[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

CN114200421B - Multi-band sub-band signal fusion method - Google Patents

Multi-band sub-band signal fusion method Download PDF

Info

Publication number
CN114200421B
CN114200421B CN202111493579.6A CN202111493579A CN114200421B CN 114200421 B CN114200421 B CN 114200421B CN 202111493579 A CN202111493579 A CN 202111493579A CN 114200421 B CN114200421 B CN 114200421B
Authority
CN
China
Prior art keywords
band
training
echo signals
signal
full
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202111493579.6A
Other languages
Chinese (zh)
Other versions
CN114200421A (en
Inventor
蒋伊琳
张莉婷
唐三强
陈涛
刘鲁涛
郭立民
赵忠凯
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.)
Harbin Engineering University
Original Assignee
Harbin Engineering 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 Harbin Engineering University filed Critical Harbin Engineering University
Priority to CN202111493579.6A priority Critical patent/CN114200421B/en
Publication of CN114200421A publication Critical patent/CN114200421A/en
Application granted granted Critical
Publication of CN114200421B publication Critical patent/CN114200421B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01SRADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G01S7/41Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
    • G01S7/417Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/14Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
    • G06F17/141Discrete Fourier transforms
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Health & Medical Sciences (AREA)
  • Discrete Mathematics (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses a multi-band sub-band signal fusion method, which comprises the following steps: receiving N 1 groups of multi-band sub-band echo signals and N 1 groups of full-band echo signals, taking the distance envelope of the multi-band sub-band echo signals as a training data set and taking the distance envelope of the full-band echo signals as a training tag set; training a training sample formed by the training data set and the training label set into a DNN neural network for training, and storing a trained DNN neural network model after training is finished; and inputting the distance envelope of the multi-band sub-band echo signals to be fused into a trained DNN neural network model to obtain the distance envelope of the fused full-band echo signals. The invention can input the multi-sub-band distance envelope, directly obtain the broadband signal distance envelope, and improve the distance resolution. Through the training network, the complicated step of estimating the signal by the traditional algorithm is omitted, and errors generated in the subband prediction process are reduced. The trained network can obtain more test results in batches and has higher precision.

Description

Multi-band sub-band signal fusion method
Technical Field
The invention belongs to the field of radar signal processing, and relates to a multi-band subband signal fusion method, in particular to a multi-band subband signal fusion method based on a Deep Neural Network (DNN).
Background
Since the resolution unit of radar imaging depends on the bandwidth of the signal, it is often necessary to transmit a signal waveform with a large time-bandwidth product or to use a wideband chirp signal for imaging in order to obtain higher resolution. However, limited by the Nyquist sampling theorem, multiple upsampling frequencies are required to boost the signal bandwidth value, thus creating redundancy in the signal. Therefore, how to increase the bandwidth to increase the resolution and increase the signal utilization rate is a problem to be solved in the radar signal processing field. In 2006, the Hinton uses a pre-training method to relieve the problem of local optimal solution, pushes the hidden layer to 7 layers, and the neural network truly has 'depth', so that the hot tide of deep learning is uncovered. In order to overcome the gradient disappearance, transfer functions of ReLU, maxout and the like replace sigmoid, and form a basic form of DNN nowadays. The DNN network is a half-theoretical, half-empirical modeling way that uses as much training data as possible and the large-scale computing power of a computer to adjust internal parameters as close as possible to the problem target, so it can be used in the processing of signals. DNN network has the advantages of strong learning ability, wide coverage, strong adaptability and good portability.
Disclosure of Invention
Aiming at the prior art, the technical problem to be solved by the invention is to provide a multi-band subband signal fusion method utilizing a Deep Neural Network (DNN) to improve the utilization rate of signals.
In order to solve the technical problems, the multi-band subband signal fusion method of the invention comprises the following steps:
S1: receiving N 1 groups of multi-band sub-band echo signals and N 1 groups of full-band echo signals, taking the distance envelope of the multi-band sub-band echo signals as a training data set and taking the distance envelope of the full-band echo signals as a training tag set;
S2: training a training sample formed by the training data set and the training label set into a DNN neural network for training, and storing a trained DNN neural network model after training is finished;
s3: and inputting the distance envelope of the multi-band sub-band echo signals to be fused into a trained DNN neural network model to obtain the distance envelope of the fused full-band echo signals.
Further, the distance envelope of the multi-band subband echo signal is specifically: multiplying N 1 groups of multi-band sub-band echo signals by the conjugate of a reference signal to obtain a baseband signal, and performing inverse discrete Fourier transform on the baseband signal to obtain the distance envelope of the multi-band sub-band echo signals, wherein the reference signal is a full-band linear frequency modulation signal transmitted by a transmitting end.
Further, the distance envelope of the full-band echo signal is specifically: multiplying N 1 groups of full-band echo signals by the conjugate of a reference signal to obtain a baseband signal, and performing inverse discrete Fourier transform on the baseband signal to obtain the distance envelope of the full-band echo signals, wherein the reference signal is a full-band linear frequency modulation signal transmitted by a transmitting end.
Further, in S2, the training data set and the training label set form a training sample, and the training sample is sent to the DNN neural network to perform training includes:
Initializing DNN neural network model parameters, comprising: the total number of layers n, the number of neurons of an input layer, each hidden layer and an output layer, an activation function, a loss function, an iteration step alpha, a maximum iteration number m and a stop iteration threshold epsilon;
The ith element x i in the training data set and the ith element y i in the training label set are combined into the ith training sample (x i,yi),i=1,2,…,N1, the training sample is obtained) Feeding the training sample into a DNN neural network;
When the change values of the linear relation coefficient matrix W and the bias vector b of each hidden layer and the output layer are smaller than the stop iteration threshold epsilon, outputting W and b as the linear relation coefficient matrix and the bias vector of the trained DNN neural network model.
The invention has the beneficial effects that: the deep neural network is used for signal fusion, a multi-subband linear frequency modulation signal is transmitted at a transmitting end, a multi-subband echo signal is multiplied by the conjugate of a reference signal to obtain a baseband signal at a receiving end, inverse Discrete Fourier Transform (IDFT) is carried out on the baseband signal to obtain a multi-subband distance envelope, the multi-subband distance envelope and a full-band distance envelope are respectively used as training data and labels to send the training data and the training data into DNN, any group of multi-subband distance envelopes can be input into a trained DNN model, the full-band distance envelope can be obtained, the distance resolution is improved, the utilization rate of the signal is greatly improved, and the high-resolution range image of the fused full-band signal is obtained.
DNN can be input into a multi-sub-band distance envelope to directly obtain a wideband signal distance envelope, so that the distance resolution is improved. Through the training network, the complicated steps required by the traditional algorithm for estimating the signal can be omitted, and errors generated in the subband prediction process are reduced. The traditional algorithm can only obtain one prediction result at a time, and the trained network can obtain more test results in batches, so that DNN is more suitable when batch test data are obtained. And the RMSE of the result consisting of the batch data of the network training is smaller than that of the single test data in the conventional algorithm, namely, the accuracy of DNN is higher.
Drawings
FIG. 1 is a multi-subband distance envelope;
FIG. 2 is a distance envelope of a full frequency band;
FIG. 3 is a RMSE for training and test sets;
FIG. 4 (a) is a double scattering point network test result one;
FIG. 4 (b) is a double scattering point network test result two;
FIG. 5 (a) is a network test result one of three scattering points;
FIG. 5 (b) is a network test result two for three scattering points;
FIG. 6 (a) is a network test result one of four scattering points;
FIG. 6 (b) is a network test result two for four scattering points;
fig. 7 is RMSE of all pole algorithm, noise reduction algorithm and DNN.
Detailed Description
The invention is further described below with reference to the drawings and specific examples.
The invention comprises the following steps:
Step one: receiving multi-band subband signals and constructing a training data set and a tag set;
Transmitting N 1 groups of multiband subband signals, receiving the multiband subband echo signals at a receiving end, multiplying the N 1 groups of multiband subband echo signals by the conjugate of a reference signal to obtain a baseband signal, performing Inverse Discrete Fourier Transform (IDFT) on the baseband signal to obtain a distance envelope of a target, multiplying the N 1 groups of full-band echo signals by the conjugate of the reference signal to obtain a baseband signal, performing inverse discrete Fourier transform on the baseband signal to obtain a distance envelope of a full-band echo signal, and taking the reference signal as a full-band linear frequency modulation signal transmitted by a transmitting end. Finally, taking the distance envelope of N 1 groups of multi-subband signals as a training data set and the distance envelope of N 1 groups of full-band echo signals as a training tag set;
step two: training DNN and fusing multi-band sub-band echo signals;
Setting up DNN with the layer number of N, setting the neuron number of an input layer as N 1, the neuron number of an output layer as N n, the neuron number of each layer of a middle hidden layer as N 2,N3,...,Nn-1 respectively, setting the training round number as m, setting the training data set number of each round as l, sending a training data set and a label set into DNN for training, and storing a DNN model after training is finished. And (3) calling a model, inputting a group of multi-band sub-band distance envelopes, and obtaining the distance envelopes of the full-band echo signals through DNN network fusion.
The examples are given in connection with specific signals and parameters:
Step one: the linear frequency modulation signal of the multiband sub-band is used as a transmitting signal, and the number of scattering points, the initial distance between the scattering points and the radar and the relative distance between the scattering points are used as variables, so that the expression of the sub-band echo signal of the receiving end is as follows:
Where m is the number of scattering points, A m is the scattering intensity of the mth scattering point, f c is the signal carrier frequency, T is time, K is the modulation frequency, determined by the ratio of the signal bandwidth B to the time width T r, τ m is the time delay, obtained by twice the ratio of the distance R m to the speed of light c.
The N 1 sets of multiband subband echo signals are multiplied by the conjugate of the reference signal to obtain a baseband signal. The expression of the reference signal is:
the baseband signal is thus expressed as:
Where N represents the number of sampling points of the baseband signal, which can be obtained by multiplying the sampling frequency f s by the time width T r.
And performing Inverse Discrete Fourier Transform (IDFT) on the baseband signals to obtain a distance envelope of the target, and finally taking the distance envelopes of N 1 groups of multi-subband signals with different signal to noise ratios as a training data set and N 1 groups of full-band transmitting signals as a training tag set, wherein the following two formulas are shown:
Glabel(n)=|IDFT{[s(n)+n(n)],Np}|
Where IDFT [ s 2i-1(n),Np ] represents an IDFT of N p points per subband, N p represents the number of samples of the wideband signal, N 0 represents the number of samples of the subband signal, and N i (N) is a noise sequence.
The second step comprises:
Step 2.1: training DNN networks
Input: total layer number N, number of nodes of input layer, hidden layer and output layer, activation function, loss function, iteration step length alpha, maximum iteration number m, stop iteration threshold epsilon, N 1 training samples
And (3) outputting: a linear relation coefficient matrix W and a bias vector b of each hidden layer and the output layer;
(1) Initializing the values of a linear relation coefficient matrix W and a bias vector b of each hidden layer and the output layer to be a random value;
(2)for epoch=1 to m:
(2-1)for i=1 to N1
(a) Setting DNN input a 1 to (x i,yi);
(b) for k=2 to n, forward propagation algorithm calculation a i,k=σ(zi,k)=σ(Wkai,k-1+bk);
(c) Calculating delta i,n of the output layer through a loss function;
(d) for k=n-1 to 2, performing a back propagation algorithm to calculate δ i,k=(Wk+1)Tδi,k+1⊙σ'(zi,k);
(2-2) for k=2 to n, update W k,bk of the k-th layer:
(2-3) if the variation values of all W, b are smaller than the stop iteration threshold epsilon, jumping out of the iteration loop to the step 3;
(3) And outputting a linear relation coefficient matrix W and a bias vector b of each hidden layer and the output layer.
In the steps: a 1 denotes an input of an input layer of DNN; (x i,yi) represents an ith training sample, an ith element x i in the training data set and an ith element y i in the training label set form the ith training sample (x i,yi);ai,k represents the output of the ith data set at the kth layer, sigma (-) represents a forward propagation activation function, z i,k represents the input of the ith data set at the kth layer, W k represents a linear relation coefficient vector of the kth layer, b k represents a bias vector of the kth layer, delta i,k represents a gradient of z i,k, symbol T represents matrix inversion, symbol +.represents matrix Hadamard product, sigma' (-) represents a derivative of the forward propagation activation function.
Step 2.2: fusion of multi-band sub-band echo signals
And calling a trained DNN network, and inputting an echo distance envelope of a group of multi-band subband signals to obtain the fused full-band echo distance envelope.
And combining specific parameters and data, taking the signal-to-noise ratio, the number of scattering points, the initial distance between the scattering points and the radar and the relative distance between the scattering points as training set variables, and generating a distance envelope training set. Parameters of training set and DNN are shown in table 1:
Table 1 training set and DNN parameters
Setting full-band distance envelope corresponding to multiple sub-bands as a tag set, setting up DNN with the layer number of 5, setting up input layer neuron number of 1500, output layer neuron number of 1500, and neuron numbers of middle 3-layer hidden layers of 1024, 1024 and 1024 respectively, setting up training round number of 15000 and training data set number of 100 in each round, sending the training data set and the tag set into DNN for training, and storing DNN model after training. And (3) calling a model, inputting a multi-sub-band distance envelope test set, and obtaining a full-band distance envelope result through DNN network fusion.
Fig. 1 is a multi-subband distance envelope, fig. 2 is a full-band distance envelope, after a data set and a label set are subjected to multiple iterative training, weights of all layers of a DNN model are adjusted according to errors of estimation results, and a training model is stored. According to the invention, the performance of the network model is evaluated by adopting Root Mean Square Error (RMSE), the data set is trained, the variation condition of the RMSE along with the number of training rounds is shown in fig. 3, test data are consistent with training data through the comparison of the RMSE of the training set and the RMSE of the test set, the difference value of the two RMSE is 0.01, and the network test can be considered to have better performance. Comparing the test results of the multi-subband distance envelope and the full-band signal distance envelope, the following conclusion can be drawn: as shown in fig. 4 (a), fig. 4 (b), fig. 5 (a), fig. 5 (b), fig. 6 (a) and fig. 6 (b), DNN still has an accurate fitting degree to a plurality of scattering points, the network can input a multi-subband distance envelope to directly obtain a full-band signal distance envelope, through training the network, the complicated steps required by estimating the signal by the traditional algorithm can be omitted, more test results can be obtained in batches by the trained network, but the traditional algorithm can obtain the test results only by performing integral calculation once every time. And the RMSE of the network training results is smaller than that of the single test data in the all-pole algorithm and the noise reduction algorithm, as shown in fig. 7, which also means that the network has higher accuracy than the conventional algorithm.

Claims (4)

1. The multi-band subband signal fusion method is characterized by comprising the following steps of:
S1: receiving N 1 groups of multi-band sub-band echo signals and N 1 groups of full-band echo signals, taking the distance envelope of the multi-band sub-band echo signals as a training data set and taking the distance envelope of the full-band echo signals as a training tag set;
S2: training a training sample formed by the training data set and the training label set into a DNN neural network for training, and storing a trained DNN neural network model after training is finished;
s3: and inputting the distance envelope of the multi-band sub-band echo signals to be fused into a trained DNN neural network model to obtain the distance envelope of the fused full-band echo signals.
2. The method for multi-band subband signal fusion according to claim 1, wherein: the distance envelope of the multi-band sub-band echo signal is specifically: multiplying N 1 groups of multi-band sub-band echo signals by the conjugate of a reference signal to obtain a baseband signal, and performing inverse discrete Fourier transform on the baseband signal to obtain the distance envelope of the multi-band sub-band echo signals, wherein the reference signal is a full-band linear frequency modulation signal transmitted by a transmitting end.
3. The method for multi-band subband signal fusion according to claim 1, wherein: the distance envelope of the full-band echo signal is specifically: multiplying N 1 groups of full-band echo signals by the conjugate of a reference signal to obtain a baseband signal, and performing inverse discrete Fourier transform on the baseband signal to obtain the distance envelope of the full-band echo signals, wherein the reference signal is a full-band linear frequency modulation signal transmitted by a transmitting end.
4. The method for multi-band subband signal fusion according to claim 1, wherein: s2, the step of forming a training sample by the training data set and the training label set and sending the training sample to the DNN neural network for training comprises the following steps:
Initializing DNN neural network model parameters, comprising: the total number of layers n, the number of neurons of an input layer, each hidden layer and an output layer, an activation function, a loss function, an iteration step alpha, a maximum iteration number m and a stop iteration threshold epsilon;
The ith element x i in the training data set and the ith element y i in the training label set are combined into the ith training sample (x i,yi),i=1,2,…,N1, the training sample is obtained) Feeding the training sample into a DNN neural network;
When the change values of the linear relation coefficient matrix W and the bias vector b of each hidden layer and the output layer are smaller than the stop iteration threshold epsilon, outputting W and b as the linear relation coefficient matrix and the bias vector of the trained DNN neural network model.
CN202111493579.6A 2021-12-08 2021-12-08 Multi-band sub-band signal fusion method Active CN114200421B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111493579.6A CN114200421B (en) 2021-12-08 2021-12-08 Multi-band sub-band signal fusion method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111493579.6A CN114200421B (en) 2021-12-08 2021-12-08 Multi-band sub-band signal fusion method

Publications (2)

Publication Number Publication Date
CN114200421A CN114200421A (en) 2022-03-18
CN114200421B true CN114200421B (en) 2024-08-23

Family

ID=80651373

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111493579.6A Active CN114200421B (en) 2021-12-08 2021-12-08 Multi-band sub-band signal fusion method

Country Status (1)

Country Link
CN (1) CN114200421B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114942415B (en) * 2022-05-17 2024-10-25 哈尔滨工程大学 Multi-band subband signal fusion method based on self-encoder

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111175719A (en) * 2020-01-08 2020-05-19 中国船舶重工集团公司第七二四研究所 Intelligent track starting method based on BP neural network

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US12032089B2 (en) * 2019-03-14 2024-07-09 Infineon Technologies Ag FMCW radar with interference signal suppression using artificial neural network

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111175719A (en) * 2020-01-08 2020-05-19 中国船舶重工集团公司第七二四研究所 Intelligent track starting method based on BP neural network

Also Published As

Publication number Publication date
CN114200421A (en) 2022-03-18

Similar Documents

Publication Publication Date Title
CN110365612B (en) Deep learning beam domain channel estimation method based on approximate message transfer algorithm
CN112073106B (en) Millimeter wave beam prediction method and device, electronic device and readable storage medium
Zhang et al. BP algorithm for the multireceiver SAS
CN108008385B (en) Interference environment ISAR high-resolution imaging method based on management loading
Liang et al. A transfer learning approach for compressed sensing in 6G-IoT
CN114745237B (en) Channel estimation method of intelligent super-surface auxiliary multi-user wireless communication system
CN108226928A (en) Based on the inverse synthetic aperture radar imaging method for it is expected propagation algorithm
CN114200421B (en) Multi-band sub-band signal fusion method
CN114201987A (en) Active interference identification method based on self-adaptive identification network
CN112162280B (en) SF ISAR one-dimensional high-resolution distance imaging method based on atomic norm minimization
CN117192548A (en) Sparse ISAR high-resolution imaging method based on depth expansion
CN114624646B (en) DOA estimation method based on model driven complex neural network
CN114942415B (en) Multi-band subband signal fusion method based on self-encoder
CN115236584A (en) Meter-wave radar low elevation angle estimation method based on deep learning
CN110471026B (en) Phase-enhanced meter-wave radar target low elevation DOA estimation method
CN109541567B (en) High-speed maneuvering target detection method based on deep learning
CN116540203B (en) Broadband radar high-speed target coherent accumulation method based on rapid sparse Bayes
CN112422208A (en) Signal detection method based on antagonistic learning under unknown channel model
CN110133577A (en) The relatively prime MIMO array DOA algorithm for estimating of single base based on time-frequency residual error network
Alimosaymer et al. Systematic approach in designing wavelet packet modulation‐orthogonal frequency division multiplexing radar signal by applying the criterion of least‐squares
Chen et al. Joint angle and range estimation for frequency diverse array using multi-layer perceptron neural network
Li et al. ISAR range alignment under sparse aperture condition based on CRAN
CN116299290A (en) Off-grid target joint parameter estimation method based on LADMM-DNN network
CN113126095B (en) Two-dimensional ISAR rapid imaging method based on sparse Bayesian learning
CN117331044A (en) Direction of arrival estimation method based on impulse neural convolutional 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
GR01 Patent grant
GR01 Patent grant