CN110991378A - Power amplifier individual identification method and device - Google Patents
Power amplifier individual identification method and device Download PDFInfo
- Publication number
- CN110991378A CN110991378A CN201911263371.8A CN201911263371A CN110991378A CN 110991378 A CN110991378 A CN 110991378A CN 201911263371 A CN201911263371 A CN 201911263371A CN 110991378 A CN110991378 A CN 110991378A
- Authority
- CN
- China
- Prior art keywords
- power amplifier
- signal
- data
- signals
- network
- 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.)
- Granted
Links
Images
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- 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)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a power amplifier individual identification method, which extracts different fingerprint characteristics by utilizing the difference of extraction characteristics of Efficientnets series heterogeneous convolution neural networks under the condition of low signal-to-noise ratio, and improves the extraction capability of the fingerprint characteristics by integrating the prediction results of a plurality of networks by using a stacking method; the invention also provides a power amplifier individual identification device, which comprises a signal receiving module, a digitization module, a data processing module and a power supply, and the device has a simple and practical structure; the method has small parameter quantity and floating point operand, and is suitable for being transplanted to embedded hardware equipment such as FPGA and the like for use; the method can accurately identify the power amplifier of a specific radiation source by receiving the steady-state signals of the radiation sources in the same batch of non-cooperative parties, further determine the specific direction of the specific radiation source by methods such as radio positioning and the like, and track the motion track of the specific radiation source.
Description
Technical Field
The invention relates to the technical field of radiation source identification, in particular to an individual identification method and an individual identification device of a steady-state signal power amplifier.
Background
Non-stationary components of some signals may exhibit periodic stationarity in terms of high order statistics, in particular, the signal may exhibit single or multi-periodic stationary variations in mean, correlation functions, and high order cumulants. These phenomena are common in artificial signals such as communications and radar, or in signals that vary periodically according to a natural law. In the memory effect power amplifier nonlinear behavior model, the periodicity of high-order statistics is an important feature in a signal, and fingerprint information can be extracted.
There are slight differences in the output signals of the same type of power amplifier due to non-linear characteristics, and these differences are called fingerprint features. The fingerprint features have shape and position inconsistency on the circular autocorrelation spectrogram.
At present, identification and classification researches aiming at steady-state signals of a radiation source are less, and a common technology at present is a document 'individual identification method of an aerial target radiation source' [ J/OL ]. Lianqia, face and aspin, Zhanglin, http:// kns.cnki.net/kcms/detail/11.2422.TN.20190822.0747.004.html.2019-08-22/2019-10-24. The method is based on time domain signal analysis, has weak anti-interference capability, obtains better results when the signal difference is obvious, but has poor performance when the signal difference is small or the signal-to-noise ratio is low, and is easily influenced by preset parameters. The document "communication radiation source individual identification based on 3D-Hibert energy spectrum and multi-scale fractal features" [ J ]. Korea, billow, Wanghuang, etc. the Communications, 2017, 38(4):99-109. processing radiation source signals using a modified Hilbert-Huang transform, verifies the effectiveness in analyzing transient signals, but does not consider the processing of steady-state signals. The literature Radio Transmitter Classification using a New Method of the strates simulations Combined with PCA [ C ]/Xu S, Huang B, Xu L, et al/Military Communications conference, Orlando, FL, USA: IEEE,2007, pp.1-5, proposes the use of SIB (rectangular-integrated bispectrum) for extracting signal spurs, but SIB contains less information and is less resistant to noise than other methods. The document 'communication radiation source nonlinear individual identification method research' [ D ]. Tang Chiling, Xian: Sian electronics technology university, 2013. a classification identification method based on high-order tensor features and a support vector machine is researched, and feasibility of the method is verified through experiments. However, these methods are relatively traditional methods, and with the rapid development of convolutional neural networks in recent years, there is a strong need for a method for processing the problem of individual classification of radiation sources by fully utilizing the feature extraction capability of convolutional networks.
Because each part in the radiation source can cause combined interference to signals, the stray noise forming mechanism of the radiation source is difficult to analyze, and a mathematical model of fingerprint characteristics can not be established for the whole radiation source.
Fingerprint features generated in the working process of the radiation source mainly come from stray noise generated by the power amplifier, and the identification rate is reduced easily due to the small difference between the fingerprint features of the radiation source and noise interference.
Because the radio frequency power amplifier usually works in a nonlinear area, signals output by the power amplifier have a new frequency component generated in the nonlinear area of the power amplifier besides the difference between a modulation mode and carried information. The fingerprint model is put to model a new component generated in a nonlinear area, wherein the Hammerstein model has the advantages of flexible structure, small calculated amount and simple solution. The Hammerstein model modeling shows that the stray noise generated by the power amplifier contains the characteristics of a cyclostationary signal.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a power amplifier individual identification method, which has the basic principle that under the condition of low signal-to-noise ratio, different fingerprint features are extracted by utilizing the difference of extraction features of heterogeneous convolutional neural networks, prediction is made according to the respective extracted features, the extraction capability of the fingerprint features is improved by integrating the prediction results of a plurality of networks by using an integrated model stacking method, and meanwhile, in order to avoid overlarge model scale caused by the improvement of classification accuracy, an Efficientnets series heterogeneous convolutional neural network with small network scale and large structural difference is used as a basic network, and the method specifically comprises the following steps:
step 1: collecting a power amplifier signal;
step 2: performing cyclic autocorrelation preprocessing on the acquired signal data of the power amplifier to obtain a two-dimensional cyclic spectrum;
and step 3: establishing a base classifier, adjusting two-dimensional cyclic spectrum data, and connecting four heterogeneous convolutional neural networks (Efficientnets) into an integrated network by using a stacking model;
and 4, step 4: segmenting the collected power amplifier signals, and training a base classifier and a stacking model of a second layer;
and 5: and predicting the individual power amplifiers by using the trained integrated network.
Further, in step 1, a receiver is used for receiving and converting output signals of a plurality of different power amplifiers into baseband signals, the baseband signals are I/Q two-path analog signals, and then the I/Q two-path analog signals are converted into two-path digital signals through analog-to-digital conversion and stored as power amplifier signal data in a dat format;
step 2, slicing the power amplifier signal data according to every 800 sampling points to obtain sliced data, then performing cyclic autocorrelation on the sliced data to obtain a two-dimensional cyclic spectrum, and storing the generated two-dimensional cyclic spectrum in a storage device;
the step 3 specifically comprises the following steps:
step 3.1: using an MBConv network structure as a backbone network, and establishing four heterogeneous convolutional neural networks in an Efficientnets manner;
step 3.2: scaling and cutting the two-dimensional cyclic spectrum slice to enable the two-dimensional cyclic spectrum slice to only contain information of a power amplifier signal period, and adjusting the size of the input ends of the four heterogeneous convolutional neural networks according to the size of the cut two-dimensional cyclic spectrum slice;
step 3.3: rotating and folding each two-dimensional cyclic spectrum picture cut in the previous step by using an image amplification method, and increasing the number of the two-dimensional cyclic spectrum pictures to obtain a new two-dimensional cyclic spectrum picture set;
step 3.4: connecting four heterogeneous convolutional neural networks (Efficientnets) into an integrated network by using a stacking model, wherein the integrated network is divided into two layers of models, the first layer of model of the integrated network is a plurality of heterogeneous Efficientnets, each EfficientNet of the first layer is a base classifier, and the second layer of model of the integrated network is the stacking model;
the step 4 specifically comprises the following steps:
step 4.1: a new two-dimensional cyclic spectrum picture set is obtained in the segmentation step 3.3;
step 4.2: training four base classifiers and a stacking model of the second layer.
Further, the collected power amplifier signal data is stored in a server hard disk in the step 1;
in step 2, the picture size of the two-dimensional cyclic spectrum is 300 pixels each of length and width, and the two-dimensional cyclic spectrum is obtained by adopting the following formula:
where R is the result of a two-dimensional cyclic spectral transformation of the signal, τ is the time delay, T is the time duration of the signal being processed, x (T) is the signal being processed, and α is the cyclic frequency that can be passedCalculated as m is a constant, T0Is the period of the cyclostationary signal;
in the step 3.1, the four heterogeneous convolutional neural network structures are determined by an Efficientnets algorithm, and the initial weight of the network is initialized randomly;
the operation of rotating the two-dimensional cyclic spectrum picture in the step 3.3 is to rotate the picture by 180 degrees according to the picture center, and the specific implementation method of the picture folding operation is to rotate by 180 degrees by taking the horizontal center line of the picture as an axis;
in step 3.4, the number of the base classifiers is n, and the base classifiers are respectively set to be M1,M2…MnThe output of the second layer model is the prediction of the individual radiation source
Step 4.1, the data set Y obtained in the step 3.3 is divided into two parts, the first part accounts for 20% of all data and is used as a test set of a second layer model of the integrated network, and Te is set2(ii) a The second part accounts for 80% of all data, is used as a training set of the first layer model of the integrated network and is set as T1(ii) a Data set T1The average is divided into 4 parts, each is S1,S2,S3,S4(ii) a Setting n different base classifiers of the first layer as M1,M2…Mn;
Step 4.2, selecting the GPU as an operation platform, training each base classifier by a gradient descent method according to data segmentation results, obtaining M weight models by each base classifier according to M data segmentation results, and training a certain base classifier MnThen, 3 parts of the data are extracted from 4 parts of the data as a basis classifier MnOne of the redundant training sets is used as a base classifier MnThe test set of (2) obtains the training results of 4 different network models of a certain base classifier, which are respectively set as Mn1,Mn2…Mn4The 4 weighted predictors are grouped together as a new training set for the second layer and set as a new training set Tr2The test set of the second layer model is 20% of data Te divided from all data2,Te2Te is obtained by predicting the trained 4 network training results2The four prediction results are averaged to obtain a test set Te 'of the second layer model'2And set as a new test set; using data sets Tr2Training an Integrated network second layer model, Using dataset Te'2And taking the prediction result X output by the second layer model as the final prediction result.
Further, the second layer model of the integrated network may be of the type XGBoost or LightGBM.
Furthermore, a method of using Test Time Amplification (TTA) is adopted in the network prediction, the picture to be predicted and the picture after the operation such as rotation and folding are carried out on the picture are simultaneously sent to the network prediction, and the prediction results of the two pictures are synthesized.
Further, cosine annealing decay is used in step 4.2 instead of the gradient descent method.
The invention also discloses a power amplifier individual identification device which comprises a signal receiving module, a digitization module, a data processing module and a power supply, wherein the signal receiving module receives radio signals, and converts the radio signals into two paths of analog signals after filtering, amplifying and orthogonalizing processing, and then sends the two paths of analog signals to the digitization module, the digitization module carries out analog-to-digital conversion on the two paths of analog signals and then stores the analog signals to a cache unit and then sends the analog signals to the data processing module, the data processing module comprises an FPGA chip for realizing the method, the signals are processed by the FPGA chip and then output prediction results, and the power supply is connected with other modules and provides voltages required by other modules when the other modules work.
Further, the signal receiving module comprises a signal receiving antenna, a band-pass filter, a low-noise power amplifier, a frequency mixer, a local oscillator, a 90-degree phase shifter and a low-pass filter, wherein the signal receiving antenna, the band-pass filter and the low-noise power amplifier are sequentially connected, the low-noise power amplifier is connected with the two frequency mixers, the local oscillator is also connected with the two frequency mixers, the 90-degree phase shifter is added between the local oscillator and the frequency mixers, two paths of I/Q analog signals are output by the two frequency mixers, and the two paths of I/Q analog signals are output to the digitizing module through the low-pass;
the digital module comprises an analog-to-digital converter and a cache unit, wherein external input signals are respectively accessed into the analog-to-digital converter to obtain two paths of orthogonal digital signals, and the two paths of orthogonal digital signals are sequentially stored in the cache unit in a paired mode;
the data processing module is an FPGA board card and is responsible for converting the received I/Q data into a two-dimensional spectrogram, cutting the spectrogram and then sending the cut spectrogram into a trained convolution network model to predict the individual category of the signal to obtain a final result;
the power supply forms a single-channel output dc/dc power supply by using the LTM4644 to supply power to the signal receiving module; the LTM4620 is used for supplying power to the data processing module where the FPGA board card is located.
Furthermore, an AMC7820 chip of TI is adopted by the analog-to-digital converter, a 16GB MLCflash memory device is adopted by the cache unit, and an Ultra96-V2FPGAFPGA board card is adopted by the FPGA board card.
The invention also provides a detection method of the individual identification method of the power amplifier, which comprises the following steps:
step 1: selecting power amplifiers of two or more signal sources as an individual radiation source to be identified;
step 2: setting parameters of a signal source for transmitting signals;
and step 3: selecting a signal receiving apparatus;
and 4, step 4: placing a signal receiver and a signal transmitter;
and 5: receiver sampling and signal conversion;
step 6: storing the acquired discrete signals and sending the discrete signals to FPGA equipment for pretreatment;
and 7: amplifying a data image;
and 8: the individual power amplifiers are predicted using a neural network.
The method can be applied to individual identification of intelligent communication countermeasure equipment and electromagnetic spectrum monitoring equipment. The power amplifier of a specific radiation source can be accurately identified by receiving steady-state signals of the radiation sources in the same batch of non-cooperative parties, and then the specific direction of the specific radiation source is determined by methods such as radio positioning and the like, and the motion track of the specific radiation source is tracked.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of data segmentation in step 2 of the method of the present invention;
FIG. 3 is a diagram of a hardware system according to the present invention;
FIG. 4 is a schematic diagram of an internal mechanism of the FPGA board card in the invention;
FIG. 5a is a graph comparing the performance of the process of the present invention with other processes;
FIG. 5b is an enlarged partial view of a graph comparing the performance of the process of the present invention with other processes.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples.
Firstly, constructing a two-dimensional structure diagram of a source signal, fingerprint characteristics and noise from original data by utilizing cyclic autocorrelation; then, extracting features such as an image structure, fine textures and the like by using a convolution network, and predicting data according to the extracted features; finally, a stacking method is used for reducing prediction errors caused by unobvious features under the condition of low signal-to-noise ratio, and fig. 1 is a flow chart of the method of the invention, which comprises the following specific steps:
step 1: and collecting a power amplifier signal.
Inside the signal transmitter, the baseband signal reaches the power amplifier after being modulated and up-converted, and the baseband signal and the modulated signal received by different power amplifiers may be different. The output signals of a plurality of different power amplifiers are received and converted into baseband signals by a receiver, the baseband signals are I/Q two-path analog signals, and then the I/Q two-path analog signals are converted into two-path digital signals through analog-to-digital conversion and stored as power amplifier signal data in a dat format. And storing the collected power amplifier signal data into a server hard disk for further processing.
Step 2: and (3) performing cyclic autocorrelation preprocessing on the acquired signal data of the power amplifier to obtain a two-dimensional cyclic spectrum.
The method comprises the steps of taking out signal data of the power amplifier from a hard disk of a server, slicing the data according to 800 sampling points to obtain sliced data, then carrying out cyclic autocorrelation on the sliced data to obtain a two-dimensional cyclic spectrum, and extracting periodic characteristics in collected signals by utilizing the cyclic autocorrelation so as to achieve the purpose of separating stray harmonic noise generated by the power amplifier.
Specifically, a two-dimensional cyclic spectrum is obtained by using the following formula, and the picture size of the two-dimensional cyclic spectrum is 300 pixels each of the length and the width:
where R is the result of a two-dimensional cyclic spectral transformation of the signal, τ is the time delay, T is the time duration of the signal being processed, x (T) is the signal being processed, and α is the cyclic frequency that can be passedCalculated as m is a constant, T0The period of the cyclostationary signal.
The generated two-dimensional cyclic spectrum picture is temporarily stored in a storage device and is further cut after waiting.
And step 3: establishing a base classifier, adjusting two-dimensional cyclic spectrum data, and connecting four heterogeneous convolutional neural networks (Efficientnets) into an integrated network by using a stacking model, wherein the method specifically comprises the following steps:
step 3.1: the MBConv network structure is used as a backbone network, four heterogeneous convolutional neural networks are established in an Efficientnets mode, the four heterogeneous convolutional neural network structures are determined by an Efficientnets algorithm, and the initial weight of the network is initialized randomly.
Step 3.2: as described in step 2, the acquired signal is preprocessed into a large image with 300 pixels each in length and width, but the base classifier training is slow due to too large image size, so that the two-dimensional cyclic spectrum image needs to be scaled and clipped to a size suitable for the base classifier. Because a large number of images with repeated shapes exist in the two-dimensional cyclic spectrum picture, when the two-dimensional cyclic spectrum picture is cut to a small size to be trained by a proper base classifier, each cutting result only contains information of one power amplifier signal period and does not contain repeated information. Considering that the size of the repeated image of different parameter signals in the two-dimensional cyclic spectrum picture is not constant, the specific cutting size needs to be determined according to the shape of the actual image in the two-dimensional cyclic spectrum picture. And adjusting the sizes of the input ends of the four heterogeneous convolutional neural networks according to the size of the cut two-dimensional cyclic spectrum.
Step 3.3: and (3) performing operations such as rotation, folding and the like on each two-dimensional cyclic spectrum picture cut in the previous step by using an image amplification method so as to increase the number of the two-dimensional cyclic spectrum pictures. In order to embody the distinction, a data set composed of a newly generated two-dimensional cyclic spectrogram is set as a 'new two-dimensional cyclic spectrum picture set', is set as a data set Y, and a data set composed of a previous two-dimensional cyclic spectrogram is set as an 'original two-dimensional cyclic spectrum picture set'. The operation of rotating the two-dimensional cyclic spectrum picture is specifically to rotate the picture by 180 degrees according to the picture center and forbid to rotate the rest angles. The specific implementation method of the two-dimensional cycle spectrogram slice folding operation is to rotate 180 degrees by taking a horizontal center line of a picture as an axis.
Step 3.4: four heterogeneous convolutional neural networks Efficientnets are connected into an integrated network by using a stacking model. The integrated network is divided into two layers of models, wherein the first layer of the integrated network is provided with a plurality of heterogeneous EffectintNet, each EffectintNet of the first layer of the integrated network is provided with a base classifier, the number of the base classifiers is n, and the base classifiers are respectively set to be M1,M2…Mn. The second layer model of the integrated network is a stacking model, and the output of the second layer model of the integrated network is prediction of the radiation source individuals.
And 4, step 4: segmenting the collected power amplifier signals, and training a base classifier and a stacking model of a second layer, wherein the training model specifically comprises the following steps:
step 4.1: the received power amplifier individual signal data set Y is divided.
Step 4.1.1: as shown in fig. 2, after obtaining the data needed by the training network, the data set Y obtained in step 3.3 is divided into two parts, the first part accounts for 20% of all the data and is used as the test set of the second layer model of the integrated network, and is set as Te2The method is used for testing the integral result of the patent algorithm after all models are trained; the second part accounts for 80% of all data, is used as a training set of the first layer model of the integrated network and is set as T1。
Step 4.1.2: data set T1Is divided into m parts, each is S1,S2…Sm. The size of m is not too large in consideration of the training cost of the model, but the purpose of fully utilizing data is achieved. The value of m is selected to be in the range of 3 to 5 in the general industry, and m is selected to be 4 in the patent. Therefore, the data set T1In this patentIn the table, the table is divided into 4 parts which are respectively set as S1,S2,S3,S4. Setting n different base classifiers of the first layer as M1,M2…Mn。
Step 4.2: training four base classifiers and a stacking model of the second layer.
Step 4.2.1: the operation of the training part is large in calculation amount, so that the operation is finished on a GPU with high calculation performance, and the trained network structure and the trained network weight model are stored in an FPGA for waiting use after the training is finished. This patent selects NVIDIA P4000GPU as the operation platform.
And training each base classifier by a gradient descent method according to the data segmentation result, wherein each base classifier obtains m weight models from m data segmentation results. Training a base classifier MnThen, 3 parts of the data are extracted from 4 parts of the data as a basis classifier MnOne of the redundant training sets is used as a base classifier MnThe test set of (1). Therefore, according to the extracted training sets of different combinations, the training weight results of 4 different network models of a certain base classifier can be obtained, and are respectively set as Mn1,Mn2…Mn4. The 4 different weight results of each base classifier respectively predict data which are selected as test sets before, and because the test sets of the 4 methods are not included when the data are segmented and the current prediction results are not included, the data can be spliced into a data set with the same size as the original data set again. The prediction results obtained by the 4 weights are spliced together as a new training set of the second layer and set as a new training set Tr2. New training set Tr2The training set is a training set of the second layer model and is used for training the stacking model of the second layer.
Step 4.2.2: the test set of the second layer model was 20% of the data Te previously divided from all the data2. First, Te2Firstly, the prediction of 4 trained network training results is needed, each result gives a prediction result, and Te is obtained in total24 kinds of prediction results. Then, averaging the four prediction results to obtain a test set Te 'of the second layer model'2And set as the new test set.
Step 4.2.3: using data sets Tr2Training an Integrated network second layer model, Using dataset Te'2And taking the prediction result X output by the second layer model as the final prediction result. And the second layer network is tested by using the new test set, which is equivalent to the test of the whole of the first layer and the second layer network. The use flow of the new test set is also the use flow of the predicted data in actual application.
And 5: and predicting the individual power amplifiers by using the trained integrated network.
The method can also improve the classification result continuously by changing the second layer model of the integrated network.
The method can also use a method of amplification during test (TTA) in network prediction, simultaneously send the picture to be predicted and the picture after the operation of rotation, folding and the like to the network prediction, and synthesize the prediction results of the two pictures, and select the result with high prediction probability. Meanwhile, cosine annealing attenuation can be used for training to replace a gradient descent method to further improve the identification capability of the network on the power amplifier individuals.
As shown in fig. 3, the present invention further provides a power amplifier individual identification device, which includes a signal receiving module, a digitizing module, a data processing module and a power supply, wherein the signal receiving module receives radio signals, and converts the radio signals after filtering, amplifying and orthogonalizing into two analog signals, which are sent to the digitizing module, the digitizing module performs analog-to-digital conversion on the two analog signals, and then stores the two analog signals in a cache unit, and then sends the two analog signals to the data processing module, the data processing module includes an FPGA chip for implementing the method of the present invention, the signals are processed by the FPGA chip and then output prediction results, and the power supply is connected to other modules to provide voltages required by the other modules during operation.
The signal receiving module comprises a signal receiving antenna, a band-pass filter, a low-noise power amplifier, frequency mixers, local oscillators, a 90-degree phase shifter and a low-pass filter, wherein the signal receiving antenna, the band-pass filter and the low-noise power amplifier are sequentially connected, the low-noise power amplifier is connected with the two frequency mixers, the local oscillators are also connected with the two frequency mixers, the 90-degree phase shifter is added between the local oscillators and the frequency mixers, and two paths of I/Q analog signals are output by the two frequency mixers and are output to the digitizing module through the low-pass filter.
The digitizing module comprises an analog-to-digital converter and a cache unit, wherein external input signals are respectively accessed into the analog-to-digital converter to obtain two paths of orthogonal digital signals, the two paths of orthogonal digital signals are sequentially stored in the cache unit in a pair-wise manner, the two paths of orthogonal digital signals are represented in a complex form in figure 4, and an imaginary part is represented by '-j'.
And the data processing module is an FPGA board card and is responsible for converting the received I/Q data into a two-dimensional spectrogram, cutting the spectrogram and then sending the cut spectrogram into a trained convolution network model to predict the individual category of the signal so as to obtain a final result. Before using the data processing module, the network model structure obtained by training and the model weight obtained by training need to be written into the FPGA chip, the running state of hardware is debugged, and the calculation precision and speed of the software part are adjusted, so that the overall performance is optimal.
The power supply uses the LTM4644 to form a single-channel output dc/dc power supply and is responsible for supplying power to the signal receiving module. Meanwhile, the LTM4620 is used for supplying power to the data processing module where the FPGA board card is located. Both the LTM4644 and LTM4620 support input voltages in the range of 5V to 14V.
The analog-to-digital converter adopts a TI AMC7820 chip, the chip integrates an 8-channel 12-bit analog-to-digital converter, three 12-bit digital-to-analog converters and 9 operational amplifiers, the communication mode is serial port communication, the application range of input voltage is wide, and the analog-to-digital converter has the advantages of low power consumption, simple design and the like.
The cache unit uses a 16GB MLC flash memory device and is used for temporarily storing the acquired digital signals in the I/Q format, waiting for the FPGA board card to read and perform subsequent processing.
The FPGA board card adopts an Ultra96-V2FPGAFPGA board card, and the type of the processor is MPSoC ZU3EG SBVA 484. The Ultra96-V2FPGAFPGA board is an FPGA board developed by Xilinx corporation and specially used for developing artificial intelligence and machine learning equipment, and has the advantages of easiness in operation, high speed and the like. The method is suitable for being used in the FPGA board card environment by using the algorithm based on the convolutional neural network. The FPGA board card is provided with serial ports such as I2C and the like, so that the chip can conveniently receive data of the analog-to-digital converter and output a prediction result of an algorithm. Meanwhile, the FPGA board card supports a 16GB MicroSD memory which is responsible for storing an operation program and a trained neural network structure and weight.
Because the neural network needs strong computing power during training, the FPGA chip is difficult to work in the part, and the amount of data needed during training the neural network is huge, so the FPGA is not suitable for storing a large amount of data generally. The GPU is used to train the neural network. And when the network training is finished, the complete convolution network structure and weight information are obtained, and the convolution network structure and weight information stored in the Python environment are converted into a protobuf file by using a conversion tool of the Xilinx FPGA board card, so that the FPGA can conveniently and quickly deduce. And then storing the file in a MicroSD storage device, and waiting for the FPGA chip to predict the individual result of the radiation source.
In order to verify the technical effect of the method, a Matlab2017a, a deep learning Pythrch frame and a sklernk library are used for building a model provided by the method, and a simulation environment is built by using an NVIDIA P4000 GPU.
The method comprises the following steps:
Step 2, signal source emission signal parameter setting: the carrier rate of the AM narrowband signal with the same type of power amplifier having the characteristic of spurious noise is 200kHz, and the code rate is 10 kHz. To mimic the noise in the channel, white gaussian noise of-5 dB to 15dB is randomly added to the signal.
Step 3, selecting signal receiving equipment: the device for receiving the fingerprint noise of the power amplifier of the radiation source should minimize the influence of other factors on the extraction of the fingerprint characteristics of the power amplifier in the transmitter. Therefore, the type, the placement position, the placement angle and the like of the receiving antenna of the signal receiver should not be changed, and the band-pass filter, the mixer, the low-pass filter and the analog-to-digital converter of the signal receiver should be the same.
Step 4, placing a signal receiver and a signal transmitter as a radiation source: the positions of the two amplifiers need to be set in advance, and the corresponding power amplifiers are selected in turn according to the sequence.
Step 6, storing the acquired discrete signals and sending the discrete signals to FPGA equipment for pretreatment: the acquired discrete signals are I/Q sequence signals, and the sequence signals need to be converted into two-dimensional image signals. And the FPGA chip performs cyclic autocorrelation transformation on the I/Q sequence signal by utilizing a pre-processing software module which is burned in before to obtain a two-dimensional cyclic autocorrelation spectrogram. And the software module cuts the two-dimensional circular autocorrelation spectrogram and outputs a cutting result to a large-capacity storage device for storage.
And 7, amplifying the data image: according to the program written in advance, the FPGA takes out the data stored in the large-capacity storage device and performs rotation and mirror image turnover transformation on the data. And the converted picture and the original picture are mixed into a larger data set and stored in the storage device again.
Step 8, predicting the power amplifier individual by using a neural network: the trained neural network weight is preset in the FPGA, and the predicted value output by the neural network can be obtained only by inputting the individual image to be predicted into the neural network. The image obtained by the amplification method in the step 7 can be used in TTA during neural network prediction, and the neural network can simultaneously predict the original image and TTA transformation thereof, so that the accuracy of neural network prediction can be improved.
The EfficientNets-based classifier used by the method has the advantages that the floating point operand per second is less than one hundred and eighty million times per second, the method belongs to a structure with a small network scale, and the method can have high accuracy in the aspect of power method individual identification. The method can effectively extract fine features in the output signal of the power amplifier by utilizing the two-dimensional spectrogram of the cyclic autocorrelation and the Efficientnets. Therefore, the method is suitable for being transplanted to embedded hardware equipment such as FPGA and the like for use due to the fact that the parameter number and the floating point operation amount per second are small.
The method can also continuously improve the overall classification result of the model by changing the second-layer model, namely adjusting the type of the second-layer classifier in the stacking method. Specific results are shown in table 1:
TABLE 1
The method of the invention may also improve the accuracy and stability of the model using data amplification and amplification-as-tested (TTA), as described in steps 7 and 8 above. In addition, the identification capability of the network to the power amplifier individual can be further improved by using a cosine annealing attenuation method during model training. Table 2 shows the overall accuracy of the model when other optimization methods are continuously added, wherein the brackets behind the accuracy represent the improvement of the network accuracy when the method is used.
TABLE 2
Experimental results show that the method can further improve the classification capability of the heterogeneous convolutional neural network by using methods such as data amplification and the like.
The method has the advantages that the average accuracy of the model in the range from-5 dB to 15dB reaches 93.08%, and the anti-interference capability is good. Meanwhile, a network with more parameter quantity and higher accuracy can be selected, and the classification capability of the whole model on the individual characteristics of the power amplifier can be further improved under the conditions of sacrificing hardware resources and reducing the prediction speed.
As shown in FIG. 5a, the curves in the graph are EfficientNet-B0 represented by model1, EffectientNet-B3 represented by model4, the variation modal decomposition method represented by VMD and the prediction accuracy curve generated by the method of the present invention, the abscissa represents the signal-to-noise ratio of-5 to 15dB, and the ordinate represents the accuracy. It can be seen from fig. 5a that the performance of the method based on the variational modal decomposition is lower than that of the method of the present invention and there is almost no prediction capability in the low snr region, especially the performance of the algorithm is rapidly reduced around 2 dB. The basic network used by the method can keep higher accuracy and still has more than 70% of accuracy in a low signal-to-noise ratio interval. FIG. 5b is a partial enlarged view of FIG. 5a, and it can be seen from FIG. 5b that the accuracy of the method of the present invention is not lower than that of the base classifier under all SNR conditions, and the accuracy is significantly improved compared to the base classifier under some SNR conditions. The method can be used for deducing that the prediction result of the network is more stable, and the prediction performance is improved to a certain extent.
The method can be applied to the algorithm part of the individual identification module of intelligent communication countermeasure equipment and electromagnetic spectrum monitoring equipment. Stray noise generated by power amplifiers in radiation sources such as radio stations and radars in the same batch can be accurately identified by receiving steady-state signals of different radiation sources of a non-partner, so that the specific direction of a certain radio station of the non-partner is positioned, the motion track of the certain radio station is tracked, and the method is a prepositive means for analyzing networking of non-partner communication equipment on a battlefield.
The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are only exemplary embodiments of the present invention, and are not intended to limit the present invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (10)
1. A power amplifier individual identification method, comprising:
step 1: collecting a power amplifier signal;
step 2: performing cyclic autocorrelation preprocessing on the acquired signal data of the power amplifier to obtain a two-dimensional cyclic spectrum;
and step 3: establishing a base classifier, adjusting two-dimensional cyclic spectrum data, and connecting four heterogeneous convolutional neural networks (Efficientnets) into an integrated network by using a stacking model;
and 4, step 4: segmenting the collected power amplifier signals, and training a base classifier and a stacking model of a second layer;
and 5: and predicting the individual power amplifiers by using the trained integrated network.
2. A power amplifier individual identification method as claimed in claim 1, characterized in that:
in the step 1, a receiver is used for receiving and converting output signals of a plurality of different power amplifiers into baseband signals, the baseband signals are I/Q two-path analog signals, and then the I/Q two-path analog signals are converted into two-path digital signals through analog-to-digital conversion and stored as power amplifier signal data in a dat format;
step 2, slicing the power amplifier signal data according to every 800 sampling points to obtain sliced data, then performing cyclic autocorrelation on the sliced data to obtain a two-dimensional cyclic spectrum, and storing the generated two-dimensional cyclic spectrum in a storage device;
the step 3 specifically comprises the following steps:
step 3.1: using an MBConv network structure as a backbone network, and establishing four heterogeneous convolutional neural networks in an Efficientnets manner;
step 3.2: scaling and cutting the two-dimensional cyclic spectrum slice to enable the two-dimensional cyclic spectrum slice to only contain information of a power amplifier signal period, and adjusting the size of the input ends of the four heterogeneous convolutional neural networks according to the size of the cut two-dimensional cyclic spectrum slice;
step 3.3: rotating and folding each two-dimensional cyclic spectrum picture cut in the previous step by using an image amplification method, and increasing the number of the two-dimensional cyclic spectrum pictures to obtain a new two-dimensional cyclic spectrum picture set;
step 3.4: connecting four heterogeneous convolutional neural networks (Efficientnets) into an integrated network by using a stacking model, wherein the integrated network is divided into two layers of models, the first layer of model of the integrated network is a plurality of heterogeneous Efficientnets, each EfficientNet of the first layer is a base classifier, and the second layer of model of the integrated network is the stacking model;
the step 4 specifically comprises the following steps:
step 4.1: a new two-dimensional cyclic spectrum picture set is obtained in the segmentation step 3.3;
step 4.2: training four base classifiers and a stacking model of the second layer.
3. A power amplifier individual identification method as claimed in claim 2, characterized in that:
storing the collected power amplifier signal data into a server hard disk in the step 1;
in step 2, the picture size of the two-dimensional cyclic spectrum is 300 pixels each of length and width, and the two-dimensional cyclic spectrum is obtained by adopting the following formula:
where R is the result of a two-dimensional cyclic spectral transformation of the signal, τ is the time delay, T is the time duration of the signal being processed, x (T) is the signal being processed, and α is the cyclic frequency that can be passedCalculated as m is a constant, T0Is the period of the cyclostationary signal;
in the step 3.1, the four heterogeneous convolutional neural network structures are determined by an Efficientnets algorithm, and the initial weight of the network is initialized randomly;
the operation of rotating the two-dimensional cyclic spectrum picture in the step 3.3 is to rotate the picture by 180 degrees according to the picture center, and the specific implementation method of the picture folding operation is to rotate by 180 degrees by taking the horizontal center line of the picture as an axis;
the number of base classifiers in step 3.4 is nThe base classifiers are respectively set to M1,M2…MnThe output of the second layer model is the prediction of the individual radiation source
Step 4.1, the data set Y obtained in the step 3.3 is divided into two parts, the first part accounts for 20% of all data and is used as a test set of a second layer model of the integrated network, and Te is set2(ii) a The second part accounts for 80% of all data, is used as a training set of the first layer model of the integrated network and is set as T1(ii) a Data set T1The average is divided into 4 parts, each is S1,S2,S3,S4(ii) a Setting n different base classifiers of the first layer as M1,M2…Mn;
Step 4.2, selecting the GPU as an operation platform, training each base classifier by a gradient descent method according to data segmentation results, obtaining M weight models by each base classifier according to M data segmentation results, and training a certain base classifier MnThen, 3 parts of the data are extracted from 4 parts of the data as a basis classifier MnOne of the redundant training sets is used as a base classifier MnThe test set of (2) obtains the training results of 4 different network models of a certain base classifier, which are respectively set as Mn1,Mn2…Mn4The 4 weighted predictors are grouped together as a new training set for the second layer and set as a new training set Tr2The test set of the second layer model is 20% of data Te divided from all data2,Te2Te is obtained by predicting the trained 4 network training results2The four prediction results are averaged to obtain a test set Te 'of the second layer model'2And set as a new test set; using data sets Tr2Training an Integrated network second layer model, Using dataset Te'2And taking the prediction result X output by the second layer model as the final prediction result.
4. A power amplifier individual identification method as claimed in claim 3, characterized in that:
the second layer model of the integrated network can also be of the type XGBoost or LightGBM.
5. A power amplifier individual identification method as claimed in claim 3, characterized in that:
when the network prediction is carried out, a method of using Test Time Amplification (TTA) is adopted, the picture to be predicted and the picture after the operation of rotating, folding and the like are simultaneously sent to the network prediction, and the prediction results of the two pictures are synthesized.
6. A power amplifier individual identification method as claimed in claim 3, characterized in that:
cosine annealing attenuation is used in the step 4.2 to replace a gradient descent method.
7. A power amplifier individual identification device comprises a signal receiving module, a digitizing module, a data processing module and a power supply, and is characterized in that: the signal receiving module receives a radio signal, and converts the radio signal into two paths of analog signals after filtering, amplifying and orthogonalizing the radio signal and sends the two paths of analog signals to the digitizing module, the digitizing module performs analog-to-digital conversion on the two paths of analog signals and then stores the two paths of analog signals to the cache unit and then to the data processing module, the data processing module comprises an FPGA chip for realizing the method, the signals are processed by the FPGA chip and then output a prediction result, and the power supply is connected with other modules and provides voltages required by other modules during working.
8. A power amplifier individual identification apparatus as claimed in claim 7, wherein:
the signal receiving module comprises a signal receiving antenna, a band-pass filter, a low-noise power amplifier, frequency mixers, local oscillators, a 90-degree phase shifter and a low-pass filter, wherein the signal receiving antenna, the band-pass filter and the low-noise power amplifier are sequentially connected, the low-noise power amplifier is connected with the two frequency mixers, the local oscillators are also connected with the two frequency mixers, the 90-degree phase shifter is added between the local oscillators and the frequency mixers, two paths of I/Q analog signals are output by the two frequency mixers, and the two paths of I/Q analog signals are output to the digitizing module through the low-pass;
the digital module comprises an analog-to-digital converter and a cache unit, wherein external input signals are respectively accessed into the analog-to-digital converter to obtain two paths of orthogonal digital signals, and the two paths of orthogonal digital signals are sequentially stored in the cache unit in a paired mode;
the data processing module is an FPGA board card and is responsible for converting the received I/Q data into a two-dimensional spectrogram, cutting the spectrogram and then sending the cut spectrogram into a trained convolution network model to predict the individual category of the signal to obtain a final result;
the power supply forms a single-channel output dc/dc power supply by using the LTM4644 to supply power to the signal receiving module; the LTM4620 is used for supplying power to the data processing module where the FPGA board card is located.
9. A power amplifier individual identification apparatus as claimed in claim 8, wherein: the analog-to-digital converter adopts an AMC7820 chip of TI, a cache unit adopts a 16GB MLC flash memory device, and an FPGA board card adopts an Ultra96-V2FPGAFPGA board card.
10. A detection method of an individual identification method of a power amplifier comprises the following steps:
step 1: selecting power amplifiers of two or more signal sources as an individual radiation source to be identified;
step 2: setting parameters of a signal source for transmitting signals;
and step 3: selecting a signal receiving apparatus;
and 4, step 4: placing a signal receiver and a signal transmitter;
and 5: receiver sampling and signal conversion;
step 6: storing the acquired discrete signals and sending the discrete signals to FPGA equipment for pretreatment;
and 7: amplifying a data image;
and 8: the individual power amplifiers are predicted using a neural network.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911263371.8A CN110991378B (en) | 2019-12-11 | 2019-12-11 | Individual identification method and device for power amplifier |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911263371.8A CN110991378B (en) | 2019-12-11 | 2019-12-11 | Individual identification method and device for power amplifier |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110991378A true CN110991378A (en) | 2020-04-10 |
CN110991378B CN110991378B (en) | 2023-06-02 |
Family
ID=70092201
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911263371.8A Active CN110991378B (en) | 2019-12-11 | 2019-12-11 | Individual identification method and device for power amplifier |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110991378B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112183279A (en) * | 2020-09-21 | 2021-01-05 | 中国人民解放军国防科技大学 | Communication radiation source individual identification method based on IQ graph characteristics |
CN115186714A (en) * | 2022-07-15 | 2022-10-14 | 中国人民解放军国防科技大学 | Network card frequency spectrum fingerprint feature amplification method based on feature correlation and self-adaptive decomposition |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107301381A (en) * | 2017-06-01 | 2017-10-27 | 西安电子科技大学昆山创新研究院 | Recognition Method of Radar Emitters based on deep learning and multi-task learning strategy |
US20180357542A1 (en) * | 2018-06-08 | 2018-12-13 | University Of Electronic Science And Technology Of China | 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method |
CN109684995A (en) * | 2018-12-22 | 2019-04-26 | 中国人民解放军战略支援部队信息工程大学 | Specific Emitter Identification method and device based on depth residual error network |
-
2019
- 2019-12-11 CN CN201911263371.8A patent/CN110991378B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107301381A (en) * | 2017-06-01 | 2017-10-27 | 西安电子科技大学昆山创新研究院 | Recognition Method of Radar Emitters based on deep learning and multi-task learning strategy |
US20180357542A1 (en) * | 2018-06-08 | 2018-12-13 | University Of Electronic Science And Technology Of China | 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method |
CN109684995A (en) * | 2018-12-22 | 2019-04-26 | 中国人民解放军战略支援部队信息工程大学 | Specific Emitter Identification method and device based on depth residual error network |
Non-Patent Citations (2)
Title |
---|
余沁;程伟;杨瑞娟;: "基于排列熵分形维数特征提取的通信辐射源个体识别" * |
耿梦婕;张君毅;: "基于神经网络的辐射源个体识别技术" * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112183279A (en) * | 2020-09-21 | 2021-01-05 | 中国人民解放军国防科技大学 | Communication radiation source individual identification method based on IQ graph characteristics |
CN112183279B (en) * | 2020-09-21 | 2022-06-10 | 中国人民解放军国防科技大学 | Communication radiation source individual identification method based on IQ graph characteristics |
CN115186714A (en) * | 2022-07-15 | 2022-10-14 | 中国人民解放军国防科技大学 | Network card frequency spectrum fingerprint feature amplification method based on feature correlation and self-adaptive decomposition |
CN115186714B (en) * | 2022-07-15 | 2023-07-21 | 中国人民解放军国防科技大学 | Network card frequency spectrum fingerprint feature amplification method based on feature correlation and self-adaptive decomposition |
Also Published As
Publication number | Publication date |
---|---|
CN110991378B (en) | 2023-06-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109344772B (en) | Ultrashort wave specific signal reconnaissance method based on spectrogram and deep convolutional network | |
CN107220606B (en) | Radar radiation source signal identification method based on one-dimensional convolutional neural network | |
CN109307862A (en) | A kind of target radiation source individual discrimination method | |
US20200143279A1 (en) | Radio frequency band segmentation, signal detection and labelling using machine learning | |
CN111541511B (en) | Communication interference signal identification method based on target detection in complex electromagnetic environment | |
CN107729916B (en) | ISODATA-based interference source classification and identification algorithm | |
CN109544536A (en) | The fast automatic analysis method of hip joint x-ray image | |
CN114154545B (en) | Intelligent unmanned aerial vehicle measurement and control signal identification method under strong mutual interference condition | |
CN110991378A (en) | Power amplifier individual identification method and device | |
Flak | Drone detection sensor with continuous 2.4 GHz ISM band coverage based on cost-effective SDR platform | |
CN111310700A (en) | Intermediate frequency sampling sequence processing method for radiation source fingerprint feature identification | |
CN116047427B (en) | Small sample radar active interference identification method | |
US20220077941A1 (en) | Frequency spectrum analyzers and devices, systems, software and methods for signal power measurement and spectrum analysis | |
CN109874104A (en) | User location localization method, device, equipment and medium | |
CN111046025B (en) | Unmanned aerial vehicle signal detection method and device | |
CN116482720A (en) | GNSS deception jamming detection method based on machine learning theory | |
CN115668786A (en) | Adaptive radio configuration in a wireless network | |
Limin et al. | Low probability of intercept radar signal recognition based on the improved AlexNet model | |
CN115600101A (en) | Unmanned aerial vehicle signal intelligent detection method and device based on priori knowledge | |
CN103607249A (en) | Method and system for testing DS/FH (Direct Sequence/Frequency Hopping) mixed spread frequency signal parameters | |
CN116318249A (en) | Short wave frequency hopping signal sorting method based on improved YOLOv5 | |
CN115345216A (en) | FMCW radar interference elimination method fusing prior information | |
CN110995365B (en) | Radio frequency gene library establishing method and illegal radio wave detection method | |
CN114492210A (en) | Hyperspectral satellite borne data intelligent interpretation system and implementation method thereof | |
Yip et al. | Efficient and effective tropical cyclone eye fix using genetic algorithms |
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 |