CN111680601A - Wireless signal modulation classifier visualization method based on long-term and short-term memory network - Google Patents

Wireless signal modulation classifier visualization method based on long-term and short-term memory network Download PDF

Info

Publication number
CN111680601A
CN111680601A CN202010484191.9A CN202010484191A CN111680601A CN 111680601 A CN111680601 A CN 111680601A CN 202010484191 A CN202010484191 A CN 202010484191A CN 111680601 A CN111680601 A CN 111680601A
Authority
CN
China
Prior art keywords
classifier
weight
modulation
input signal
points
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
Application number
CN202010484191.9A
Other languages
Chinese (zh)
Inventor
黄亮
张友
潘伟建
陈晋音
钱丽萍
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN202010484191.9A priority Critical patent/CN111680601A/en
Publication of CN111680601A publication Critical patent/CN111680601A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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

Abstract

A wireless signal modulation classifier visualization method suitable for a long-term and short-term memory network comprises the following steps: 1) introducing a group of weight vectors, measuring the importance degree of each signal point in the input signal to the output result of the classifier, reducing the loss function to update the weight vectors by defining the loss function consisting of constraint terms related to the classification precision and disturbance smoothness of the classifier, and finally normalizing the obtained weight vectors to be in a [0,1] range; 2) drawing a constellation diagram of the input signal, and endowing constellation points with different colors according to the weight obtained in the step 2); 3) setting a weight threshold, and if the weights of two continuous sampling points are greater than the threshold, connecting the two points through a green line segment. The method can be suitable for the wireless signal modulation classifier based on the LSTM, the distribution state of the feature points can be visually shown, and the change rule of the feature points is extracted by the classifier on the time sequence.

Description

Wireless signal modulation classifier visualization method based on long-term and short-term memory network
Technical Field
The invention belongs to the field of deep learning visualization and wireless signal automatic modulation classification, and relates to a visualization method of a wireless signal modulation classifier based on a long-short term memory network (LSTM).
Background
The wireless signal automatic modulation identification technology is a very key technology in the field of wireless communication, and the identification of a wireless signal modulation mode has very important significance in the military field and the civil field. With the continuous development of deep learning, the method based on deep learning obtains excellent performance in various classification tasks. However, the deep learning based classification method is different from the interpretable conventional feature-based modulation recognition method, and its classifier predicts the modulation class in an end-to-end manner, so it lacks interpretability. In order to solve the problem, the characteristics extracted by a classifier based on deep learning in the field of modulation and recognition are visualized by utilizing a visualization technology based on deep learning in the field of image processing, and the characteristics distribution and the time sequence characteristics are obtained, so that the decision made by the model is explained. Ensuring trust in the actual deployment of deep learning based classifiers. Therefore, it is necessary to design a visualization method suitable for the wireless signal modulation classifier based on deep learning.
Disclosure of Invention
In the wireless signal modulation classification method based on deep learning, the feature extraction of the classifier is invisible and inexplicable. To solve this problem, a set of weight vectors is introduced, whose shape and size are consistent with the size of the input signal. Wherein each weight value reflects the importance degree of the same position signal value to the classification result. In order to obtain the characteristic that the classifier extracts the features on the time sequence, a weight threshold value is introduced. And connecting the two or more continuous signal points by a line segment when the corresponding weight values of the two or more continuous signal points are greater than a given threshold value. The invention researches a feature extraction method of an LSTM-based classifier aiming at a wireless signal modulation classification task based on deep learning.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a visualization method for a wireless signal modulation classifier based on a long-short term memory network, the method comprising the steps of:
1) giving a trained wireless signal modulation classifier based on a long-short term memory network (LSTM) and consisting of m LSTM layers and n fully-connected layers, outputting a prediction vector through a normalized exponential function, and recording an input signal of the modulation classifier as
Figure BDA0002518521400000021
Wherein
Figure BDA0002518521400000022
NxRepresenting the length of the signal, each xiFrom a set of amplitudes AiAnd phase value phiiComposition, input signal x belongs to NyOne of the different modulation classes, a set of inputs x is mapped by the classifier into a set of vectors
Figure BDA0002518521400000023
The mapping process is denoted y ═ pi (x), where
Figure BDA0002518521400000024
yjBelongs to the j-th class, and the class with the highest probability is
Figure BDA0002518521400000025
2) Input x in amplitude and phase formi＝(Ai,φi) From inphase and quadrature forms of input xi＝(Ii,Qi) The conversion formula is as follows:
Figure BDA0002518521400000026
3) a group of weight vectors is introduced to visualize the LSTM-based wireless signal modulation classifier, each element in the weight vectors represents the importance degree of the signal value to the modulation classification result at the corresponding moment, and the steps of obtaining the weight vectors are as follows:
step 3.1: defining a width which is consistent with the length of the input signalWeight vector with degree of 1
Figure BDA0002518521400000027
All elements wiInitializing to 1;
step 3.2: the prediction probability of the input signal x through the LSTM classifier is noted as
Figure BDA0002518521400000028
Step 3.3: defining a function Φ (x, w) for masking the input signal, reducing the output probability of the modulation class for which the probability of the modulation classifier is the greatest, as follows:
Φ(x,w)＝(1-w)⊙x+ξw (2)
wherein [ ] indicates the multiplication operation of corresponding elements of the matrix, and [ ] is a constant deletion value;
step 3.4: and taking phi (x, w) as a new input of the modulation classifier, and obtaining a new prediction vector y ═ pi (phi (x, w)), wherein the maximum probability predicted by the classifier is recorded as
Figure BDA0002518521400000033
Step 3.5: defining a loss function for optimizing the weight vector, wherein the loss function is as follows:
Figure BDA0002518521400000034
wherein | · | purplepDenotes p-norm operation, λ1And λ2Is two regulating parameters, λ1||w||1In order for the weight vector to produce more 0-valued elements,
Figure RE-GDA0002629449650000033
to make the disturbance smoother;
step 3.6: setting a learning rate gamma, and continuously updating the weight vector w by reducing a loss function until the weight vector w is optimal, as follows:
Figure BDA0002518521400000037
step 3.7: normalizing the optimized weight vector w to be in the range of [0,1], and the following steps are carried out:
Figure BDA0002518521400000038
4) for better visualization, a constellation diagram of an input signal is drawn on a cartesian coordinate system, different colors (yellow to red) are given to each constellation point according to the weight vector obtained in the step 3) to distinguish the importance degree of different signal points to the final classification result, then a weight threshold is set to filter secondary feature points, and important and continuous feature points are connected by line segments, and the steps are as follows:
step 4.1: according to the input signal x at each moment (I)i,Qi) Drawing a constellation diagram in a Cartesian coordinate system, wherein the abscissa represents an in-phase value, and the ordinate represents an orthogonal phase value;
step 4.2: according to the obtained weight vector w', each constellation point x is processediAccording to the corresponding weight value wiGiving colors from one color to another with a greater weight wiI.e. the weight value wiSample points near 1 are painted another color, indicating that it is important for the classification result;
step 4.3 setting threshold ηwIf two successive sample points xiAnd xi+1Weight value w ofiAnd wi+1Greater than ηwThen the two points are connected by a line segment of the third color.
The technical idea of the invention is as follows: the method comprises the steps of firstly, inputting various elements in signals, wherein the influence of the elements in the signals on classification results is different, indicating the importance degree of signal points at different moments on the classification results by introducing weight vectors, mapping constellation stores at different moments into different colors from yellow to red on the basis of the obtained weight vectors, more intuitively representing different importance degrees, and connecting continuous feature points higher than a threshold value by introducing a feature point with a lower threshold value filtering weight value so as to find out the distribution characteristics of the feature points.
The invention has the following beneficial effects: the invention can visually represent the distribution state and the continuously changing trend of the characteristic points by introducing the weight vector and processing the characteristic points in the constellation diagram on the basis, thereby explaining the basis of correct or wrong classification results of the classifier and increasing the trust of the model in actual deployment.
Drawings
Fig. 1 is a flowchart of the optimization for obtaining weight vectors.
Fig. 2 is a schematic diagram of a visualization method framework of an LSTM-based wireless signal modulation classifier.
Detailed Description
The present invention is described in further detail below with reference to the attached drawing figures.
Referring to fig. 1 and 2, a visualization method for a wireless signal modulation classifier based on a long-term and short-term memory network is used for reducing loss by defining a loss function consisting of classifier classification precision and constraint terms of disturbance smoothness, setting a learning rate and continuously optimizing a weight vector. Meanwhile, processing the input signal according to the obtained weight vector, and visualizing important feature points (as shown in fig. 2), the method comprises the following steps:
1) giving a trained wireless signal modulation classifier based on a long-short term memory network (LSTM) and consisting of m LSTM layers and n fully-connected layers, outputting a prediction vector through a normalized exponential function, and recording an input signal of the modulation classifier as
Figure BDA0002518521400000051
Wherein
Figure BDA0002518521400000052
NxRepresenting the length of the signal, each xiFrom a set of amplitudes AiAnd phase value phiiComposition, input signal x belongs to NyOne of the different modulation classes, a set of inputs x is mapped by the classifier into a set of vectors
Figure BDA0002518521400000053
The mapping process is denoted y ═ pi (x), where
Figure BDA0002518521400000054
yjBelongs to the j-th class, and the class with the highest probability is
Figure BDA0002518521400000055
2) Input x in amplitude and phase formi＝(Ai,φi) From inphase and quadrature forms of input xi＝(Ii,Qi) The conversion formula is as follows:
Figure BDA0002518521400000056
3) a group of weight vectors is introduced to visualize the LSTM-based wireless signal modulation classifier, each element in the weight vectors represents the importance degree of the signal value to the modulation classification result at the corresponding moment, and the steps of obtaining the weight vectors are as follows:
step 3.1: defining a weight vector with the length consistent with the input signal and the width of 1
Figure BDA0002518521400000057
All elements wiInitializing to 1;
step 3.2: the prediction probability of the input signal x through the LSTM classifier is noted as
Figure BDA0002518521400000058
Step 3.3: defining a function Φ (x, w) for masking the input signal, reducing the output probability of the modulation class for which the probability of the modulation classifier is the greatest, as follows:
Φ(x,w)＝(1-w)⊙x+ξw (2)
wherein [ ] indicates the multiplication operation of corresponding elements of the matrix, and [ ] is a constant deletion value;
step 3.4: and taking phi (x, w) as a new input of the modulation classifier, and obtaining a new prediction vector y ═ pi (phi (x, w)), wherein the maximum probability predicted by the classifier is recorded as
Figure BDA0002518521400000063
Step 3.5: defining a loss function for optimizing the weight vector, wherein the loss function is as follows:
Figure BDA0002518521400000064
wherein | · | purplepDenotes p-norm operation, λ1And λ2Is two regulating parameters, λ1||w||1In order for the weight vector to produce more 0-valued elements,
Figure RE-GDA0002629449650000063
to make the disturbance smoother;
step 3.6: setting a learning rate gamma, and continuously updating the weight vector w by reducing a loss function until the weight vector w is optimal, as follows:
Figure BDA0002518521400000067
step 3.7: normalizing the optimized weight vector w to be in the range of [0,1], and the following steps are carried out:
Figure BDA0002518521400000068
4) for better visualization, a constellation diagram of an input signal is drawn on a cartesian coordinate system, different colors (yellow to red) are given to each constellation point according to the weight vector obtained in the step 3) to distinguish the importance degree of different signal points to the final classification result, then a weight threshold is set to filter secondary feature points, and important and continuous feature points are connected by line segments, and the steps are as follows:
step 4.1: according to the input signal x at each momentIs (I)i,Qi) Drawing a constellation diagram in a Cartesian coordinate system, wherein the abscissa represents an in-phase value, and the ordinate represents an orthogonal phase value;
step 4.2: according to the obtained weight vector w', each constellation point x is processediAccording to the corresponding weight value wiGiving colors from yellow to red with a greater weight wiI.e. the weight value wiSamples close to 1 were colored red, indicating that it is important for the classification results;
step 4.3 setting threshold ηwIf two successive sample points xiAnd xi+1Weight value w ofiAnd wi+1Greater than ηwThen connecting the two points through a green line segment;
the above is a complete process of calculating weight vectors and feature visualizations.

Claims (1)

1. A method for visualizing a wireless signal modulation classifier based on a long-short term memory network, the method comprising the steps of:
1) giving a trained wireless signal modulation classifier based on a long-short term memory network (LSTM), which consists of m LSTM layers and n fully-connected layers, outputting a prediction vector through a normalized exponential function, and recording an input signal of the modulation classifier as
Figure FDA0002518521390000011
Wherein
Figure FDA0002518521390000012
NxRepresenting the length of the signal, each xiFrom a set of amplitudes wAiAnd phase value phiiComposition, input signal x belongs to NyOne of the different modulation classes, a set of inputs x is mapped into a set of vectors by a classifier
Figure FDA0002518521390000013
The mapping process is denoted y ═ pi (x), where
Figure FDA0002518521390000014
yjBelongs to the j-th class, and the class with the highest probability is
Figure FDA0002518521390000015
2) Input x in amplitude and phase formi＝(Ai,φi) From inphase and quadrature forms of input xi＝(Ii,Qi) The conversion formula is as follows:
Figure FDA0002518521390000016
3) a group of weight vectors is introduced to visualize the LSTM-based wireless signal modulation classifier, each element in the weight vectors represents the importance degree of the signal value to the modulation classification result at the corresponding moment, and the steps of obtaining the weight vectors are as follows:
step 3.1: defining a weight vector with the length consistent with the input signal and the width of 1
Figure FDA0002518521390000017
All elements wiInitializing to 1;
step 3.2: the prediction probability of the input signal x through the LSTM classifier is noted as
Figure FDA0002518521390000018
Step 3.3: defining a function Φ (x, w) for masking the input signal, reducing the output probability of the modulation class for which the probability of the modulation classifier is maximal, as follows:
Φ(x,w)＝(1-w)⊙x+ξw (2)
wherein [ ] indicates the multiplication operation of corresponding elements of the matrix, and [ ] is a constant deletion value;
step 3.4: taking phi (x, w) as a new input of the modulation classifier, obtaining a new prediction vector y ═ pi (phi (x, w)),wherein the maximum probability predicted by the classifier is recorded as
Figure FDA0002518521390000021
Step 3.5: defining a loss function for optimizing the weight vector, wherein the loss function is as follows:
Figure FDA0002518521390000022
wherein | · | purplepDenotes p-norm operation, λ1And λ2Is two regulating parameters, λ1||w||1In order for the weight vector to produce more 0-valued elements,
Figure FDA0002518521390000023
to make the disturbance smoother;
step 3.6: setting a learning rate gamma, and continuously updating the weight vector w by reducing a loss function until the weight vector w is optimal, as follows:
Figure FDA0002518521390000024
step 3.7: normalizing the optimized weight vector w to be in the range of [0,1], and the following steps are carried out:
Figure FDA0002518521390000025
4) drawing a constellation diagram of an input signal on a Cartesian coordinate system, endowing each constellation point with different colors according to the weight vector obtained in the step 3) so as to distinguish the importance degree of different signal points to a final classification result, then setting a weight threshold value to filter secondary feature points, and connecting important and continuous feature points by line segments, wherein the steps are as follows:
step 4.1: according to the input signal x at each moment (I)i,Qi) The constellation diagram is drawn in a Cartesian coordinate system, the abscissa represents the in-phase value and the ordinate represents the quadrature phase value；
Step 4.2: according to the obtained weight vector w', each constellation point x is processediAccording to the corresponding weight value wiGiving colors from one color to another with a greater weight wiI.e. the weight value wiSample points near 1 are painted another color, indicating that it is important for the classification result;
step 4.3 setting threshold ηwIf two successive sample points xiAnd xi+1Weight value w ofiAnd wi+1Greater than ηwThen the two points are connected by a line segment of the third color.
CN202010484191.9A 2020-06-01 2020-06-01 Wireless signal modulation classifier visualization method based on long-term and short-term memory network Pending CN111680601A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010484191.9A CN111680601A (en) 2020-06-01 2020-06-01 Wireless signal modulation classifier visualization method based on long-term and short-term memory network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010484191.9A CN111680601A (en) 2020-06-01 2020-06-01 Wireless signal modulation classifier visualization method based on long-term and short-term memory network

Publications (1)

Publication Number Publication Date
CN111680601A true CN111680601A (en) 2020-09-18

Family

ID=72453048

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010484191.9A Pending CN111680601A (en) 2020-06-01 2020-06-01 Wireless signal modulation classifier visualization method based on long-term and short-term memory network

Country Status (1)

Country Link
CN (1) CN111680601A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112884122A (en) * 2021-02-23 2021-06-01 杭州弈鸽科技有限责任公司 Signal modulation type recognition model interpretable method and device based on neuron activation
CN114254680A (en) * 2022-02-28 2022-03-29 成都大公博创信息技术有限公司 Deep learning network modulation identification method based on multi-feature information

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101667252A (en) * 2009-10-15 2010-03-10 哈尔滨工业大学 Classification and identification method for communication signal modulating mode based on ART2A-DWNN
CN109873779A (en) * 2019-01-30 2019-06-11 浙江工业大学 A kind of grading type wireless identification of signal modulation method based on LSTM
CN110133599A (en) * 2019-01-08 2019-08-16 西安电子科技大学 Intelligent radar emitter Signals classification method based on long memory models in short-term
CN111026847A (en) * 2019-12-09 2020-04-17 北京邮电大学 Text emotion recognition method based on attention network and long-short term memory network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101667252A (en) * 2009-10-15 2010-03-10 哈尔滨工业大学 Classification and identification method for communication signal modulating mode based on ART2A-DWNN
CN110133599A (en) * 2019-01-08 2019-08-16 西安电子科技大学 Intelligent radar emitter Signals classification method based on long memory models in short-term
CN109873779A (en) * 2019-01-30 2019-06-11 浙江工业大学 A kind of grading type wireless identification of signal modulation method based on LSTM
CN111026847A (en) * 2019-12-09 2020-04-17 北京邮电大学 Text emotion recognition method based on attention network and long-short term memory network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIANG HUANG ETAL.: ""Visualizing Deep Learning-based Radio Modulation Classifier"", 《ARXIV》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112884122A (en) * 2021-02-23 2021-06-01 杭州弈鸽科技有限责任公司 Signal modulation type recognition model interpretable method and device based on neuron activation
CN112884122B (en) * 2021-02-23 2022-07-05 杭州弈鸽科技有限责任公司 Signal modulation type recognition model interpretable method and device based on neuron activation
CN114254680A (en) * 2022-02-28 2022-03-29 成都大公博创信息技术有限公司 Deep learning network modulation identification method based on multi-feature information

Similar Documents

Publication Publication Date Title
Zhao et al. Discriminative feature learning for unsupervised change detection in heterogeneous images based on a coupled neural network
Huang et al. Visualizing deep learning-based radio modulation classifier
CN107194336B (en) Polarized SAR image classification method based on semi-supervised depth distance measurement network
CN110516718A (en) The zero sample learning method based on depth embedded space
CN111680601A (en) Wireless signal modulation classifier visualization method based on long-term and short-term memory network
CN113259331B (en) Unknown abnormal flow online detection method and system based on incremental learning
CN110232448A (en) It improves gradient and promotes the method that the characteristic value of tree-model acts on and prevents over-fitting
Liang et al. Comparison detector for cervical cell/clumps detection in the limited data scenario
CN110969186B (en) Channel detection-based attack-resisting defense method and device facing wireless signal identification
CN113298186A (en) Network abnormal flow detection method for confluent flow model confrontation generation network and clustering algorithm
CN110991500A (en) Small sample multi-classification method based on nested integrated depth support vector machine
CN113392731A (en) Modulated signal classification method and system based on graph neural network
CN113642378A (en) Signal countermeasure sample detector design method and system based on N +1 type countermeasure training
Fuchs et al. Towards more specific estimation of membership functions for data-driven fuzzy inference systems
CN111582373A (en) Radiation source identification method based on weighted migration extreme learning machine algorithm
US20230162041A1 (en) Neural network model, method, electronic device, and readable medium
CN114707635A (en) Model construction method and device based on network architecture search and storage medium
CN115705393A (en) Radar radiation source grading identification method based on continuous learning
CN116486238B (en) Target fine granularity identification method combining point set representation and graph classification
CN111666861A (en) Wireless signal modulation classifier visualization method based on convolutional neural network
CN116958767A (en) Intelligent contract security analysis method and system based on multi-mode technology
CN116089812A (en) Fault diagnosis method based on semi-supervised contrast domain generalization intelligent model
Sun et al. Smart phone-based intelligent invoice classification method using deep learning
Cai et al. Mobile signal modulation recognition based on multimodal feature fusion
Lin et al. Features fusion based automatic modulation classification using convolutional neural network

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
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200918