burst signal detection method Download PDFInfo
- Publication number
- CN114584227B CN114584227B CN202210033937.3A CN202210033937A CN114584227B CN 114584227 B CN114584227 B CN 114584227B CN 202210033937 A CN202210033937 A CN 202210033937A CN 114584227 B CN114584227 B CN 114584227B
- Authority
- CN
- China
- Prior art keywords
- signal
- frequency
- time
- neural network
- burst
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract 25
- 238000003062 neural network model Methods 0.000 claims abstract 12
- 238000005070 sampling Methods 0.000 claims abstract 9
- 239000011159 matrix material Substances 0.000 claims abstract 7
- 238000000605 extraction Methods 0.000 claims abstract 5
- 238000006243 chemical reaction Methods 0.000 claims abstract 4
- 230000005670 electromagnetic radiation Effects 0.000 claims abstract 4
- 238000012545 processing Methods 0.000 claims abstract 3
- 230000003321 amplification Effects 0.000 claims abstract 2
- 238000001914 filtration Methods 0.000 claims abstract 2
- 238000012806 monitoring device Methods 0.000 claims abstract 2
- 238000003199 nucleic acid amplification method Methods 0.000 claims abstract 2
- 238000012805 post-processing Methods 0.000 claims abstract 2
- 238000012549 training Methods 0.000 claims abstract 2
- 238000013528 artificial neural network Methods 0.000 claims 4
- 238000013527 convolutional neural network Methods 0.000 claims 4
- 230000004913 activation Effects 0.000 claims 3
- 238000004364 calculation method Methods 0.000 claims 2
- 238000001228 spectrum Methods 0.000 claims 2
- 239000000284 extract Substances 0.000 claims 1
- 238000002372 labelling Methods 0.000 claims 1
- 238000005259 measurement Methods 0.000 claims 1
- 238000010606 normalization Methods 0.000 claims 1
- 238000007781 pre-processing Methods 0.000 claims 1
- 238000013215 result calculation Methods 0.000 claims 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/0082—Monitoring; Testing using service channels; using auxiliary channels
- H04B17/0087—Monitoring; Testing using service channels; using auxiliary channels using auxiliary channels or channel simulators
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B1/00—Details of transmission systems, not covered by a single one of groups H04B3/00 - H04B13/00; Details of transmission systems not characterised by the medium used for transmission
- H04B1/0003—Software-defined radio [SDR] systems, i.e. systems wherein components typically implemented in hardware, e.g. filters or modulators/demodulators, are implented using software, e.g. by involving an AD or DA conversion stage such that at least part of the signal processing is performed in the digital domain
- H04B1/0007—Software-defined radio [SDR] systems, i.e. systems wherein components typically implemented in hardware, e.g. filters or modulators/demodulators, are implented using software, e.g. by involving an AD or DA conversion stage such that at least part of the signal processing is performed in the digital domain wherein the AD/DA conversion occurs at radiofrequency or intermediate frequency stage
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/20—Monitoring; Testing of receivers
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/309—Measuring or estimating channel quality parameters
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/373—Predicting channel quality or other radio frequency [RF] parameters
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Signal Processing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Electromagnetism (AREA)
- Quality & Reliability (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Monitoring And Testing Of Transmission In General (AREA)
Abstract
The automatic burst signal detection method disclosed by the invention has the advantages of high detection speed, high detection stability and high precision. The invention is realized by the following technical scheme: in a space electromagnetic signal detection discovery scene, a monitoring device captures a space electromagnetic radiation signal by adopting an antenna and converts the space electromagnetic radiation signal into an electric signal, the electric signal is output to an acquisition module through radio frequency channel analog frequency conversion, filtering and amplification, a single-channel analog signal output by a receiving channel is subjected to AD sampling and digital down-conversion to generate a section of signal sampling data, STFT is performed, broadband time-frequency matrix data are output to a detection processing module, a training depth neural network model is loaded, after post-processing of an inference result, confidence probability, signal starting time, signal duration, signal center frequency, signal bandwidth parameters of whether the electromagnetic signal exists in the burst signal contained in the signal sampling data of the corresponding section of the time-frequency matrix are formed, and detection discovery of the electromagnetic space burst signal and parameter extraction of signal occupation time, frequency and the like are completed.
Description
Technical Field
The invention relates to a signal monitoring technology in a complex electromagnetic environment, in particular to an automatic burst signal detection method of a broadband electromagnetic spectrum monitoring receiver.
Background
In modern communication systems, such as satellite communication and mobile communication, data transmission is performed based on a wireless channel, and due to the severe environment of the wireless channel, channel parameters often show relatively severe changes with time during communication, so that burst mode is generally adopted during signal demodulation, that is, the parameters of the channel are considered to be constant in a short time. Before the transmitting end transmits data, a preamble for estimating channel parameters is typically added before the data to be transmitted. The core task of the receiving end is to detect the starting point of the received data, namely burst signal detection. The physical meaning of the method is to find out the burst starting point of the burst signal in the received signal, and find out the burst starting point of the signal by the burst signal detection technology. In general, there are several different signal patterns in an electromagnetic environment, 1) continuous fixed frequency signals. The continuous fixed frequency signal means that the signal continuously appears near a specific frequency point, the frequency is rarely changed along with the time, the signal approximately appears as a vertical line on a time-frequency diagram, and the direction of the signal is relatively stable. In a practical electromagnetic environment, more fixed frequency signals exist in a channel. A typical fixed frequency signal has a short class radius with a limited bandwidth occupation and a long duration, and a class, i.e., a class crossing portion, is constructed. 2) Intermittent fixed frequency signals. The intermittent fixed-frequency signal refers to a signal in which the signal intermittently appears at a specific frequency point, the signal is intermittent at the specific frequency point, and the direction of the signal is relatively stable. 3) Random noise signals. The random noise signal refers to various channel noise and spatial interference signals. The signals are presented as random scattered points on the time-frequency diagram. According to practical experience, the random noise signal amplitude is generally relatively small, but the random noise signal amplitude is extremely large in number, has a large relation with the frequency range and the time range, is strong in some frequency bands, and is weak in some frequency bands. 4) Burst signals, which are signals that appear randomly, last for a period of time, and then disappear, are also represented as shorter straight lines on a time-frequency diagram. The burst signal has the characteristics of burst and transient, a certain method is needed to judge whether the received signal is noise or a signal carrying data before demodulation, otherwise, the signal is lost. For burst signals, if a demodulator for continuous signals is used for demodulating burst signals, the following problems must occur that when signals exist, the demodulator accurately locks and the signals can be correctly demodulated; when the signal disappears (only noise is present), the demodulator will drift. 5) A frequency hopping signal. The transmitting power of the frequency hopping signals on different channels is consistent, and the characteristics of an automatic detection model of the frequency hopping signals of the received signals on the whole frequency band are also basically consistent. To locate and interfere with a particular frequency-hopped communication signal, an interfering party first has to detect the frequency-hopped signal and attempt to sort out the frequency set of each frequency-hopped station. The electromagnetic environment is very complex, so that the signal distribution is dense, the data volume to be processed is large, various signals are mixed together, and the characteristics bring inconvenience to the detection of the frequency hopping signals in the short wave band. The main frequency hopping signal detection and analysis methods at present comprise an autocorrelation detection algorithm, an adaptive threshold denoising algorithm and an algorithm combining an adaptive threshold and a fixed threshold, and a time-frequency analysis method. The time-frequency analysis method has great advantages in analyzing the frequency hopping signals, however, modern communication environments are more and more complex, the intercepted data volume is more and more huge due to the fact that short-wave band electromagnetic environments are more and more complex, the frequency hopping signals also have the trends of frequency band broadening, frequency hopping speed increasing and the like, and it is difficult to directly detect the frequency hopping signals from mass data by using a time-frequency amplitude three-dimensional correlation method. The methods do not emphasize the incoming wave azimuth information of the signals, have large calculation amount and are difficult to meet the requirement of rapid detection.
The burst signal detection technology includes sliding window method, double sliding window method, self-adaptive threshold energy regulation detection method, short-time autocorrelation method and frequency domain detection algorithm. Common main methods are energy methods, autocorrelation methods, and the like. The sliding window method is the most commonly used method in burst signal detection algorithms based on the received signal energy. However, the above algorithm is either poor in performance or large in calculation amount, and in any case, it is difficult to meet the requirements of high-speed burst signal detection. Among them, the sliding window method has a disadvantage in that the threshold Th value is difficult to determine.
The application range of the current development of the radio communication technology is also wider and wider, and the problem is that the demand of various industries for radio spectrum resources is rapidly increased, and the radio communication environment is also gradually complicated. In this case, the monitoring and management of radio spectrum resources are imperative, and the monitoring receiver is used as a core device for monitoring and management of the frequency spectrum and has an irreplaceable effect on purifying electromagnetic environment and improving radio spectrum utilization rate. Based on the wide application of burst communication mechanisms nowadays, electromagnetic signal monitoring equipment is a device for capturing electromagnetic signals in the environment and analyzing and processing attention signals to acquire target signal parameters. Aiming at the situation that burst communication is more and more, electromagnetic signal monitoring equipment extracts parameters from a target signal of interest, firstly, signals are detected in a complex multi-signal electromagnetic environment, the time occupation range and the frequency occupation range of each signal are determined, and single signals are filtered one by one based on the occupation ranges and used for subsequent detailed analysis and parameter extraction. In the conventional wideband electromagnetic spectrum monitoring equipment, bandwidth and center frequency detection of a signal is performed in a signal frequency domain, namely in a signal one-dimensional frequency spectrum, filtering processing is performed based on a frequency detection result, and then detection of signal starting time and burst duration is performed in a time domain. The method has the defects of low detection and discovery probability of the burst electromagnetic radiation signals, inaccurate time-frequency parameter extraction and the like.
In recent years, in the development of technologies such as power control and access modes in wireless communication, signal detection is performed by adopting a frequency and time step-by-step detection method, and the discovery probability of electromagnetic signals is lower. Therefore, a method for automatically acquiring signal time-frequency parameters by performing time-frequency conversion on signals, namely performing signal detection on a two-dimensional time-frequency diagram is proposed. The current method for detecting the signals based on the time-frequency diagram mainly adopts the steps of setting a threshold, and realizing the detection of the time-frequency parameters of the signals by fusing spectral lines which cross the threshold through the threshold of the signal amplitude. In the above detection scheme, one link is to use an adaptive threshold correlation detection algorithm to compare the peak value after matching correlation with a set threshold, and construct a decision statistic by using the ratio of signal power estimation before and after correlation, that is, to detect whether a burst signal is coming or not by judging whether the decision variable exceeds a predetermined threshold. The correlation detection algorithm has high implementation complexity, and the absolute threshold is sensitive to signal level or noise intensity. Since the decision threshold is determined by the false alarm probability P of the system f And the length of the local unique code sequenceAnd (5) determining. In the sample correlation values, more than one may be greater than the decision threshold, so that a maximum sample correlation value is selected to determine the starting boundary of the unique code in the burst signal. In the detection of unique codes by the correlation method, if a fixed threshold is adopted, the detection is difficult to be correctly performed. The method for setting the hard threshold has the problems of high signal-to-noise ratio requirement, low detection and discovery probability, high false alarm rate and the like.
Disclosure of Invention
Aiming at the defects of low detection and discovery probability, inaccurate time-frequency parameter extraction and the like of the broadband electromagnetic spectrum monitoring equipment in the prior art on the burst electromagnetic radiation signals, the invention provides an automatic burst signal detection method which can simplify the operation of the monitoring equipment, has the advantages of high burst signal detection speed, small detection error, high detection and discovery probability and high time-frequency parameter estimation precision. The invention solves the technical problems with small calculation amount, high speed and small detection error, and adopts the technical proposal that: an automatic burst signal detection method has the following technical characteristics: in a space electromagnetic signal detection discovery scene, a broadband electromagnetic spectrum monitoring device receives signals by adopting a receiving antenna, captures space electromagnetic radiation signals and converts the space electromagnetic radiation signals into electric signals, and the converted electric signals are subjected to analog frequency conversion, matched filtering and amplification by a radio frequency channel module and then output to an acquisition module for time sampling and frequency sampling; the acquisition module performs automatic AD sampling, digital channel preprocessing and digital down conversion according to a single-channel analog signal output by the radio frequency channel module, generates a section of signal sampling data, performs short-time Fourier transform (STFT) on the signal sampling data, the STFT divides the data into sections and windows, selects a time-frequency localized window function, calculates power spectrums of different moments, obtains frequency information of the function near the moment tau and a broadband time-frequency matrix of a short-time Fourier transform result, extracts a frame of broadband time-frequency matrix data and outputs the frame of broadband time-frequency matrix data to the detection processing module; the detection processing module adopts a high-precision floating point calculation operator to construct a deep neural network model, trains the deep neural network model by high-precision labeling data, loads the trained deep neural network model, takes broadband time-frequency matrix data as burst multi-signal detection and inputs the burst multi-signal detection data into the deep neural network model, and automatically generates parameter information such as confidence probability of whether burst signals exist in an electromagnetic space, signal starting time, signal duration, signal center frequency, signal bandwidth and the like through reasoning of the deep neural network model and post-processing of reasoning results, so as to form the confidence probability of whether the burst signals exist in signal sampling data of a corresponding section of a time-frequency matrix; the deep neural network model adopts a basic network and combines a multi-layer feature extraction network, utilizes a time-frequency boundary to estimate feature information under different resolution scales output by different feature network layers of the network, carries out confidence probability and time-frequency parameter estimation of whether signals exist, merges estimation results of detection results under different scales, combines detection results corresponding to the same target burst signal, sets a threshold, carries out burst signal detection on signal starting time, signal duration, signal center frequency and signal bandwidth parameters of which the probability exceeds the threshold according to the confidence probability of whether the burst signal exists, and completes detection discovery of electromagnetic space burst signals and extraction of signal occupation time and frequency parameters.
Compared with the prior art, the invention has the following beneficial effects.
The operation of the monitoring device is simplified. The invention aims at the automatic detection and discovery of the broadband electromagnetic spectrum monitoring equipment on the burst electromagnetic radiation signals, and in the scene of the detection and discovery of the space electromagnetic signals, the broadband electromagnetic spectrum monitoring equipment adopts a receiving antenna to receive the signals, and automatically adopts the antenna to capture the space electromagnetic radiation signals and convert the space electromagnetic radiation signals into electric signals; automatically carrying out analog frequency conversion, filtering and amplification on the signals through a radio frequency channel and outputting the signals to an acquisition module; the acquisition module automatically performs AD sampling and digital down-conversion on a single-channel analog signal output by a receiving channel to generate a section of signal sampling data, performs short-time Fourier transform on the signal sampling data, and outputs broadband time-frequency matrix data to the detection processing module; the detection processing model utilizes the deep neural network model to automatically extract existing signals, time and frequency parameters, a time-frequency diagram visualized by a time-frequency matrix is not needed to be observed manually, the signal parameters are not needed to be determined manually according to burst signals selected by a time-frequency diagram frame, parameters such as a signal time occupation range, a frequency occupation range and the like are automatically extracted by a machine, the design of monitoring equipment is greatly simplified, the manual operation is reduced, and the reliability of engineering realization is improved.
The burst signal detection speed is high. The invention aims at space electromagnetic radiation observation, discovers burst electromagnetic signals through automatic comprehensive processing, extracts confidence probability, signal starting time, signal duration, signal center frequency and signal bandwidth parameters, performs automatic AD sampling, digital channel preprocessing and digital down-conversion according to a single-channel analog signal output by a radio frequency channel module, generates a section of signal sampling data, performs short-time Fourier transform (STFT) on the signal sampling data, and performs data segmentation windowing on the STFT, selects a time-frequency localized window function, calculates power spectrums at different moments, obtains frequency information of the function near a moment tau and a broadband time-frequency matrix of a short-time Fourier transform result, and extracts a frame of broadband time-frequency matrix data and outputs the broadband time-frequency matrix data to a detection processing module; after down-conversion and short-time Fourier transformation, broadband time-frequency matrix data are formed, the method is efficient, convenient and fast, the calculated amount is small, parallel calculation is easy, the speed is high, a deep neural network model can be built through a basic calculation unit, and automatic reasoning is achieved through the deep neural network model. Meanwhile, the deep neural network model can be deployed on a high-performance GPU or a special AI processing chip to perform quick parallel reasoning, so that the deep neural network model can quickly calculate and broadband time-frequency matrix data can quickly process and output the parameter information such as confidence probability, signal starting time, signal duration, signal center frequency, signal bandwidth and the like of whether the parameter information exists. The detection speed of the broadband spectrum monitoring equipment on the burst electromagnetic radiation signals is greatly improved.
The detection error is small, and the detection discovery probability is high. According to the invention, a high-precision floating point calculation operator is adopted to construct a deep neural network model aiming at various space radiation electromagnetic signals, the deep neural network model is trained by high-precision labeling data, the deep neural network model is trained by a supervised training mode based on the deep neural network model, the trained deep neural network model is utilized to process broadband time-frequency matrix data, and parameter information such as confidence probability, signal starting time, signal duration, signal center frequency, signal bandwidth and the like of whether burst signals exist in an electromagnetic space are automatically generated. Floating point calculation and high-precision reasoning are adopted, broadband time-frequency data is taken as input, and a neural network is utilized to automatically estimate confidence probability, signal starting time, signal duration, signal center frequency and signal bandwidth parameters of whether a plurality of burst signals exist in the time-frequency data. The high-stability and high-precision detection of the electromagnetic radiation signals by the broadband electromagnetic spectrum monitoring equipment is realized. The method has the advantages of high detection stability, high detection precision and high time-frequency parameter estimation precision. The invention can be used for verification from the aspects of frequency offset capturing range, signal-to-noise ratio working threshold, level receiving range and the like, and the verification result shows that the invention can reliably work at lower signal-to-noise ratio, has moderate channel adaptability and constant false alarm probability characteristics, and is an effective detection method suitable for detecting burst signals of a TDMA system.
Drawings
For the purpose of further explanation and not limitation of the above-described implementations of the present invention, the following description gives the best embodiments with reference to the accompanying drawings so as to make the details and advantages of the present invention more apparent.
Fig. 1 is a schematic diagram of an automatic burst signal detection principle based on a multi-layer deep neural network.
Fig. 2 is a schematic diagram of a reference bounding box for each scale of the signal detection arrangement of the present invention.
Fig. 3 is a schematic diagram of a residual network structure in the deep neural network model of the present invention.
Detailed Description
See fig. 1. According to the invention, in a space electromagnetic signal detection discovery scene, a broadband electromagnetic spectrum monitoring device adopts a receiving antenna to receive signals, captures space electromagnetic radiation signals and converts the space electromagnetic radiation signals into electric signals, and the converted electric signals are subjected to analog frequency conversion, matched filtering and amplification through a radio frequency channel module and then output to an acquisition module for time sampling and frequency sampling; the acquisition module performs automatic AD sampling, digital channel preprocessing and digital down conversion according to a single-channel analog signal output by the radio frequency channel module, generates a section of signal sampling data, performs short-time Fourier transform (STFT) on the signal sampling data, the STFT divides the data into sections and windows, selects a time-frequency localized window function, calculates power spectrums of different moments, obtains frequency information of the function near the moment tau and a broadband time-frequency matrix of a short-time Fourier transform result, extracts a frame of broadband time-frequency matrix data and outputs the frame of broadband time-frequency matrix data to the detection processing module; the detection processing module adopts a high-precision floating point calculation operator to construct a deep neural network model, trains the deep neural network model by high-precision labeling data, loads the trained deep neural network model, takes broadband time-frequency matrix data as burst multi-signal detection and inputs the burst multi-signal detection data into the deep neural network model, and automatically generates parameter information such as confidence probability of whether burst signals exist in an electromagnetic space, signal starting time, signal duration, signal center frequency, signal bandwidth and the like through reasoning of the deep neural network model and post-processing of reasoning results, so as to form the confidence probability of whether the burst signals exist in signal sampling data of a corresponding section of a time-frequency matrix; the deep neural network model adopts a basic network and combines a multi-layer feature extraction network, utilizes a time-frequency boundary to estimate feature information under different resolution scales output by different feature network layers of the network, carries out confidence probability and time-frequency parameter estimation of whether signals exist, merges estimation results of detection results under different scales, combines detection results corresponding to the same target burst signal, sets a threshold, carries out burst signal detection on signal starting time, signal duration, signal center frequency and signal bandwidth parameters of which the probability exceeds the threshold according to the confidence probability of whether the burst signal exists, and completes detection discovery of electromagnetic space burst signals and extraction of signal occupation time and frequency parameters.
When the deep neural network model is trained, the combination of confidence coefficient errors of the target signals and time-frequency parameter estimation errors is adopted as a loss function for training the deep neural network model, and the cost function is calculated in the following way:
wherein N is the number of effective signals in the time-frequency data, lambda code is the weight of time-frequency parameter estimation error, L loc An error is estimated for the time-frequency parameter.
The deep neural network model is weighted in a manner of taking cross entropy as a loss measure and focalloss, confidence errors are calculated,the time-frequency boundary estimation network calculates the time-frequency parameter estimation error by adopting the following calculation mode,
wherein y is i Representing the true result of classification,Representing the prediction result of the classification, f i 、t i 、b i 、l i Andand respectively representing the real result and the predicted result of the time-frequency parameter. The loss function reduces the weight of a large number of simple negative samples in training, and solves the problem of serious unbalance of the proportion of the positive and negative samples in one-stage target detection.
In an alternative embodiment, the input 1-channel amplitude data of the neural network is 1×128×1024, 128-dimensional time data and 1024-dimensional frequency data, the input data is processed through the feature extraction network to form at least 4 different-scale feature information, wherein the dimension of the scale 1 feature information is 32 channels 16×128, the dimension of the scale 2 feature information is 48 channels 8×64, the dimension of the scale 3 feature information is 64 channels 4×32, and the dimension of the scale 4 feature information is 128 channels 2×16.
See fig. 2. The deep neural network model detects signals of each Cell in the characteristics by taking 8 default bounding boxes with different proportions as references according to the characteristic information of each scale, and sets the length-width ratio of the 8 default bounding boxes according to the bandwidth and the duration equal-time frequency characteristics of electromagnetic signalsExamples are 1:1, 1.5:1.5, 1:2, 2:1, 1:3, 3:1, 1:5, 5:1. For each Cell in the feature information under each scale, the time-frequency boundary estimation network forms 8 estimation results, namely, each default boundary box corresponds to one estimation result, and each estimation result contains information recorded as [ P, f ] c ，t c ，b c ，l c ]Where P represents the confidence of the presence and absence of targets. f (f) c Representing the relative center frequency, t, of the signal c Representing the relative central time of the signal (start time + half the signal duration), b c Representing relative signal bandwidth, l c Indicating the relative duration of the signal.
The time-frequency boundary estimation network adopts a relative estimation result calculation mode to calculate:
target signal relative center frequency
Target signal relative to center time
Relative signal bandwidth
Signal relative time length
The relative estimation result output by the time-frequency boundary estimation network can obtain the original center frequency f of the target signal after the inverse transformation c Time t of center g Bandwidth b g Sum duration l g ，
Wherein f g ,t g ,b g ,l g For the true center frequency, center time, bandwidth and duration of the target signal, f b ,t b ,b b ,l b Center frequency, center time, bandwidth, duration represented by the reference bounding box.
The deep neural network detects the task for the bursty target signal,the multi-layer convolutional neural network is adopted to extract the characteristics of the input time-frequency data, forecast the information result of the target signal, and the residual network is used as a main network model component to better complete training and forecast. The multilayer characteristic extraction network adopts 26 layers of neural networks to extract the characteristics, and usesRepresenting two layers of neural networks connected in sequence as a residual network block and overlapping 3 in succession, wherein [16,3×3]Representing a convolutional neural network outputting 16 channels 3 x 3.
The feature extraction network design parameters in the deep neural network model are shown below:
see fig. 3. The residual block network block inputs multi-channel feature data using layer normalization (LayerNorm) and activation function (layrelu), first performs a 3×3 convolutional neural network, then sequentially superimposes the layer normalized LayerNorm layer, the layrelu activation layer, the 3×3 convolutional neural network, the LayerNorm layer, and then adds with the input through a 1×1 convolutional neural network layer, and finally superimposes the layrelu activation layer.
While the present invention has been described in detail with reference to the drawings, it should be noted that the foregoing examples are merely preferred examples of the present invention, and are not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art, such as the number of layers of different neural networks, the number of channels of each layer of neural network, the size parameters of the convolution kernel, etc. may be selected in connection with specific engineering projects. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.
Claims (10)
1. An automatic burst signal detection method is characterized in that: in a space electromagnetic signal detection discovery scene, a broadband electromagnetic spectrum monitoring device receives signals by adopting a receiving antenna, captures space electromagnetic radiation signals and converts the space electromagnetic radiation signals into electric signals, and the converted electric signals are subjected to analog frequency conversion, matched filtering and amplification by a radio frequency channel module and then output to an acquisition module for time sampling and frequency sampling; the acquisition module performs automatic AD sampling, digital channel preprocessing and digital down conversion according to a single-channel analog signal output by the radio frequency channel module, generates a section of signal sampling data, performs short-time Fourier transform (STFT) on the signal sampling data, the STFT divides the data into sections and windows, selects a time-frequency localized window function, calculates power spectrums of different moments, obtains frequency information of the function near the moment tau and a broadband time-frequency matrix of a short-time Fourier transform result, extracts a frame of broadband time-frequency matrix data and outputs the frame of broadband time-frequency matrix data to the detection processing module; the detection processing module adopts a high-precision floating point calculation operator to construct a deep neural network model, trains the deep neural network model by high-precision labeling data, loads the trained deep neural network model, inputs wideband time-frequency matrix data into the deep neural network model, and automatically generates parameter information of confidence probability, signal starting time, signal duration, signal center frequency and signal bandwidth of whether burst signals contained in an electromagnetic space exist or not through reasoning of the deep neural network model and post-processing of reasoning results, so as to form the confidence probability of whether the burst signals contained in signal sampling data of a corresponding section of the time-frequency matrix; the deep neural network model adopts a basic network and combines a multi-layer feature extraction network, utilizes feature information under different resolutions output by a time-frequency boundary estimation network to estimate confidence probability and time-frequency parameters of whether signals exist, merges estimation results of detection results under different resolutions, combines detection results corresponding to the same target burst signal, sets a threshold, and performs burst signal detection on signal starting time, signal duration, signal center frequency and signal bandwidth parameters with the confidence probability value exceeding the threshold according to the confidence probability of whether the burst signal exists, thereby completing detection discovery of the electromagnetic space burst signal and extraction of signal occupation time and frequency parameters.
2. As claimed in claim 1The automatic burst signal detection method is characterized by comprising the following steps of: when the deep neural network model is trained, the combination of confidence coefficient errors of the target signals and time-frequency parameter estimation errors is adopted as a loss function for training the deep neural network model, and the cost function is calculated in the following way:；
wherein N is the number of effective signals in the time-frequency data, lambda coord To weight the time-frequency parameter estimation error, L loc An error is estimated for the time-frequency parameter.
3. The automated burst signal detection method of claim 1, wherein: the deep neural network model weights by adopting cross entropy as a loss measurement and a loss function Focal loss on the basis of the cross entropy loss function, calculates a confidence error,；
the time-frequency boundary estimation network calculates the time-frequency parameter estimation error by adopting the following calculation mode,
wherein y is i Representing the true result of classification,Representing the prediction result of the classification, f i 、t i 、b i 、l i And->And respectively representing the real result and the predicted result of the time-frequency parameter.
4. The automated burst signal detection method of claim 1, wherein: the input data of the deep neural network is 1×128×1024, 128-dimensional time data and 1024-dimensional frequency data, the input data forms at least 4 different-scale feature information through a feature extraction network, wherein the dimension of the scale 1 feature information is 32 channels 16×128, the dimension of the scale 2 feature information is 48 channels 8×64, the dimension of the scale 3 feature information is 64 channels 4×32, and the dimension of the scale 4 feature information is 128 channels 2×16.
5. The automated burst signal detection method of claim 1, wherein: the deep neural network model detects signals of each Cell in the characteristics by taking 8 default bounding boxes with different proportions as references according to characteristic information of each scale, and the aspect ratio of the 8 default bounding boxes is set to be 1:1, 1.5:1.5, 1:2, 2:1, 1:3, 3:1, 1:5 and 5:1 according to the bandwidth and time-frequency characteristics of electromagnetic signals.
6. The automated burst signal detection method of claim 1, wherein: for each Cell in the feature information under each scale, the time-frequency boundary estimation network forms 8 estimation results, namely, each default boundary box corresponds to one estimation result, and each estimation result contains information recorded as [ P, f ] c ,t c ,b c ,l c ]Wherein P represents confidence of the presence and absence of the target, f c Representing the relative center frequency of the signal, t c Representing the relative central time of the signal, b c Representing relative signal bandwidth, l c Indicating the relative duration of the signal.
7. The automated burst signal detection method of claim 1, wherein: the time-frequency boundary estimation network adopts a relative estimation result calculation mode to calculate:
target signal relative center frequency
Target signal relative to center time
Relative signal bandwidth
Signal relative time length
The relative estimation result output by the time-frequency boundary estimation network is inverse transformed to obtain the absolute center frequency f of the target signal g Time t of center g Bandwidth b g Sum duration l g ，
Wherein f d ,t d ,b d ,l d Center frequency, center time, bandwidth, duration represented by the reference bounding box.
8. The automated burst signal detection method of claim 1, wherein: aiming at the burst target signal detection task, the deep neural network model adopts a multi-layer convolution neural network to extract the input time-frequency data characteristics, predicts the target signal information result, and uses a residual network as a main network model component.
9. The automated burst signal detection method of claim 1, wherein: the multilayer characteristic extraction network adopts 26 layers of neural networks to extract the characteristics, and usesRepresenting a residual network block connected in sequence by two layers of neural networks and overlapping 3 in succession, wherein [16,3×3]Representing a convolutional neural network outputting 16 channels 3 x 3.
10. The automated burst signal detection method of claim 9, wherein: the residual network block inputs multi-channel feature data by using layer normalization (LayerNorm) and activation function (layerelu), firstly carries out 3×3 convolutional neural network, then sequentially superimposes normalized LayerNorm layer, layerelu activation layer, 3×3 convolutional neural network, layerNorm layer and input 1×1 convolutional neural network layer for addition, and finally superimposes the layerelu activation layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210033937.3A CN114584227B (en) | 2022-01-12 | 2022-01-12 | Automatic burst signal detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210033937.3A CN114584227B (en) | 2022-01-12 | 2022-01-12 | Automatic burst signal detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114584227A CN114584227A (en) | 2022-06-03 |
| CN114584227B true CN114584227B (en) | 2023-08-22 |
Family
ID=81769094
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210033937.3A Active CN114584227B (en) | 2022-01-12 | 2022-01-12 | Automatic burst signal detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114584227B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116244637B (en) * | 2023-05-12 | 2023-08-01 | 中星联华科技（北京）有限公司 | Burst signal acquisition method and device |
| CN118631662B (en) * | 2024-08-15 | 2024-11-26 | 深圳市友恺通信技术有限公司 | A communication data management system and method based on data analysis |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH1139495A (en) * | 1997-07-24 | 1999-02-12 | Nec Corp | Image supervisory device |
| WO2001048952A1 (en) * | 1999-12-24 | 2001-07-05 | Ntt Docomo, Inc. | Method and device for transmitting burst signal in mobile communication system, information distribution method, and information distribution controller |
| CN105207965A (en) * | 2015-08-14 | 2015-12-30 | 成都中安频谱科技有限公司 | Automatic VHF/UHF frequency range modulation identification method |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
| CN109271975A (en) * | 2018-11-19 | 2019-01-25 | 燕山大学 | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification |
| CN112735435A (en) * | 2020-12-25 | 2021-04-30 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Voiceprint open set identification method with unknown class internal division capability |
| CN112751633A (en) * | 2020-10-26 | 2021-05-04 | 中国人民解放军63891部队 | Broadband spectrum detection method based on multi-scale window sliding |
-
2022
- 2022-01-12 CN CN202210033937.3A patent/CN114584227B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH1139495A (en) * | 1997-07-24 | 1999-02-12 | Nec Corp | Image supervisory device |
| WO2001048952A1 (en) * | 1999-12-24 | 2001-07-05 | Ntt Docomo, Inc. | Method and device for transmitting burst signal in mobile communication system, information distribution method, and information distribution controller |
| CN105207965A (en) * | 2015-08-14 | 2015-12-30 | 成都中安频谱科技有限公司 | Automatic VHF/UHF frequency range modulation identification method |
| CN107122738A (en) * | 2017-04-26 | 2017-09-01 | 成都蓝色起源科技有限公司 | Automatic Communication Signals Recognition based on deep learning model and its realize system |
| CN109271975A (en) * | 2018-11-19 | 2019-01-25 | 燕山大学 | A kind of electrical energy power quality disturbance recognition methods based on big data multi-feature extraction synergetic classification |
| CN112751633A (en) * | 2020-10-26 | 2021-05-04 | 中国人民解放军63891部队 | Broadband spectrum detection method based on multi-scale window sliding |
| CN112735435A (en) * | 2020-12-25 | 2021-04-30 | 西南电子技术研究所（中国电子科技集团公司第十研究所） | Voiceprint open set identification method with unknown class internal division capability |
Non-Patent Citations (1)
| Title |
|---|
| 冉雨.自适应突发通信系统的决策技术研究.《中国优秀硕士学位论文全文数据库信息科技辑(月刊)，2020年第07期》.2020,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114584227A (en) | 2022-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yang et al. | Cyclostationary feature detection based spectrum sensing algorithm under complicated electromagnetic environment in cognitive radio networks | |
| US10879946B1 (en) | Weak signal processing systems and methods | |
| CN114584227B (en) | Automatic burst signal detection method | |
| CN107911183B (en) | Unmanned aerial vehicle identification method and system based on real-time signal characteristic analysis | |
| CN111693944A (en) | Radar active interference signal parameter extraction and interference pattern identification method and device | |
| CN102244548B (en) | Cooperative spectrum detection method, cooperative spectrum sensing apparatus and spectrum detection system | |
| CN110099019A (en) | LoRa Modulation Signal Detection Method based on deep learning | |
| CN108063641B (en) | Single-channel multi-signal detection method | |
| CN105743612A (en) | Method for real-time blind demodulation of large-frequency-shift short-time burst signal | |
| CN106170139B (en) | A kind of frequency spectrum detecting method and system | |
| Kumar et al. | MDI-SS: matched filter detection with inverse covariance matrix-based spectrum sensing in cognitive radio | |
| CN112003803B (en) | Detection and reception equipment for VHF and UHF band aviation radio station signals | |
| CN112929141A (en) | Unmanned aerial vehicle detection and identification method and system based on graph-borne signal matching | |
| CN103905129B (en) | The input analyzed based on spectral pattern and signal message interpretation method | |
| CN106656372B (en) | Frequency band interference detection method of frequency hopping system | |
| CN108718223B (en) | Blind spectrum sensing method for non-cooperative signals | |
| CN108401255B (en) | A Two-Stage Blind Spectrum Sensing Scheme | |
| CN115632724B (en) | Anti-interference channelized detection method based on weighted Gaussianity test | |
| CN117318853A (en) | Burst signal detection method under low signal-to-noise ratio | |
| CN117335864A (en) | A dual-window filtering detection method based on digital channelization technology | |
| Ding et al. | Data Fusion Network-Based Time–Frequency Enhancement Algorithm for Doppler Through-Wall Radar Tracking | |
| CN114401526B (en) | Narrow-band interference position detection method and system based on double threshold decision | |
| CN115270844A (en) | AI (Artificial intelligence) identification method for signal characteristics of unmanned aerial vehicle | |
| CN119807861B (en) | A UAV detection method and system based on pilot synchronization sequence matching | |
| CN115099289B (en) | System and method for distinguishing identity of radio radiation source based on weak fingerprint characteristics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |