CN113962262A - Radar signal intelligent sorting method based on continuous learning - Google Patents

Radar signal intelligent sorting method based on continuous learning Download PDF

Info

Publication number
CN113962262A
CN113962262A CN202111226205.8A CN202111226205A CN113962262A CN 113962262 A CN113962262 A CN 113962262A CN 202111226205 A CN202111226205 A CN 202111226205A CN 113962262 A CN113962262 A CN 113962262A
Authority
CN
China
Prior art keywords
radar signal
data
sorting
softmax
batch
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
Application number
CN202111226205.8A
Other languages
Chinese (zh)
Other versions
CN113962262B (en
Inventor
杨承志
吴宏超
邴雨晨
王美玲
许冰
王龙
周一鹏
易仁杰
王鸿超
吴焕欣
商犇
刘焕鹏
李吉民
石礼盟
曹鹏宇
陈泽盛
苏琮智
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.)
PLA AIR FORCE AVIATION UNIVERSITY
Original Assignee
PLA AIR FORCE AVIATION UNIVERSITY
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 PLA AIR FORCE AVIATION UNIVERSITY filed Critical PLA AIR FORCE AVIATION UNIVERSITY
Priority to CN202111226205.8A priority Critical patent/CN113962262B/en
Publication of CN113962262A publication Critical patent/CN113962262A/en
Application granted granted Critical
Publication of CN113962262B publication Critical patent/CN113962262B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F2218/02—Preprocessing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Signal Processing (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention relates to a radar signal intelligent sorting method based on continuous learning, which comprises the following steps: radar signal data are obtained and preprocessed to obtain a training set; combining a full convolution self-encoder with a logistic regression classifier to construct a combined deep learning model, and optimizing by adopting an orthogonal weight modification method to obtain a radar signal sorting model capable of being continuously learned; dividing the radar signals to be sorted into IQ data, inputting the IQ data into a trained radar signal sorting model for sorting, performing data dimension reduction processing on the IQ data by using the radar signal sorting model to obtain the sparse representation after dimension reduction, and giving the sparse representation after dimension reduction to a softmax function for radar signal sorting to finally output a corresponding label. The invention effectively improves the learning training efficiency and the real-time performance of the radar signal sorting network, and can quickly adapt to the dynamic change of the radar signal modulation format of the enemy, thereby obtaining the electromagnetic space control right in the electronic countermeasure of the enemy and the my.

Description

Radar signal intelligent sorting method based on continuous learning
Technical Field
The invention belongs to the technical field of electronic countermeasure technology and artificial intelligence, and particularly relates to a radar signal intelligent sorting method based on continuous learning.
Background
Artificial intelligence is one of the leading scientific fields at present, and deep learning is taken as a module of artificial intelligence. What is "intelligent" is that it has high adaptability to complex, dynamically changing environments. Deep Neural Networks (DNNs) are powerful tools for learning complex but fixed mapping rules between inputs and outputs, which remain unchanged but limit their application in more complex and dynamically changing situations. In the aspect of radar signal sorting and identification in the field of electronic countermeasures, only limited types of enemy radar signals can be collected. When the communication modulation format of an enemy radar is changed continuously, after new radar data are obtained continuously, the traditional DNN model with well-trained original data needs to start learning and training again to identify the new radar signal modulation format, but when the data volume is large, a large amount of time is consumed by a conventional deep learning and training mode, the learning and training efficiency is low, the real-time performance is poor, and the realization in actual operation is complex.
Disclosure of Invention
In order to solve the problems that in the prior art, when the communication modulation format of an enemy radar is continuously changed, a traditional deep learning model needs to restart learning and training, the learning and training efficiency is low, and the real-time performance is poor, the invention provides a radar signal intelligent sorting method based on continuous learning. When the continuous learning method is used for learning a new task, data characteristics in an old task do not need to be learned again, and the learned characteristics in the old task can be applied to the new task, so that the intelligent radar signal sorting method can show good performance in all tasks, and the continuity of the intelligent radar signal sorting method is reflected.
In order to achieve the purpose, the invention adopts the following technical scheme:
a radar signal intelligent sorting method based on continuous learning comprises the following steps:
the method comprises the following steps: acquiring radar signal data, preprocessing the radar signal data to obtain a data set, and randomly sampling from the data set to obtain a training set;
step two: combining a full convolution self-encoder with a logistic regression classifier, constructing a combined deep learning model CAE _ SOFTMAX, and optimizing a full convolution self-encoder network in the combined deep learning model CAE _ SOFTMAX by adopting an orthogonal weight modification method to perform continuous learning to obtain a radar signal sorting model CAE _ SOFTMAX _ OWM capable of being continuously learned;
step three: training the radar signal sorting model CAE _ SOFTMAX _ OWM by using the training set;
step four: acquiring a radar signal to be sorted, and dividing the radar signal to be sorted into IQ two-path data;
step five: inputting the IQ two paths of data into a trained radar signal sorting model CAE _ SOFTMAX _ OWM for sorting, performing data dimension reduction processing on the input IQ two paths of data by adopting a full convolution self-encoder to obtain a sparse representation after dimension reduction, and giving the sparse representation after dimension reduction to a SOFTMAX function for radar signal sorting to finally output a corresponding label.
The invention has the beneficial effects that: when the communication modulation format of an enemy radar is continuously changed, the traditional deep learning model DNN needs to restart learning training, but the radar signal sorting model CAE _ SOFTMAX _ OWM can adapt to the dynamically changed radar signal modulation format through continuous learning, the learning does not need to be started again, and meanwhile, the features learned in the old task can be applied to the new modulation format recognition task, so that the good sorting effect can be shown in all the tasks. The invention effectively improves the learning training efficiency and the real-time performance of the radar signal sorting network, and can quickly adapt to the dynamic change of the radar signal modulation format of the enemy, thereby obtaining the electromagnetic space control right in the electronic countermeasure of the enemy and the my.
Drawings
FIG. 1 is a diagram of a full convolution self-encoder CAE network architecture of the present invention;
FIG. 2 is a diagram of a joint deep learning model CAE _ SOFTMAX network architecture of the present invention;
FIG. 3 is a diagram of a CAE _ SOFTMAX _ OWM network structure of a radar signal sorting model according to the present invention;
FIG. 4 is a graph comparing the classification accuracy of a traditional deep learning model DNN and a radar signal sorting model CAE _ SOFTMAX _ OWM of the present invention in an experiment;
fig. 5 is a T-SNE visualization comparison graph of a conventional deep learning model DNN and a radar signal sorting model CAE _ software _ OWM of the present invention in an experiment, fig. 5(a) is a T-SNE visualization graph of the conventional deep learning model DNN, and fig. 5(b) is a T-SNE visualization graph of the radar signal sorting model CAE _ software _ OWM of the present invention.
Detailed Description
The technical solution of the present invention will be described in detail with reference to the accompanying drawings and preferred embodiments.
Traditional deep learning methods (abbreviated as DNN) learn complex and fixed mapping rules between inputs and outputs that can be used to sort radar signals. However, once the learning stage is finished, the mapping rule learned by the method is solidified, so that not only is it difficult to conveniently learn new mapping, but also flexible response to situation information existing in the actual environment, such as self state, environment change, task change and the like, cannot be made, and the sorting requirement of complex and variable radar signals is difficult to meet. The deep learning based stacked self-encoder (SAE for short) adopts implicit information in multi-layer neural network modeling data. Compared with a shallow model, a deep model can better approximate a nonlinear function, and a stacked self-encoder is applied to signal feature extraction and dimension reduction, and has the defect that the calculation amount of training network parameters is huge due to the adoption of a fully-connected network. The full convolution self-encoder (CAE for short) is SAE stacked self-encoder variant structure, and the CAE adopts local connection, weight sharing convolution layer, pooling operation and the like to replace SAE full connection layer, thereby greatly reducing network weight parameters, reducing network parameter scale and calculated amount, and consequently, CAE can be used to perform feature extraction on radar modulation signals to achieve the purpose of dimension reduction.
The invention provides a radar signal intelligent sorting method based on continuous learning, which comprises the following steps:
the method comprises the following steps: firstly, radar signal data used for training and testing a model are obtained, for example, the radar signal data can be obtained through a radar signal receiver, and received radar signals are divided into IQ two-path data to be stored; then, preprocessing the acquired radar signal data to obtain a data set, wherein the data set comprises 24 modulation formats, each modulation format comprises 26 signal-to-noise ratios, each signal-to-noise ratio comprises 3000 pieces of data, each piece of data comprises IQ two-path signals, and each path of signal comprises 1024 points, so that the size of the data set is as follows: 2555904 × 2 × 1024, data set sizes up to 18.01G. And taking a single IQ sample signal in the data set as a picture to reduce the dimension, and obtaining a training set and a test set by random sampling from the data set, wherein the training set is used for training a radar signal sorting model CAE _ SOFTMAX _ OWM, and after the model is trained, the test set is adopted for carrying out experimental testing.
Step two: and combining the full convolution self-encoder with the logistic regression classifier to construct a combined deep learning model CAE _ SOFTMAX, and optimizing a full convolution self-encoder network in the combined deep learning model CAE _ SOFTMAX by adopting an orthogonal weight modification method to perform continuous learning to obtain a radar signal sorting model CAE _ SOFTMAX _ OWM capable of being continuously learned.
Specifically, a full convolution self-Encoder (for short, CAE) is adopted to perform data dimension reduction on 1024 × 2 IQ modulation signals, the structure of a dimension reduction compression network CAE is shown in fig. 1, a single signal vector is input by a CAE Encoder (Encoder), features are extracted through operations such as down-sampling of a plurality of convolution layers and nonlinear activation, and dimension reduction representation is output through a full connection layer. The CAE Decoder (Decoder) input is a compressed representation of an encoder, and signal pixel vectors are reconstructed through operations of one full connection layer, a plurality of deconvolution up-sampling, nonlinear activation and the like. The full convolution self-encoder adopts the root mean square error as the cost function of the CAE, and the calculation formula of the cost function is as follows:
Figure BDA0003314430960000051
wherein m is the size of the mini-batch training adopted, xijRepresenting the ith input signal pixel x in a mini-batchi∈RnThe (j) th element of (a),
Figure BDA0003314430960000052
representing the i-th reconstructed signal pixel in a mini-batch
Figure BDA0003314430960000053
N is a vector xiAnd
Figure BDA0003314430960000054
length of (d).
The method is characterized in that a pooling layer and an anti-pooling layer are removed from a traditional convolution automatic encoder, convolution and transposition convolution are used for realizing up-sampling and down-sampling, the traditional convolution automatic encoder becomes a full convolution self encoder, in addition, an activation function of a coding output layer of the full convolution self encoder is replaced by a Leaky ReLU, and output is normalized to a range of [ -1,1 ]. The full convolution self-encoder sequentially comprises an input layer, five convolution layers, a folding layer, two full-connection layers, an unfolding layer and two transposition convolution layers, wherein the activation functions of the two full-connection layers are both Leaky ReLU. The CAE network parameters are shown in table 1.
TABLE 1 CAE network parameters
Figure BDA0003314430960000055
In table 1, In denotes an input layer of the network, Conv and Deconv denote a convolutional layer and a transposed convolutional layer, FC denotes a fully-connected layer, and Fold and unfold denote an unrolled layer and a folded layer between the fully-connected layer and the convolutional layer or the transposed convolutional layer. M represents the number of encoded output neurons from the encoder, dropout is used to prevent overfitting, and the LeakRelu activation function is used for normalization before binarization.
And adding a logistic regression classifier into the dimension reduction compression network CAE, wherein the logistic regression part uses a full connection layer to obtain a joint deep learning model CAE _ SOFTMAX shown in FIG. 2.
And training a combined deep learning model CAE _ SOFTMAX by adopting a batch processing updating strategy, wherein a cost function of the combined deep learning model CAE _ SOFTMAX is as follows:
Figure BDA0003314430960000061
where m is the size of the mini-batch taken for batch training, yijRepresenting the ith input vector x in a mini-batchiTrue category label y ofi∈RkThe (j) th element of (a),
Figure BDA0003314430960000062
representing the ith input vector x in a mini-batchiPrediction class label of
Figure BDA0003314430960000063
K is a vector yiAnd
Figure BDA0003314430960000064
length of (d).
The CAE _ SOFTMAX network parameters are shown in table 2.
TABLE 2 CAE _ SOFTMAX network parameters
Figure BDA0003314430960000065
In table 2, the softmax function is used for radar signal sorting directly using the compressed representation features extracted by the CAE network, and M represents the number of categories sorted.
And (3) combining an orthogonal weight updating method with a full convolution automatic encoder network, designing a CAE _ SOFTMAX _ OWM model, performing learning training by taking the CAE _ SOFTMAX _ OWM model as a whole, and optimizing a reconstruction error and a sorting error.
When a new task is learned, the orthogonal weight modification method (OWM for short) only modifies CAE _ SOFTMAX network weight parameters in the orthogonal direction of the input space of the old task, and the weight iterative computation hardly acts with the input of the previous task, so that the solution searched by the network in the training process of the new task is still in the solution space of the previous task. Mathematically, the OWM achieves its goal by the weight increment effect obtained by the orthogonal projection operator P and the error back propagation algorithm.
Consider a feed-forward network of L +1 layers, with the layer number L0, 1. In all hidden layers, the activation function g (-) is used. WlDenotes the connection W between the l-th layer and the (l-1) -th layerl∈Rs ×m。xlAnd ylRespectively representing the l-th layer output and input, where xl＝g(yl)，yl＝Wl Txl-1，xl-1∈Rs，yl∈Rm. Where s and m represent the dimensions of the input and output, respectively. In OWM, an orthogonal projection P is defined in the l-th layer input spacelIs the key for overcoming the catastrophic forgetting by continuous learning. In practice, P may be updated recursively for each tasklThe method is similar to the calculation of the correlation inverse matrix OWM method in RLS algorithms allowing to depend on the current input PlTo determine P of the last taskl. It also avoids the original definition of PlThe matrix inversion operation in (1).
The following provides specific calculation process implementation steps of the OWM method:
step 1: initializing parameters: a regularization constant β and a learning rate are input, and for L1, 2l(0) And a projection matrix Pl(0) Setting up Pl(0)＝Il/β。
Step 2: the ith batch input propagated forward at the jth task, then the loss propagated and the weight W used with the standard BP algorithml(i-1, j) calculating a weight modification value Δ Wl BP(i,j)。
And step 3: updating the weight matrix of each layer:
Wl(i,j)＝W(i-1,j)+κ(i,j)ΔWl BP(i,j) j＝1 (3)
Wl(i,j)＝W(i-1,j)+κ(i,j)Pl(j-1)ΔWl BP(i,j) j＝2,3,... (4)
where κ (i, j) is a predefined learning rate.
And 4, step 4: each batch repeats steps 2 through 3.
And 5: if the jth task has completed training, then the average of each batch input is propagated forward in turn in the jth task:
Figure BDA0003314430960000081
Figure BDA0003314430960000082
wherein the content of the first and second substances,
Figure BDA0003314430960000083
is the response of layer l-1 to the average of the jth task, ith batch input, and Pl(0,j)＝Pl(j-1)，i＝1,2,...,nj。
Step 6: projection P of an iterative computation matrixlAnd updating is performed.
If the orthogonal projection P for each batch is given by the formula (2)lPerforming update, alpha adopts
Figure BDA0003314430960000084
The same performance can be obtained with the algorithm with fading. This approach may be understood as treating each batch as a different task. It avoids the extra memory space and data reloading in step 4, thus significantly speeding up the processing. The learning rate is set as:
Figure BDA0003314430960000085
and 7: repeating steps 2 to 6 for the newly added task.
The orthogonal weight modifying method realizes effective protection of the existing weight value in the network, mainly acts on weight adjustment in the forward propagation process, can be completely compatible with the existing BP reverse gradient propagation algorithm, and shows good performance in a continuous learning test task.
The orthogonal weight modification method OWM is combined with the CAE _ SOFTMAX network to obtain a radar signal sorting model CAE _ SOFTMAX _ OWM of the present invention, and a network structure thereof is shown in fig. 3. In the CAE _ SOFTMAX _ OWM network architecture, examples of the full-convolution self-encoder, examples of the feature extraction module of the full-convolution self-encoder, and examples of the continuity learning module in the feature extraction module of the full-convolution self-encoder are included. x is the number of1 (1)Representation of full convolution reconstruction of original signal x from encoder1Other xn (1)Similarly. h islAn L-th layer feature extraction module representing a full convolution self-encoder. owm1Represents h1And the orthogonal weight modification method of the continuity learning module in the corresponding L-th layer feature extraction module. Essentially fig. 1 is a self-encoder structure, but a continuity learning method with orthogonal weight modifications embedded therein.
The data for each modulation format is equivalent to a task when sorting actual radar signal data. When real-time requirements are high, the network needs to be retrained for each additional signal of one modulation type. In order to reduce this limitation, the invention proposes a continuous learning method for radar signal sorting, which embodies its flexibility in the task of radar signal sorting.
To further illustrate the technical effects of the present invention, the following experiment analysis is performed on the conventional deep learning model DNN and the radar signal sorting model CAE _ SOFTMAX _ OWM of the present invention, the experiment operating environment is shown in table 3, a CAE _ SOFTMAX _ OWM network is used in the experiment to perform continuous learning training on 24 modulation format signals, and new tasks are sequentially added according to the modulation type label sequence during the continuous learning of the model.
TABLE 3 Experimental Environment configuration
Deep learning framework Tensorflow Keras
Operating system Ubuntu16.04
CPU Inter(R)Core(TM)i7-6700K CPU@4.00GHz
GPU Nvidia GeForce GTX 1080Ti(11GB)
RAM 16GB
Fig. 4 shows a comparison chart of classification accuracy on a test set of radar signal sorting by the radar signal sorting model CAE _ SOFTMAX _ OWM and the conventional deep learning model DNN, and table 4 shows a comparison of learning training time of the two models. As can be seen from fig. 4 and table 4, the classification accuracy of both classification models approaches as the modulation category of the new task increases, but the accuracy decreases. However, in the model learning training time, the continuous learning CAE _ SOFTMAX _ OWM model is significantly lower than the conventional deep learning model DNN.
TABLE 4 CAE _ SOFTMAX _ OWM versus DNN model training time (in seconds)
Figure BDA0003314430960000101
Further, the sparse representation after dimension reduction of the CAE _ SOFTMAX _ OWM model and the DNN model is visualized by a T-SNE method, the T-SNE method is called as a T-distribution field embedding algorithm, the feature layer of the output classification result of the two models is drawn into a two-dimensional scatter diagram, and the shapes and colors of the scatter diagram corresponding to the 24 modulation formats are given in the table 5. Fig. 5 shows a visual comparison of T-SNE method of the two models, and it can be seen from fig. 5 that both classification models have better sorting effect for 24 modulation formats.
Table 524 scatter diagram shape table corresponding to modulation format
Figure BDA0003314430960000111
In summary, the radar signal intelligent sorting method based on continuous learning provided by the invention adopts a full convolution automatic encoder network to perform dimensionality reduction processing on radar signals with different modulation formats, so as to obtain sparse representation of the radar signals; and then constructing a combined deep learning model CAE _ SOFTMAX for dimension reduction and sorting, and then combining the combined deep learning model CAE _ SOFTMAX with an orthogonal weight modification method to design a radar signal intelligent sorting model CAE _ SOFTMAX _ OWM capable of continuous learning. Compared with the traditional deep learning method (DNN for short), the method solves the problems that when the communication modulation format of an enemy radar is continuously changed, the traditional deep learning model DNN needs to restart learning training, the learning training efficiency is low, and the real-time performance is poor. Experimental results show that the method effectively improves the learning training efficiency and the real-time performance of the radar signal sorting network, and can quickly adapt to the dynamic change of the modulation format of the enemy radar signal, so that the electromagnetic space control right is obtained in electronic countermeasure of enemy and my.
The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims (7)

1. A radar signal intelligent sorting method based on continuous learning is characterized by comprising the following steps:
the method comprises the following steps: acquiring radar signal data, preprocessing the radar signal data to obtain a data set, and randomly sampling from the data set to obtain a training set;
step two: combining a full convolution self-encoder with a logistic regression classifier, constructing a combined deep learning model CAE _ SOFTMAX, and optimizing a full convolution self-encoder network in the combined deep learning model CAE _ SOFTMAX by adopting an orthogonal weight modification method to perform continuous learning to obtain a radar signal sorting model CAE _ SOFTMAX _ OWM capable of being continuously learned;
step three: training the radar signal sorting model CAE _ SOFTMAX _ OWM by using the training set;
step four: acquiring a radar signal to be sorted, and dividing the radar signal to be sorted into IQ two-path data;
step five: inputting the IQ two paths of data into a trained radar signal sorting model CAE _ SOFTMAX _ OWM for sorting, performing data dimension reduction processing on the input IQ two paths of data by adopting a full convolution self-encoder to obtain a sparse representation after dimension reduction, and giving the sparse representation after dimension reduction to a SOFTMAX function for radar signal sorting to finally output a corresponding label.
2. The continuous learning-based radar signal intelligent sorting method according to claim 1, wherein the full convolution self-encoder comprises an input layer, five convolution layers, a folding layer, two full-connected layers, an unfolding layer and two transposed convolution layers in sequence, wherein the activation functions of the two full-connected layers are both Leaky ReLU.
3. The radar signal intelligent sorting method based on continuous learning of claim 1 or 2, wherein a full convolution self-encoder adopts a root mean square error as a cost function, and the cost function calculation formula is as follows:
Figure FDA0003314430950000021
wherein m is the size of the mini-batch training adopted, xijRepresenting the ith input signal pixel x in a mini-batchi∈RnThe (j) th element of (a),
Figure FDA0003314430950000022
representing the i-th reconstructed signal pixel in a mini-batch
Figure FDA0003314430950000023
N is a vector xiAnd
Figure FDA0003314430950000024
length of (d).
4. The continuous learning-based radar signal intelligent sorting method according to claim 1 or 2, wherein a batch update strategy is adopted to train the joint deep learning model CAE _ SOFTMAX, and a cost function of the joint deep learning model CAE _ SOFTMAX is:
Figure FDA0003314430950000025
where m is the size of the mini-batch taken for batch training, yijRepresenting the ith input vector x in a mini-batchiTrue category label y ofi∈RkThe (j) th element of (a),
Figure FDA0003314430950000026
representing the ith input vector x in a mini-batchiPrediction class label of
Figure FDA0003314430950000027
K is a vector yiAnd
Figure FDA0003314430950000028
length of (d).
5. The radar signal intelligent sorting method based on continuous learning according to claim 1 or 2, wherein the calculation process of the orthogonal weight modification method comprises the following steps:
step 1: inputting a regularization constant beta and a learning rate, and randomly initializing the weight W of each layerl(0) And a projection matrix Pl(0)，l＝0,...,L；
Step 2: calculating the ith batch input of the jth task in a forward direction, solving by using a BP algorithm to obtain a weight Wl(i-1, j) calculating a weight modification value Δ Wl BP(i,j)；
And step 3: updating the weight matrix of each layer;
and 4, step 4: repeating the step 2 to the step 3 for each batch;
and 5: after the training of the task j is completed, the average value of each batch input is propagated forward in sequence;
step 6: projection P of an iterative computation matrixlAnd updating is performed.
6. The method for intelligently sorting radar signals based on continuous learning according to claim 1 or 2, wherein the data set comprises 24 modulation formats, each modulation format comprises 26 signal-to-noise ratios, each signal-to-noise ratio comprises 3000 pieces of data, each piece of data comprises IQ two-way signals, and each signal comprises 1024 points.
7. The continuous learning-based radar signal intelligent sorting method according to claim 1 or 2, wherein the radar signal data is acquired by a radar signal receiver.
CN202111226205.8A 2021-10-21 2021-10-21 Continuous learning-based intelligent radar signal sorting method Active CN113962262B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111226205.8A CN113962262B (en) 2021-10-21 2021-10-21 Continuous learning-based intelligent radar signal sorting method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111226205.8A CN113962262B (en) 2021-10-21 2021-10-21 Continuous learning-based intelligent radar signal sorting method

Publications (2)

Publication Number Publication Date
CN113962262A true CN113962262A (en) 2022-01-21
CN113962262B CN113962262B (en) 2024-05-14

Family

ID=79465829

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111226205.8A Active CN113962262B (en) 2021-10-21 2021-10-21 Continuous learning-based intelligent radar signal sorting method

Country Status (1)

Country Link
CN (1) CN113962262B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115508795A (en) * 2022-10-20 2022-12-23 中国人民解放军空军航空大学 Detection and interference integrated shared signal dynamic generation method

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5162805A (en) * 1975-02-19 1992-11-10 The United States Of America As Represented By The Secretary Of The Navy Frequency diversity sidelobe canceller
CN101587546A (en) * 2009-06-12 2009-11-25 浙江大学 Identification method of lower limb action pattern based on supporting vector multivariate classification
WO2014205231A1 (en) * 2013-06-19 2014-12-24 The Regents Of The University Of Michigan Deep learning framework for generic object detection
CN109116312A (en) * 2018-07-20 2019-01-01 电子科技大学 Radar signal classification method based on QMFB and convolutional neural networks
CN110569752A (en) * 2019-08-20 2019-12-13 河海大学 convolutional neural network-based radar signal category determination method
US20200110809A1 (en) * 2018-02-01 2020-04-09 Jungle Disk, L.L.C. Collating Information From Multiple Sources To Create Actionable Categories And Associated Suggested Actions
CN110991513A (en) * 2019-11-27 2020-04-10 佛山科学技术学院 Image target recognition system and method with human-like continuous learning capability
CN111948622A (en) * 2020-08-07 2020-11-17 哈尔滨工程大学 Linear frequency modulation radar signal TOA estimation algorithm based on parallel CNN-LSTM
CN112859025A (en) * 2021-01-05 2021-05-28 河海大学 Radar signal modulation type classification method based on hybrid network
CN113033473A (en) * 2021-04-15 2021-06-25 中国人民解放军空军航空大学 ST2DCNN + SE-based radar overlapped signal identification method

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5162805A (en) * 1975-02-19 1992-11-10 The United States Of America As Represented By The Secretary Of The Navy Frequency diversity sidelobe canceller
CN101587546A (en) * 2009-06-12 2009-11-25 浙江大学 Identification method of lower limb action pattern based on supporting vector multivariate classification
WO2014205231A1 (en) * 2013-06-19 2014-12-24 The Regents Of The University Of Michigan Deep learning framework for generic object detection
US20200110809A1 (en) * 2018-02-01 2020-04-09 Jungle Disk, L.L.C. Collating Information From Multiple Sources To Create Actionable Categories And Associated Suggested Actions
CN109116312A (en) * 2018-07-20 2019-01-01 电子科技大学 Radar signal classification method based on QMFB and convolutional neural networks
CN110569752A (en) * 2019-08-20 2019-12-13 河海大学 convolutional neural network-based radar signal category determination method
CN110991513A (en) * 2019-11-27 2020-04-10 佛山科学技术学院 Image target recognition system and method with human-like continuous learning capability
CN111948622A (en) * 2020-08-07 2020-11-17 哈尔滨工程大学 Linear frequency modulation radar signal TOA estimation algorithm based on parallel CNN-LSTM
CN112859025A (en) * 2021-01-05 2021-05-28 河海大学 Radar signal modulation type classification method based on hybrid network
CN113033473A (en) * 2021-04-15 2021-06-25 中国人民解放军空军航空大学 ST2DCNN + SE-based radar overlapped signal identification method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
PING LANG 等: "A Comprehensive Survey of Machine Learning Applied to Radar Signal Processing", 《ARXIV:2009.13702V1》, 29 September 2020 (2020-09-29), pages 1 - 49 *
刘松涛 等: "认知电子战研究进展", 《探测与控制学报》, vol. 42, no. 5, 31 October 2020 (2020-10-31), pages 1 - 15 *
文茜: "基于神经网络的复杂雷达信号分类方法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》, 15 May 2019 (2019-05-15), pages 136 - 637 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115508795A (en) * 2022-10-20 2022-12-23 中国人民解放军空军航空大学 Detection and interference integrated shared signal dynamic generation method
CN115508795B (en) * 2022-10-20 2024-04-02 中国人民解放军空军航空大学 Method for dynamically generating detection interference integrated shared signal

Also Published As

Publication number Publication date
CN113962262B (en) 2024-05-14

Similar Documents

Publication Publication Date Title
CN111882040B (en) Convolutional neural network compression method based on channel number search
CN109711426B (en) Pathological image classification device and method based on GAN and transfer learning
KR101326914B1 (en) Method for training neural networks
CN111461322B (en) Deep neural network model compression method
CN112926641B (en) Three-stage feature fusion rotating machine fault diagnosis method based on multi-mode data
CN112288011A (en) Image matching method based on self-attention deep neural network
CN112101364B (en) Semantic segmentation method based on parameter importance increment learning
CN112766496B (en) Deep learning model safety guarantee compression method and device based on reinforcement learning
CN112200733B (en) Grid denoising method based on graph convolution network
CN112634149A (en) Point cloud denoising method based on graph convolution network
CN113962262B (en) Continuous learning-based intelligent radar signal sorting method
KR102149355B1 (en) Learning system to reduce computation volume
Jang et al. Deep neural networks with a set of node-wise varying activation functions
CN114692830A (en) Self-strengthening image and voice deep learning model of promotion network
CN113516019A (en) Hyperspectral image unmixing method and device and electronic equipment
CN110472668B (en) Image classification method
CN112132207A (en) Target detection neural network construction method based on multi-branch feature mapping
CN116246110A (en) Image classification method based on improved capsule network
US20220253709A1 (en) Compressing a Set of Coefficients for Subsequent Use in a Neural Network
CN112488117B (en) Point cloud analysis method based on direction-induced convolution
US11657282B2 (en) Efficient inferencing with fast pointwise convolution
CN115205308A (en) Fundus image blood vessel segmentation method based on linear filtering and deep learning
CN112102156A (en) Controllable automobile image synthesis method based on causal flow model
Gautam et al. A sequential greedy approach for training implicit deep models
CN113837360B (en) DNN robust model reinforcement method based on relational graph

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