CN115296759A - Interference identification method based on deep learning - Google Patents
Interference identification method based on deep learning Download PDFInfo
- Publication number
- CN115296759A CN115296759A CN202210831105.6A CN202210831105A CN115296759A CN 115296759 A CN115296759 A CN 115296759A CN 202210831105 A CN202210831105 A CN 202210831105A CN 115296759 A CN115296759 A CN 115296759A
- Authority
- CN
- China
- Prior art keywords
- interference
- information
- network
- layer
- block
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 238000013135 deep learning Methods 0.000 title claims abstract description 12
- 230000002776 aggregation Effects 0.000 claims abstract description 24
- 238000004220 aggregation Methods 0.000 claims abstract description 24
- 230000004927 fusion Effects 0.000 claims abstract description 24
- 238000004364 calculation method Methods 0.000 claims abstract description 17
- 238000012549 training Methods 0.000 claims description 30
- 238000000605 extraction Methods 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 17
- 238000011176 pooling Methods 0.000 claims description 15
- 238000010586 diagram Methods 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 10
- 230000003993 interaction Effects 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 238000005192 partition Methods 0.000 claims description 4
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 230000006835 compression Effects 0.000 claims description 3
- 238000007906 compression Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 230000001737 promoting effect Effects 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims description 3
- 230000011218 segmentation Effects 0.000 abstract description 6
- 239000012634 fragment Substances 0.000 abstract 1
- 238000004891 communication Methods 0.000 description 18
- 238000013527 convolutional neural network Methods 0.000 description 18
- 238000004088 simulation Methods 0.000 description 9
- 238000005516 engineering process Methods 0.000 description 7
- 238000007476 Maximum Likelihood Methods 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 241000209094 Oryza Species 0.000 description 2
- 235000007164 Oryza sativa Nutrition 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 235000009566 rice Nutrition 0.000 description 2
- 238000001228 spectrum Methods 0.000 description 2
- 230000001629 suppression Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000010363 phase shift Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000008054 signal transmission Effects 0.000 description 1
Images
Classifications
-
- 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
- H04B17/345—Interference values
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Electromagnetism (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)
Abstract
The invention discloses an interference identification method based on deep learning. The main principle of the method is as follows: 1) The local attention fusion layer is adopted, a plurality of input segmentation information blocks with the same quantity are divided into different regions, global features in the regions are independently calculated in each region, the quantity of the segmentation information blocks in each region is far smaller than that of all the segmentation information blocks, at the moment, the calculation quantity of the global features independently calculated in all the regions is far smaller than that of all the input information blocks simultaneously calculated by multi-head self-attention, 2) a block aggregation module is introduced and alternately appears with the local attention fusion layer, the module can convert a plurality of input information blocks into one information block and aggregate the information of the information blocks, so that output information fragments are gradually reduced along with the increase of the number of layers.
Description
Technical Field
The invention belongs to the technical field of communication, and particularly relates to a method for identifying interference signals in a wireless communication anti-interference system.
Background
With the coming of modern information age, the wireless communication technology is continuously developed, and the wireless communication is widely applied to various fields, but in the process of wireless signal transmission, a wireless communication system is susceptible to various interferences, and has system internal interferences caused by system characteristics and system external interferences caused by artificial malicious damage, wherein the external interferences with attack or damage intentions can greatly reduce the safety and reliability of the communication system, so that the improvement of the anti-interference capability of the wireless communication system is an important guarantee of the wireless communication and also an internal requirement of the quality and efficiency of the modern wireless communication; in a complex electromagnetic environment, in order to reduce the risk of influencing communication and the probability of being intercepted by interference, communication anti-interference technologies are developed, and common communication anti-interference technologies comprise interference avoidance, interference suppression, interference cancellation and the like; interference is first detected and identified before Interference avoidance, interference suppression and Interference cancellation, and thus, radio Interference identification (Wireless Interference) is performedWII) is a precondition and a basis for interference resistance of wireless communication, the technology aims to identify the type of a received interference signal without any prior information, thereby providing information for technologies such as interference parameter estimation, interference avoidance and the like, and is widely applied in military wireless communication, for example, a wireless interference identification technology is used for discovering and identifying the interference signal from an enemy; in addition, in civil wireless communication, the wireless interference identification technology can improve transmission efficiency by avoiding mutual interference among users; therefore, the wireless interference identification is more suitable for military field and civil fieldApplications such as cognitive radio, spectrum management, secure communications, communications electronics warfare, and the like.
The currently disclosed wireless interference identification method mainly comprises the following steps:
1) The interference identification method based on Maximum Likelihood (ML): the method constructs test statistics according to Bayesian minimum misjudgment criteria, and obtains optimal performance by comparing the test statistics with an optimal decision threshold, however, the ML-based method requires sufficient knowledge of wireless channel information, which is unavailable in most communication scenarios, and in addition, the method has high computational complexity and is difficult to deploy on resource-limited equipment;
2) An interference identification method based on Feature Extraction (FE) comprises the following steps: the method identifies the signal type by extracting features, firstly, the characteristics of interference signals are extracted in a feature extraction stage, then, a classifier is applied to classify according to the signal characteristics, the extracted features comprise high-order statistical characteristics, amplitude, a cyclic spectrum and the like, the feature extraction method is widely used in the early stage due to low calculation complexity, however, the feature extraction method needs expert knowledge and manually-made feature engineering, and the method adopts a method of respectively optimizing the feature extraction and the classifier, so that the globally optimal identification result cannot be achieved;
3) An interference identification method based on Convolutional Neural Networks (CNN): in recent years, deep Learning (DL) has made remarkable progress in many fields, and due to excellent data processing capability, a wireless interference identification method based on Deep Learning has made performance superior to a feature extraction method, wherein the interference identification method of a Convolutional Neural Network (CNN) is widely applied, and an end-to-end optimization method is adopted, so that an automatic feature extraction part and a classifier can be jointly optimized, and the interference identification performance is remarkably improved; however, because the CNN in the scheme adopts the sliding trainable convolution kernel to automatically extract features, the CNN has inherent limitations of focusing on extracting local features, cannot comprehensively utilize global features, causes the network to continuously stack more convolution layers to improve performance, and obviously increases the amount of calculation;
4) An interference identification method based on a converter network (Transformer) comprises the following steps: compared with a CNN (compressed natural network), the Transformer network has the capability of extracting global features and obtains remarkable performance in the aspect of wireless interference identification, and the identification performance can reach or even exceed that of a CNN-based method, however, a multi-head self-attention module in the Transformer network needs to calculate the correlation between each divided information block and other information blocks in input, the calculation amount is in direct proportion to the square of the number of the divided information blocks, and therefore, the network has huge calculation overhead, and the network is prevented from being deployed on resource-limited equipment.
By combining the existing wireless interference identification method, the ML-based method needs channel prior information and is difficult to be practical; compared with the traditional method based on feature extraction, the existing Interference identification method based on deep learning can effectively improve the identification performance, for example, a CNN (convolutional neural network) and a Transformer network update trainable parameters through back propagation, end-to-end network optimization is realized, and the Interference identification performance at low dry-to-noise ratio (INR) is obviously superior to that of the feature extraction method; however, the CNN method cannot comprehensively utilize global characteristics, and the identification performance still has certain limitation; in addition, the multilayer network structure in the CNN and Transformer network brings huge calculation cost, and the calculation complexity is high; therefore, the existing interference identification method based on deep learning has a further space for improvement in interference identification performance and complexity.
Disclosure of Invention
In order to solve the defects of the prior art, the invention provides a Wireless Interference identification converter network (Wireless Interference) basedTransformer, WII-Transformer), the method not only obviously improves the recognition performance compared with the existing interference recognition method based on deep learning, but also can reduce the calculation cost requirement, and the main principle is as follows: 1) Local-attention Fusion Layers (LAFL) are usedConsists of a Region-based Multi-head Self-Attention (RMSA) and an Information Exchange Module (IEM); the multi-head self-attention based on the region improves a multi-head self-attention module in a traditional transform network, divides a plurality of input segmentation information blocks with the same quantity into different regions, and independently calculates global characteristics in each region, because the quantity of the segmentation information blocks in each region is far less than that of all the segmentation information blocks, the calculation quantity of the global characteristics in all the regions is far less than that of all the input information blocks calculated by the multi-head self-attention at the same time, and the calculation quantity is not directly proportional to the square of the number of the information blocks any more, but is reduced to be directly proportional to the number of the information blocks; however, after multi-head self-attention processing based on regions, each region cannot interact with information of other regions, so that after each region is independently processed, an information block is randomly selected from each region and input to an information exchange module, the information block information from different regions is fused by the module, and the fused information is transmitted to each region, so that information interaction between the regions is promoted; the complexity is effectively reduced, meanwhile, the information transmission and interaction are guaranteed, and each region is easier to extract effective features due to the fact that the number of information blocks is small, and the identification performance is further improved; 2) In order to further reduce the calculation amount, the invention introduces a Patch Aggregation Module (PAM) which alternately appears with the local attention fusion layer, the Module can convert several input information blocks into one information block and aggregate the information of the information blocks, so that the output information segments are gradually reduced along with the increase of the layer number, namely the information segments input into the local attention fusion layer are gradually reduced, which can gradually reduce the calculation amount close to the final output layer.
The interference identification method based on WII-Transformer is given below.
The WII-Transformer method comprises the following specific steps in a training phase:
s1, constructing training dataWherein x i ∈C M×1 For the sampling complex signal of the receiving end, the ith sample is represented, M is the number of sampling points, y i E is 1,2, wherein E represents a label corresponding to the ith sample, E represents the type number of interference, and the number of samples is N in total;
extracting a time-frequency diagram of the time-frequency distribution of each sample through short-time Fourier transform to obtainWherein W 0 ×H 0 Representing resolution, C in Is the number of channels;
s2, constructing a wireless interference identification network, wherein the wireless interference identification network comprises a block embedding layer, a feature extraction module, a global average pooling layer and a full connection layer, and the feature extraction module is formed by alternately connecting a plurality of local attention fusion layers and a plurality of block aggregation modules; the specific processing mode of the wireless interference identification network on the samples is as follows:
for input sample X 0 By partitioning the block-embedded layer, X 0 The uniform partition is converted into T information blocks X with C channels, namely X belongs to R T×C Wherein T = W 0 ×H 0 /p 2 ,C=C in ×p 2 P × p is the block size; x belongs to R T×C Each information block in (a) is subjected to an embedding layer to obtain a corresponding embedded signature, i.e. O c =f c (X,θ c ) WhereinIs represented by having C 0 An embedded feature of the channel, f c And theta c Representing embedding functions and their parameters, respectively, embedding layer f c Adopting a convolution layer operation;
o obtained by embedding blocks into a layer c Performing region division into T/r 2 A non-overlapping region is input into the feature extraction module, where r 2 Indicating the size of the area, i.e. output conversionPerforming local attention fusion operation on all regions through a local attention fusion layer, namely performing region-multiple-head-based self-attention operation extraction on global features O 'in the regions in each region, promoting information interaction between the regions through an information exchange module, wherein the output of the information exchange module is O', splicing information blocks of all the regions together through a block aggregation module, and fusing information through projection, wherein the information is represented as O p =f pa (reshape (o ")), where f pa Represents the projection function, adopts the deep convolution as the operation of the block aggregation module, reshape (-) represents the stitching operation, O' represents the output of the information exchange module, O p Representing the output of the block aggregation module;
after training samples pass through all local attention fusion layers and a block aggregation module in a feature extraction module, the feature extraction module obtains extracted features, the extracted features are input into a global average pooling layer and a full connection layer, and the dimensionality of the features is obtained through global average pooling compression and is a one-dimensional vector z = [ z ] z = 1 ,z 2 ,...,,z U ]U represents the dimension of the features obtained after global average pooling, and the network output p (x) is obtained by adopting a full connection layer and a softmax function i ) Is the ith sample x i The number of neurons in the fully connected layer is E, i.e., the number of types of interference;
s3, training the wireless interference recognition network constructed in the S2 by adopting the training data constructed in the S1, setting training cycle times K, setting the batch size as B, and defining a loss function as follows:
wherein q (x) i ) And p (x) i ) For the ith sample x i By reverse direction and predicted distributionPropagating updated quantized network parameters, i.e.Wherein theta is all trainable parameters of the network, and eta is the learning rate;
stopping training to obtain a trained parameter theta when the training cycle number is reached, so as to obtain a trained wireless interference identification network;
s4, interference identification is carried out: after the quasi-complex baseband interference signal obtained by the interference detection equipment is sampled r (n), the quasi-complex baseband interference signal is sent to a trained wireless interference identification network to obtain an identification result output by the network, namely the identification result isWherein P (-) represents a one-dimensional probability vector derived from the network output, e * Indicating the result obtained by the recognition.
In step S2, the specific implementation method of the local attention fusion layer is as follows:
extracting global features O' within a region by regional multi-headed self-attention operations: the calculation mode is a multi-head self-attention mode, and the expression isThe method for calculating the query matrix Q, the key matrix K and the value matrix V is Q = O c W q 、K=Ο c W k 、V=Ο c W v Wherein W is q 、W k 、Is a projection matrix obtained by training, d represents the vector size, and the multi-head self-attention consists of h parallel self-attention structures, namely RMSA (Q, K, V) = concat (hd) 1 ,hd 2 ,...,hd h )W,hd h =Att(Q h ,K h ,V h ) Wherein W is a matrix obtained by training; meanwhile, the short connection operation introduced by the multi-head self-attention of the region increases the forward propagation path, that is, the original information block is added to the information block after the multi-head self-attention processing of the region, and the sum is O' = RMSA (O) c )+Ο c O' is the output of multi-head self-attention in the region;
after the blocks of each area are processed independently, one processed information block is randomly selected in each area and input to the information exchange module, the information exchange module is composed of one fully connected layer, and after the information exchange is performed through the operation, the information block is added to the rest of the information blocks of the corresponding area, where o "= FC (o ') + o', where o" denotes the output of the information exchange module.
The invention has the beneficial effects that:
the invention provides a WII-Transformer interference identification scheme, which adopts a network to introduce a region division idea, so that region characteristics are calculated by multi-head self-attention based on regions in each local region, the complexity is effectively reduced, and in order to overcome the bottleneck of information interaction between the regions, an information exchange module is adopted to promote the flow of region information; in order to further reduce complexity and extract local features, a new block aggregation module is introduced to fuse adjacent feature information; compared with the existing interference identification method based on deep learning, the method provided by the method further enhances the capability of extracting the strong features, and the identification performance is obviously improved when the calculation complexity is close; the method effectively reduces the deployment cost of the network in interference identification, and the calculation complexity is lower than that of the existing method; compared with the prior art, the interference identification optimization method has the advantages of obvious advantages, easy realization and strong application value
Drawings
FIG. 1 is a WII-Transformer interference identification network according to the present invention;
FIG. 2 is a flow chart for interference identification;
FIG. 3 shows an RMSA block in the WII-Transformer according to the present invention;
FIG. 4 illustrates a PAM module in the WII-Transformer of the present invention;
FIG. 5 illustrates an IEM module in the WII-Transformer of the present invention;
FIG. 6 is a comparison diagram of the performance of the identification accuracy simulation of various networks as INR changes;
FIG. 7 is a comparison diagram of simulation performance of identification accuracy of various interference signal types of the ResNeXt network along with the change of INR;
FIG. 8 is a comparison graph of simulation performance of identification accuracy as a function of INR for various interfering signal types of the SWIN network;
FIG. 9 is a comparison diagram of simulation performance of identification accuracy of various interference signal types of a BGCNN network along with the change of INR;
FIG. 10 is a comparison graph of simulation performance for the identification accuracy of various interference signal types of the WII-Transformer network as a function of INR;
FIG. 11 is a schematic diagram of a confusion matrix simulation at INR = -16dB for a WII-Transformer network;
FIG. 12 is a schematic diagram of a confusion matrix simulation for a WII-Transformer network at INR = -10 dB.
Detailed Description
The technical solution of the present invention will be described in detail below with reference to the embodiments and the accompanying drawings.
The invention considers the scene of air-to-ground interference, such as an interference signal released by an unmanned aerial vehicle, selects a rice single path as a wireless channel, sets the rice factor to be 10dB, and the pattern of the interference signal comprises 8 types: single-Tone interference (ST), binary Frequency Shift Keying (BFSK), binary Phase Shift Keying (BPSK), narrowband Noise (NBN), multi-Tone interference (MT), sinusoidal Frequency Modulation (SFM), partial Band Noise (PBN), linear Frequency Modulation (LFM); the center frequency of each interference sample ranges from-25 MHz to 25MHz,2FSK, BPSK, NBN with bandwidth from 1.5MHz to 5MHz, MT, SFM, PBN, LFM with bandwidth interference from 5MHz to 50MHz, SFM and LFM with period from 0.01ms to 1ms; the training data and the test data included 16000 samples and 1600 samples for each category, respectively; INR ranges from-20 dB to 10dB with a 2dB separation.
The structural parameters of the model are shown in table 1, and it can be seen from the table that the network comprises a block embedding layer, a local attention fusion layer, a block aggregation module, etc., in the invention, the embedding layer is composed of two 3 × 3 convolutional layers, has 32 channels and 2 steps, "RMSA (5,32) + IEM (32)" represents a local attention layer having 5 × 5 local region and 32 channels, and the block aggregation module layer is implemented by using a depth convolutional layer, wherein the depth convolutional layer (3 fusion 3,32, 1) represents a 3 × 3 depth convolutional layer, has 32 channels and 1 step, and the network is identified by using global average pooling and a full connection layer having softmax activation function.
TABLE 1 network architecture
Number of layers | Structure of the |
1 | |
2 | Local attention fusion layer 1: RMSA (5,32) + IEM (32) |
3 | Block aggregation module 1: deep convolution layer (3 x 3,32, 1) + deep convolution layer (3 x 3,64, 2) |
4 | Local attention fusion layer 2: RMSA (5,64) + IEM (64) |
5 | Block aggregation module 2: deep convolution layer (3 x 3,64, 1) |
6 | Global average pooling + full-link layer + softmax function |
Fig. 1 shows the general structure of the method, where interference signals are processed by a block embedding layer, and obtained features are divided into several regions, the network independently calculates global features of each region by using multi-head self-attention based region, and makes up for the difficulty in exchanging information between regions due to multi-head self-attention based region by using a block aggregation module, and in addition, the block aggregation module further reduces complexity and fuses adjacent local information, and finally, identifies interference by using a full connection layer of a softmax activation function; fig. 2 shows a flow of interference identification, which includes time-frequency transform processing and network processing, and finally obtains an output result.
With the structure of table 1, the following method specifically introduces interference identification based on WII-Transformer, which specifically comprises the following steps:
s1, constructing training dataWherein x is i ∈C M×1 For the sampling complex signal of the receiving end, the ith sample is represented, M is the number of sampling points, y i E1,2., E denotes the label corresponding to the i-th sample, E denotes the number of types of interference, the number of samples is N, in this example E =8, the training data and the test data respectively include 16000 samples and 1600 samples for each class, INR ranges from-20 dB to 10dB, and the interval is 2dB;
extracting a time-frequency distribution time-frequency graph of each sample through short-time Fourier transform to obtainWherein W 0 ×H 0 Represents resolution, C in Is the number of channels, in this example, X 0 ∈R 200×200×1 To represent the input, where 200 x 200 represents the size, and 1 represents a single-channel grayscale image, as the input to the network;
s2, constructing a wireless interference identification network, wherein the wireless interference identification network comprises a block embedding layer, a feature extraction module, a global average pooling layer and a full connection layer, and the feature extraction module is formed by alternately connecting a plurality of local attention fusion layers and a plurality of block aggregation modules; the specific processing mode of the wireless interference identification network on the sample is as follows:
for input sample X 0 The block is divided by the block embedding layer, the size of the block is defined as 1 × 1, i.e., p =1, and X is set as 0 The uniform partition is converted into T information blocks X with C channels, namely X belongs to R T×C Wherein T = W 0 ×H 0 /p 2 =200×200/1 2 =40000,C=C in ×p 2 =1,p × p is the block size; x is formed by R T×C Each information block in (a) is subjected to an embedding layer to obtain a corresponding embedded signature, i.e. O c =f c (X,θ c ) WhereinIs represented by having C 0 An embedded feature of the channel, f c And theta c Representing embedding functions and their parameters, respectively, embedding layer f c A convolutional layer operation is adopted, having two 3 x 3 convolutional layer compositions, having 32 channels and 2 steps;
o obtained by embedding blocks into a layer c Performing region division into T/r 2 A non-overlapping region is input into the feature extraction module, where r 2 =25 area size, i.e. output conversionPerforming local attention fusion operation on all regions through a local attention fusion layer, namely performing region-multiple-head-based self-attention operation extraction on global features O 'in the regions in each region, promoting information interaction between the regions through an information exchange module, wherein the output of the information exchange module is O', splicing information blocks of all the regions together through a block aggregation module, and fusing information through projection, wherein the information is represented as O p =f pa (reshape (O ")), where f pa Representing projection functions, using depth convolution as an operation of the block aggregation module, reshape (-) tableO' represents the output of the information exchange module, O p Representing the output of the block aggregation module;
after training samples pass through all local attention fusion layers and a block aggregation module in a feature extraction module, the feature extraction module obtains extracted features, the extracted features are input into a global average pooling layer and a full connection layer, and the dimensionality of the features is obtained through global average pooling compression and is a one-dimensional vector z = [ z ] z = 1 ,z 2 ,...,,z U ]U represents the dimension of the features obtained after global average pooling, and the network output p (x) is obtained by adopting a full connection layer and a softmax function i ) Is the ith sample x i The number of neurons of the fully connected layer is E, i.e. the number of types of interference;
s3, training the wireless interference recognition network constructed in the S2 by adopting the training data constructed in the S1, setting training cycle times K, setting the batch size as B, and defining a loss function as follows:
wherein q (x) i ) And p (x) i ) For the ith sample x i By updating the quantized network parameters by back-propagation, i.e. true label distribution and predicted distributionWherein theta is all trainable parameters of the network, and eta is the learning rate;
stopping training to obtain a trained parameter theta when the training cycle number is reached, so as to obtain a trained wireless interference identification network;
s4, interference identification is carried out: after the quasi-complex baseband interference signal obtained by the interference detection equipment is sampled r (n), the quasi-complex baseband interference signal is sent to a trained wireless interference identification network to obtain an identification result output by the network, namely the identification result isWhere P (-) represents the one-dimensional probability vector derived from the network output,e * indicating the result of the recognition.
The specific implementation method of the local attention fusion layer comprises the following steps:
as shown in fig. 3, extracting the global feature O' in the region by the region-multi-head self-attention operation: the calculation mode is a multi-head self-attention mode, and the expression isThe method for calculating the query matrix Q, the key matrix K and the value matrix V is Q = O c W q 、K=Ο c W k 、V=Ο c W v Wherein W is q 、W k 、Is a projection matrix obtained by training, d represents the vector size, and the multi-head self-attention consists of h =4 parallel self-attention structures, namely RMSA (Q, K, V) = concat (hd) 1 ,hd 2 ,...,hd h )W,hd h =Att(Q h ,K h ,V h ) Wherein W is a matrix obtained by training; meanwhile, the short connection operation introduced by the multi-head self-attention of the region increases the forward propagation path, that is, the original information block is added to the information block after the multi-head self-attention processing of the region, and the sum is O' = RMSA (O) c )+Ο c O' is the output of multi-head self-attention in the region;
as shown in fig. 5, after the blocks of each area are processed independently, one processed information block is randomly selected in each area and input to the information exchange module, which is composed of one fully connected layer, and after the information exchange operation, the information block is added to the rest of the information blocks of the corresponding area, where o "= FC (o ') + o', where o denotes the output of the information exchange module.
The computational complexity of MSA + MLP can be expressed as Ψ (MSA + MLP) =5TC 0 2 +2T 2 C 0 ∝T 2 Thus, it can be concluded that the computational complexity is proportional to the square of the number of blocks, when the input resolution is very highThe amount of computation increases greatly, which is unacceptable for devices with limited computational resources; the computational complexity of the LAFL can be expressed as Ψ (LAFL) = (4+1/r 2 )TC 0 2 +2r 2 TC 0 Is proportional to the number of blocks, making the complexity low; therefore, the number of partitions through the subsequent LAFA layer gradually decreases, which further reduces the computational complexity of the network; in addition, PAM uses a lightweight deep convolution operation, as shown in fig. 4, which can fuse several adjacent blocks and convert them into one block, so that the number of information blocks input to the LAFA module through the PAM network is gradually reduced, which further reduces the computational complexity.
TABLE 2 computational complexity of the various networks
The invention compares five traditional networks, including a convolution network (CNN), a residual error network (ResNet), a convergence residual error (ResNeXt) of a deep neural network, a layered vision converter (SWIN) of a shift window and a global convolution neural network (BGCNN), wherein the CNN is composed of two 3 x 3 convolution layers with 16 and 32 channels; resNet consists of a 3 x 3 convolutional layer and a residual building block, with dimensions of 32 and 37, respectively, where the residual block consists of convolutional and short connections; resNeXt is composed of a 3 x 3 convolutional layer with 32 channels and a ResNeXt block with 128 channels, wherein the ResNeXt block uses a block convolution mode on the basis of a residual block, the block convolution groups input characteristics on the basis of convolution operation, the convolution kernel also carries out corresponding grouping, and the convolution operation is independently carried out in each block, so that the computational complexity is reduced; at the end of these networks, global average pooling and full connectivity layers are used to convert features into recognition probabilities; the BGCNN network is formed by cascading a convolutional layer and a Transformer network to identify interference, and uses two 3 multiplied by 3 convolutional layers comprising 16 and 32 channels and two Transformer layers; in SWIN, it considers multi-headed self-attention under window partitioning, calculates multi-headed self-attention operations at each fixed window size, with block size and window size set to 20 and 5, dimensions and depth set to 104 and 2, and attention head numbers for both models set to 4; the batch size is set to 64, and the learning rate is initialized to 0.001; table 2 shows the Floating-point Operations Per Second (FLOPs) and trainable parameters of different models, and it can be seen from the table that the FLOPs of the method of the present invention is the lowest.
The recognition performance of 6 networks is compared in fig. 6, and it can be seen from the figure that the recognition performance of all models is gradually increased with the increase of INR, while the recognition performance of CNN is lower than that of other networks under different INRs, because the CNN network has a simple structure and poor recognition performance; the identification performance of the WII-Transformer reaches the best under different INRs, and when the INR is equal to-15 dB, the identification performance of the WII-Transformer is about 7 percent better than BGCNN, about 15 percent better than SWIN, about 17 percent better than ResNeXt and ResNet, and about 30 percent better than CNN; when INR is-10 dB, the identification performance of WII-Transformer is about 8% better than BGCNN, SWIN and ResNeXt and about 35% better than CNN, which proves the effectiveness of the invention; at an INR of about-5 dB, the recognition accuracy of the WII-transducer is close to 100%.
Fig. 7 reports the performance of the renex network for identifying different interference types with INR, and it can be seen from the graph that the scheme has poor performance for identifying NBN interference, at-10 dB, the type interference identification performance is about 60%, while at high INR, all PBN interference still does not reach 100% accuracy.
Figure 8 reports the performance of the SWIN network for the identification of different interference types with INR, and it can be seen from the figure that this scheme performs poorly for PBN and BPSK interference, with an interference identification performance of about 55% for this type at-10 dB, and at high INR, all interference identification rates approach 100%.
Fig. 9 reports the identification performance of the BGCNN network for different interference types with INR, and it can be seen from the figure that the scheme is poor for BPSK interference identification, at-10 dB, the type interference identification performance is about 40%, and at high INR, all interference identification rate is close to 100%.
Fig. 10 reports the performance of the WII-Transformer network for identifying different interference types along with INR, and it can be seen from the figure that the scheme has better performance for identifying various types of interference, at-10 dB, the performance of identifying all types of interference reaches more than 70%, and at high INR, the rate of identifying all types of interference approaches 100%.
Fig. 11 shows a simulation diagram of the confusion matrix of the WII-Transformer network when INR = -16dB, and it can be seen from the diagram that confusion occurs in different interference types, and confusion of single-tone interference is small, and the time-frequency diagram characteristics are different from other interferences greatly.
Fig. 12 shows a simulation diagram of the confusion matrix of the WII-Transformer network when INR = -10dB, and it can be seen that the confusion of different interference types is less, because the noise power is lower, and the identification of each type of interference becomes accurate.
Claims (2)
1. An interference identification method based on deep learning is characterized by comprising the following steps:
s1, constructing training dataWherein x i ∈C M×1 For the sampling complex signal of the receiving end, the ith sample is represented, M is the number of sampling points, y i E is 1,2, wherein E represents a label corresponding to the ith sample, E represents the type number of interference, and the number of samples is N in total;
extracting a time-frequency diagram of the time-frequency distribution of each sample through short-time Fourier transform to obtainWherein W 0 ×H 0 Representing resolution, C in Is the number of channels;
s2, constructing a wireless interference identification network, wherein the wireless interference identification network comprises a block embedding layer, a feature extraction module, a global average pooling layer and a full connection layer, and the feature extraction module is formed by alternately connecting a plurality of local attention fusion layers and a plurality of block aggregation modules; the specific processing mode of the wireless interference identification network on the sample is as follows:
for input sample X 0 By partitioning the block-embedded layer, X 0 The uniform partition is converted into T information blocks X with C channels, namely X belongs to R T×C Wherein T = W 0 ×H 0 /p 2 ,C=C in ×p 2 P × p is the block size; x is formed by R T×C Each information block in (a) is subjected to an embedding layer to obtain a corresponding embedded signature, i.e. O c =f c (X,θ c ) WhereinIs represented by having C 0 An embedded feature of the channel, f c And theta c Representing embedding functions and their parameters, respectively, embedding layer f c Adopting a convolution layer operation;
o obtained by embedding blocks into a layer c Performing region division into T/r 2 A non-overlapping region is input into the feature extraction module, where r 2 Indicating the size of the area, i.e. output conversionPerforming local attention fusion operation on all regions through a local attention fusion layer, namely performing region-multi-head-based extraction of global features O 'in the regions from attention operation in each region, promoting information interaction between the regions through an information exchange module, wherein the output of the information exchange module is O', splicing information blocks of all the regions together through a block aggregation module, and fusing information through projection, wherein the information is represented as O p =f pa (reshape (O ")), where f pa Representing a projection function, adopting a depth convolution as the operation of the block aggregation module, reshape (-) representing a splicing operation, and O' representing the output of the information exchange module, O p Representing the output of the block aggregation module;
after the training sample passes through all local attention fusion layers and block aggregation modules in the feature extraction module, the feature extraction module obtains extracted features, and the extracted features are input into global average pooling and full connectionIn a layer, the dimension of a feature is obtained by global average pooling compression, so that the feature is a one-dimensional vector z = [ z = 1 ,z 2 ,...,,z U ]U represents the dimension of the features obtained after global average pooling, and the network output p (x) is obtained by adopting a full connection layer and a softmax function i ) Is the ith sample x i The number of neurons of the fully connected layer is E, i.e. the number of types of interference;
s3, training the wireless interference recognition network constructed in the S2 by adopting the training data constructed in the S1, setting training cycle times K, setting the batch size as B, and defining a loss function as follows:
wherein q (x) i ) And p (x) i ) For the ith sample x i By updating the quantized network parameters by back-propagation, i.e. true label distribution and predicted distributionWherein theta is all trainable parameters of the network, and eta is the learning rate;
stopping training to obtain a trained parameter theta when the training cycle number is reached, so as to obtain a trained wireless interference identification network;
s4, interference identification is carried out: after the quasi-complex baseband interference signal obtained by the interference detection equipment is sampled r (n), the quasi-complex baseband interference signal is sent to a trained wireless interference identification network to obtain an identification result output by the network, namely the identification result isWherein P (-) represents a one-dimensional probability vector derived from the network output, e * Indicating the result obtained by the recognition.
2. The interference identification method based on deep learning of claim 1, wherein in step S2, the specific implementation method of the local attention fusion layer is as follows:
extracting global feature O' within a region by region-multi-headed self-attentive operations: the calculation mode is a multi-head self-attention mode, and the expression isWherein the method for calculating the query matrix Q, the key matrix K and the value matrix V is Q = O c W q 、K=Ο c W k 、V=Ο c W v Wherein W is q 、W k 、Is a projection matrix obtained by training, d represents the size of a vector, and the multi-head self-attention is composed of h parallel self-attention structures, namelyWherein W is a matrix obtained by training; meanwhile, the short connection operation introduced by the multi-head self-attention of the region increases the forward propagation path, that is, the original information block is added to the information block after the multi-head self-attention processing of the region, and the sum is O' = RMSA (O) c )+Ο c O' is the output of multi-head self-attention in the region;
after the blocks of each area are processed independently, one processed information block is randomly selected in each area and input to the information exchange module, the information exchange module is composed of one fully connected layer, and after the information exchange is performed through the operation, the information block is added to the rest of the information blocks of the corresponding area, where o "= FC (o ') + o', where o" denotes the output of the information exchange module.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210831105.6A CN115296759A (en) | 2022-07-15 | 2022-07-15 | Interference identification method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210831105.6A CN115296759A (en) | 2022-07-15 | 2022-07-15 | Interference identification method based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115296759A true CN115296759A (en) | 2022-11-04 |
Family
ID=83821692
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210831105.6A Pending CN115296759A (en) | 2022-07-15 | 2022-07-15 | Interference identification method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115296759A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116318470A (en) * | 2023-01-09 | 2023-06-23 | 中国电子科技集团公司第三十六研究所 | Method and device for estimating communication interference signal power under non-Gaussian noise |
CN116628473A (en) * | 2023-05-17 | 2023-08-22 | 国网上海市电力公司 | Power equipment state trend prediction method based on multi-factor neural network algorithm |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2007122188A1 (en) * | 2006-04-20 | 2007-11-01 | Wireless Audio I.P. B.V. | System and method for interference identification and frequency allocation |
CN104158611A (en) * | 2014-08-20 | 2014-11-19 | 西安电子科技大学 | System and method of detecting interference of wireless signal based on spectral analysis |
WO2018009476A1 (en) * | 2016-07-06 | 2018-01-11 | Booz Allen Hamilton Inc. | System and method for mitigating narrow-band interference |
WO2021008032A1 (en) * | 2019-07-18 | 2021-01-21 | 平安科技(深圳)有限公司 | Surveillance video processing method and apparatus, computer device and storage medium |
CN113033310A (en) * | 2021-02-25 | 2021-06-25 | 北京工业大学 | Expression recognition method based on visual self-attention network |
CN113435247A (en) * | 2021-05-18 | 2021-09-24 | 西安电子科技大学 | Intelligent identification method, system and terminal for communication interference |
CN114584440A (en) * | 2022-01-27 | 2022-06-03 | 西安电子科技大学 | Small sample AMC method based on Transformer coding network |
-
2022
- 2022-07-15 CN CN202210831105.6A patent/CN115296759A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2007122188A1 (en) * | 2006-04-20 | 2007-11-01 | Wireless Audio I.P. B.V. | System and method for interference identification and frequency allocation |
CN104158611A (en) * | 2014-08-20 | 2014-11-19 | 西安电子科技大学 | System and method of detecting interference of wireless signal based on spectral analysis |
WO2018009476A1 (en) * | 2016-07-06 | 2018-01-11 | Booz Allen Hamilton Inc. | System and method for mitigating narrow-band interference |
WO2021008032A1 (en) * | 2019-07-18 | 2021-01-21 | 平安科技(深圳)有限公司 | Surveillance video processing method and apparatus, computer device and storage medium |
CN113033310A (en) * | 2021-02-25 | 2021-06-25 | 北京工业大学 | Expression recognition method based on visual self-attention network |
CN113435247A (en) * | 2021-05-18 | 2021-09-24 | 西安电子科技大学 | Intelligent identification method, system and terminal for communication interference |
CN114584440A (en) * | 2022-01-27 | 2022-06-03 | 西安电子科技大学 | Small sample AMC method based on Transformer coding network |
Non-Patent Citations (4)
Title |
---|
PENGYU WANG: "WIR-Transformer: Using Transformers for Wireless Interference Recognition", 《IEEE WIRELESS COMMUNICATIONS LETTERS ( VOLUME: 11, ISSUE: 12, DECEMBER 2022)》, 12 July 2022 (2022-07-12), pages 2472 - 2476, XP011929857, DOI: 10.1109/LWC.2022.3190040 * |
梁金弟;程郁凡;杜越;王鹏宇: "联合多维特征的干扰识别技术研究", 《信号处理》, 25 December 2017 (2017-12-25) * |
王鹏宇: "基于卷积神经网络联合多域特征提取的干扰识别算法", 《信号处理》, 31 May 2022 (2022-05-31) * |
魏迪: "基于LSTM网络和特征融合的通信干扰识别", 《电讯技术》, 30 April 2022 (2022-04-30) * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116318470A (en) * | 2023-01-09 | 2023-06-23 | 中国电子科技集团公司第三十六研究所 | Method and device for estimating communication interference signal power under non-Gaussian noise |
CN116318470B (en) * | 2023-01-09 | 2024-05-10 | 中国电子科技集团公司第三十六研究所 | Method and device for estimating communication interference signal power under non-Gaussian noise |
CN116628473A (en) * | 2023-05-17 | 2023-08-22 | 国网上海市电力公司 | Power equipment state trend prediction method based on multi-factor neural network algorithm |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110086737B (en) | Communication signal modulation mode identification method based on graph neural network | |
CN115296759A (en) | Interference identification method based on deep learning | |
CN109450834A (en) | Signal of communication classifying identification method based on Multiple feature association and Bayesian network | |
CN112818891B (en) | Intelligent identification method for communication interference signal type | |
CN112039820B (en) | Communication signal modulation and identification method for quantum image group mechanism evolution BP neural network | |
CN113435247B (en) | Intelligent recognition method, system and terminal for communication interference | |
CN110659684A (en) | Convolutional neural network-based STBC signal identification method | |
CN112702294A (en) | Modulation recognition method for multi-level feature extraction based on deep learning | |
CN114465855A (en) | Attention mechanism and multi-feature fusion based automatic modulation recognition method | |
CN112910811B (en) | Blind modulation identification method and device under unknown noise level condition based on joint learning | |
CN111628833B (en) | MIMO antenna number estimation method based on convolutional neural network | |
CN105656826A (en) | Modulation recognizing method and system based on order statistics and machine learning | |
CN112307987B (en) | Method for identifying communication signal based on deep hybrid routing network | |
Zhang et al. | Lightweight automatic modulation classification via progressive differentiable architecture search | |
Stankowicz et al. | Unsupervised emitter clustering through deep manifold learning | |
Wei et al. | Differentiable architecture search-based automatic modulation classification | |
Wang et al. | Residual learning based RF signal denoising | |
CN111901267B (en) | Multi-antenna blind modulation identification method based on short-time Fourier transform time-frequency analysis | |
Liu et al. | A novel wireless interference identification and scheduling method based on convolutional neural network | |
Feng et al. | FCGCN: Feature Correlation Graph Convolution Network for Few-Shot Individual Identification | |
Yadav et al. | Application of Machine Learning Framework for Next‐Generation Wireless Networks: Challenges and Case Studies | |
CN115563485A (en) | Low-complexity interference identification method based on deep learning | |
CN112702132B (en) | Broadband spectrum sensing method based on convolutional neural network classifier | |
Wang et al. | Deep learning method for generalized modulation classification under varying noise condition | |
Yin et al. | Few-Shot Domain Adaption-Based Specific Emitter Identification Under Varying Modulation |
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 |