Channelized intelligent receiving method based on BDVNN

The BDVNN-based channelized receiver method efficiently separates and recognizes time-domain aliasing signals across platforms, addressing power and integration issues in traditional receivers, ensuring real-time and accurate signal processing.

CN119070933APending Publication Date: 2024-12-03HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411072222.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-06
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Traditional electronic warfare receivers face challenges with high power consumption, low integration, and information loss, and require improved signal processing capabilities to adapt to diverse signal types and frequency ranges, especially with the integration of deep learning technology for real-time signal recognition and interference strategies.

Method used

A channelized intelligent receiving method based on BDVNN, utilizing an SDR device to receive and separate time-domain aliasing signals, performing spectrum sensing and BDVNN recognition in a Matlab environment, with data interaction through Python for cross-platform recognition, reducing preprocessing time and memory overhead.

Benefits of technology

Achieves rapid and accurate recognition of time-domain mixed signals with reduced development and verification cycles, improving real-time performance and accuracy while overcoming platform mismatches, and simplifying model development.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119070933A_ABST
    Figure CN119070933A_ABST
Patent Text Reader

Abstract

The invention discloses a channelized intelligent receiving method based on a BDVNN, and the method comprises the steps: carrying out the signal receiving and collection of a specific frequency band through SDR equipment, carrying out the separation of time domain aliasing signals through a channelized structure in a Matlab environment of an upper computer, carrying out the spectrum sensing of each separated signal according to an energy detection method, completing the training of a BDVNN recognition network model through Python, and carrying out the recognition of a BDVNN recognition network model. And finally, cross-platform pattern recognition of the time domain aliasing signals is realized by using the trained BDVNN model in a Matlab and Python interaction mode. The method can effectively process a plurality of aliasing signals arriving at the same time and signals with diversified types in real time, overcomes mismatching between Matlab environment data acquisition and Python environment model development, remarkably simplifies development and verification periods, and avoids tedious work of rebuilding and training a model at a Matlab end.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of software radio channelized receivers, and in particular to a channelized intelligent receiving method based on BDVNN. Background Art

[0002] With the rapid development of communication and signal processing, especially the advancement of electronic warfare and communication systems, modern receiver technology is facing increasingly severe challenges. Traditional electronic warfare receivers are usually built with analog devices, but they have problems such as high power consumption, low integration, and easy information loss. With the advancement of digital technology, especially the improvement of analog-to-digital converter technology, digital channelized receivers designed based on software radio technology are gradually replacing traditional analog front ends, realizing efficient digital processing of signals, and are widely used in radio monitoring, signal reconnaissance, and tactical communication systems, which can quickly and accurately acquire and process enemy communications and radar signals.

[0003] With the continuous advancement of communication technology, the widespread application of new electronic systems has led to an increase in technical requirements and a diversification of signal types. This trend has put forward higher requirements for receiver technology: it needs to adapt to a wider frequency range, process aliased signals of different strengths, and recognize diverse signal patterns. The introduction of deep learning technology provides a new solution for the intelligent development of channelized receivers. Deep learning networks can learn complex signal patterns and features from a large amount of data, thereby improving the efficiency and accuracy of signal processing. In the field of military reconnaissance, the application of deep learning technology enables receivers to quickly identify enemy signal sources, provide real-time support for battlefield intelligence, and assist in the formulation of effective interference and protection strategies. In the commercial field, the signal pattern recognition capability enhanced by deep learning helps to realize intelligent spectrum management. Therefore, the development of intelligent channelized receiver technology based on deep learning is of great significance for adapting to dynamic and complex electromagnetic environments and meeting diverse application needs, and realizing real-time monitoring of multiple scenarios of a single device. Summary of the invention

[0004] In view of the limitations in the prior art, the present invention provides a channelized intelligent receiving method based on BDVNN, which can effectively separate, perceive and quickly and automatically identify time-domain aliasing signals; a specific frequency band is received and collected by an SDR device (receiver), and in the Matlab environment of the host computer, the time-domain aliasing signals are separated by using a channelized structure, and spectrum perception is performed on each separated signal according to an energy detection method, and the BDVNN recognition network model is trained through Python. Finally, the trained BDVNN model is used to realize cross-platform style recognition of time-domain aliasing signals through the interaction of Matlab and Python, which can effectively process multiple aliasing signals arriving simultaneously and signals of diverse types in real time.

[0005] To achieve the above objectives, the present invention provides a channelized intelligent receiving method based on BDVNN, and the specific implementation steps are as follows:

[0006] Step 1: Construct the polyphase filter coefficient matrix

[0007] 1.1. Design a FIR low-pass filter and obtain the FIR low-pass filter coefficient vector h = [h1,h2,···,h t+1 ];

[0008] 1.2. Perform polyphase decomposition on the designed FIR low-pass filter and obtain:

[0009] h i,n =h nD+i-D ,i=1,2,…,D; n=1,2,…,N

[0010] in D is the number of channels, t is the filter order of the FIR low-pass filter, an even number of channels is taken and N≥1 is satisfied. It means rounding down, and finally the decomposed polyphase filter coefficient matrix H is as follows:

[0011]

[0012] Step 2: Build and train the BDVNN network model. The structure of the BDVNN network model is as follows: the input data undergoes preliminary feature extraction through parallel structures A and B respectively. The features of the parallel structures A and B are merged and then passed through serial structures C and D. Finally, a fully connected layer and a Softmax layer are passed to obtain the classification probability value. The class with the largest classification probability value is taken as the judgment category.

[0013] Step 3: Data collection: After receiving the RF signal, the channelized receiver outputs a complex baseband signal.

[0014] Step 4: Perform complex baseband signal processing

[0015] 4.1. Decimate the complex baseband signal by 2 times to obtain r = [r1, r2, ···, r n ],r k =I k +jQ k ,k=1,2,…,n, where n is the number of sampling points, I k is the real part of the complex baseband signal, Q k is the imaginary part of the complex baseband signal, and j is the imaginary unit.

[0016] 4.2. For r = [r1, r2, ···, r n ] to perform serial-to-parallel conversion, the results are as follows:

[0017] r i,s =r sD+i-D ,i=1,2,…,D; s=1,2,…,M

[0018] Then we get the matrix R:

[0019]

[0020] in D is the number of channels, Indicates rounding down.

[0021] Step 5: Channelize and filter the complex baseband signal processed in step 4

[0022] 5.1. Expand the matrix R to get R′:

[0023]

[0024] where r i, ' l The definition is as follows:

[0025]

[0026] Then we get the output y′ of the i-th row i =(y′ i,1 ,y′ i,2 ,,…,y′ i,m ), i=1,2,…,D; m=1,2,…,M+N-1, where y i,m The expression is as follows:

[0027]

[0028] Where j represents an imaginary number.

[0029] 5.2. For y i ′ performs discrete Fourier transform to obtain D channel signals y ρ (m),ρ=1,2,…,D; m=1,2,…,M+N-1:

[0030]

[0031] Step 6: Perform spectrum sensing

[0032] Perform spectrum sensing on each channel and calculate the total energy E of the ρth channel signal ρ :

[0033]

[0034] where conj(y ρ (m)) means taking yρ The complex conjugate of (m).

[0035] Set the energy threshold λ to determine the channel condition: if E ρ <λ, then the channel ρ is a hole channel; if E ρ ≥λ, then the channel ρ is a non-hole channel.

[0036] Step 7: Energy Normalization

[0037] For non-hole channels, intercept each y ρ (m),ρ=1,2,…,n0 consecutive points in the middle of D, get y′ ρ (i), i = 1, 2, ..., n0, where n0 matches the input data length of the BDVNN network model; then y′ ρ (i), i=1,2,…,n0 perform energy normalization processing to obtain the normalized signal o ρ (i)

[0038]

[0039] Step 8: Use the trained BDVNN model to normalize the signal o ρ (i) Perform recognition and classification based on the normalized signal o ρ (i) Classification and recognition results The corresponding complex baseband signal is classified.

[0040] Preferably, the remezord function of the Matlab environment is called to design the FIR low-pass filter.

[0041] Preferably, the passband cutoff frequency ω of the FIR low-pass filter is designed ρ =Fs / (2r), stopband cutoff frequency ω st =Fs / 2, where the sampling frequency Fs=500KHz, the low-pass filter rectangular coefficient r=1.56; the passband ripple δ1 of the FIR low-pass filter is 0.001, the stopband ripple δ2=0.001, and the filter order t=207.

[0042] Preferably, in the BDVNN network model, the parallel structure A is composed of a convolution layer with a convolution kernel size of 1×7, a convolution layer with a convolution kernel size of 1×3, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel size of 1×1 and an expansion rate of 2, and a convolution layer with a convolution kernel size of 1×3 and an expansion rate of 2, which are connected in sequence to extract coarse features; the parallel structure B is composed of a convolution layer with a convolution kernel size of 1×19, a convolution layer with a convolution kernel size of 1×11, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel size of 1×5 and an expansion rate of 2, and a convolution layer with a convolution kernel size of 1×9 and an expansion rate of 2, which are connected in sequence to extract fine features; the serial structure C is composed of a convolution layer with a convolution kernel size of 1×3 and a 2×2 maximum pooling layer, which are connected in sequence to fuse feature information; the serial structure D is composed of 2 LSTM layers to extract time series features.

[0043] Preferably, the data set for training the BDVNN network model contains 11 signal modulation types: WBFM, DSB-AM, SSB-AM, BPSK, CPFSK, GFSK, PAM, 16-QAM, 64-QAM, QPSK and 8PSK, and each signal sample contains 128 sampling points, represented by X = [X1, X2, ..., X 128 ], and the energy is normalized by the following formula

[0044]

[0045] Among them, X i is the i-th sampling point, conj(X i ) means taking X i The conjugate complex number of ; the energy-normalized sequence XN is used as the input of the BDVNN network model to train the BDVNN network model.

[0046] Preferably, after the BDVNN network model is trained, a model recognition calling script is constructed in a Python environment.

[0047] More preferably, Matlab converts the normalized signal o ρ (i) The data is transmitted to the Python end. Python converts the dimension of the data format and sends the converted data to the BDVNN model trained in step 2 for recognition and classification. The recognition result is converted into a data format and sent back to Matlab. Matlab converts the normalized signal into ρ (i) Classification and recognition results The corresponding complex baseband signal is classified.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention receives and collects signals in a specific frequency band through an SDR device, separates time-domain aliasing signals using a channelized structure in a Matlab environment of a host computer, performs spectrum sensing on each separated signal according to an energy detection method, completes BDVNN recognition network model training through Python, and finally uses the trained BDVNN model to implement cross-platform style recognition of time-domain aliasing signals through the interaction of Matlab and Python. Only less signal preprocessing time and memory overhead are required to achieve rapid and accurate recognition of time-domain mixed signals, with good real-time performance and high accuracy. Moreover, the present invention overcomes the mismatch between Matlab environment data acquisition and Python environment model development, significantly simplifies the development and verification cycle, avoids the tedious work of rebuilding and training models on the Matlab side, improves overall development efficiency and reduces costs, and provides a reliable and efficient solution for rapid development and verification of real-time recognition functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the present invention;

[0051] Figure 2 is the amplitude-frequency response diagram of the FIR low-pass filter in the present invention;

[0052] Figure 3 It is a structural schematic diagram of the BDVNN network model in the present invention;

[0053] Figure 4 This is a schematic diagram of the structure of the biological visual pathway;

[0054] Figure 5 This is a comparison chart of the recognition accuracy of the BDVNN network model in the present invention and the MCLDNN, CNN+LSTM and other models;

[0055] Figure 6 is a power spectrum diagram of a complex baseband signal in an embodiment of the present invention;

[0056] Figure 7 It is a channel allocation diagram of the separated D-channel multiplexed signal in the present invention;

[0057] Figure 8 is a total energy distribution diagram of each channel in an embodiment of the present invention;

[0058] Fig. 9 Schematic diagram of the power spectrum of the complex baseband signal and its identification results in each channel in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0060] In this embodiment, the baseband sampling rate of the adopted channelized receiver (receiving device) is 8 MHz, the number of channels D=16, and the bandwidth of each channel is 500 kHz.

[0061] like Figure 1 As shown, the present invention provides a channelized intelligent receiving method based on BDVNN, which is specifically as follows:

[0062] Step 1: Construct the polyphase filter coefficient matrix

[0063] 1.1. Design a FIR low-pass filter with a passband cutoff frequency ω ρ =Fs / (2r), where sampling frequency Fs = 500KHz, low-pass filter rectangular coefficient r = 1.56, stopband cutoff frequency ω st =Fs / 2, passband ripple δ1=0.001, stopband ripple δ2=0.001, filter order t=207, call the remezord function of Matlab environment to design the filter and get the FIR low-pass filter coefficient vector h=[h1,h2,···,h t+1 ], its amplitude-frequency response is as follows Figure 2 shown.

[0064] 1.2. Perform polyphase decomposition on the designed FIR low-pass filter and obtain:

[0065] h i,n =h nD+i-D ,i=1,2,…,D; n=1,2,…,N

[0066] in D is the number of channels, an even number of channels is taken and N≥1 is satisfied. It means rounding down, and finally the decomposed polyphase filter coefficient matrix H is as follows:

[0067]

[0068] Step 2: Build and train the BDVNN network model

[0069] The structure of the BDVNN network model is as follows Figure 3As shown in the figure, the input data is respectively subjected to preliminary feature extraction by parallel structures A and B. The output features of parallel structures A and B are merged and then passed through serial structures C and D. Finally, they are passed through a fully connected layer (FC) and a Softmax layer to obtain the classification probability value, and the class with the largest classification probability value is taken as the judgment category. The parallel structure A is composed of a convolution layer (Conv) with a convolution kernel size of 1×7, a convolution layer with a convolution kernel size of 1×3, a 2×2 maximum pooling layer (Maxpooling), a convolution layer with a convolution kernel size of 1×1 and a dilation rate of 2 (AtrousConv), and a convolution layer with a convolution kernel size of 1×3 and a dilation rate of 2, which are connected in sequence to extract large-scale coarse features; the parallel structure B is composed of a convolution layer with a convolution kernel size of 1×19, a convolution layer with a convolution kernel size of 1×11, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel size of 1×5 and a dilation rate of 2, and a convolution layer with a convolution kernel size of 1×9 and a dilation rate of 2, which are connected in sequence to extract small-scale fine features. The serial structure C is composed of a convolution layer with a convolution kernel size of 1×3 and a 2×2 maximum pooling layer connected in sequence, which is used to fuse feature information and further extract high-level features. The serial structure D is composed of 2 LSTM layers, which is used to extract time series features. The BDVNN network structure is a bionic Figure 4 The design is based on the biological visual pathway structure shown.

[0070] The BDVNN network model has better performance than MCLDNN, CNN+LSTM and other models, such as Figure 5 As shown. Relying on Python's rich library ecology, powerful model building and training capabilities, in the Python environment Figure 3 The model is built with the structure shown in the figure, and the open source dataset RadioMl2016.10a is used to train the model. The dataset contains 11 signal modulation types: WBFM, DSB-AM, SSB-AM, BPSK, CPFSK, GFSK, PAM, 16-QAM, 64-QAM, QPSK and 8PSK, with a total of 220,000 signal samples. Each signal sample contains 128 sampling points, expressed as X = [X1, X2, …, X 128 ], and the energy is normalized by the following formula

[0071]

[0072] Among them, X i is the i-th sampling point, conj(X i ) means taking X iThe conjugate complex number of; the energy-normalized sequence XN is used as the input of the BDVNN network model to train the BDVNN network model; compared with the input forms such as spectrum diagram, constellation diagram, time-frequency diagram, etc., the energy-normalized sequence of 128 sampling points greatly reduces the complexity of data preprocessing and reduces the requirements of the receiving device for the number of sampling points, which helps to improve the operating efficiency of the receiving device and enhance the real-time processing capability of the receiving device. The complexity of various data preprocessing methods is shown in Table 1. It can be seen that the use of energy-normalized sequences greatly reduces the data preprocessing time and has less memory overhead. It relies on the powerful multi-scale feature extraction capability of the BDVNN network model to achieve better results. Compared with MCLDNN, VGG, CNN+LSTM and other network models with excellent performance, BDVNN has fewer parameters and floating-point operations, and the training time is shorter. The complexity of each network model is shown in Table 2. After this step, the trained BDVNN model is obtained, and the model recognition call script is constructed in the Python environment.

[0073] Table 1 Complexity of data preprocessing methods

[0074] Preprocessing method 100 signal processing time (s) Memory overhead (MB) Energy Normalization Sequence 0.0012 0.1 Spectrum 9.2147 117.1875 EEMD decomposition sequence 117.7814 0.98 Spectral Amplitude Sequence 0.0009 0.1 Time-frequency diagram (wavelet transform) 0.3688 6.25 Constellation map 9.1259 117.1875

[0075] Table 2 Model complexity

[0076] Network Model Parameter quantity (M) Floating point operations (GFLOPs) Total training time(s) BDVNN 0.197 0.141 828 MCLDNN 0.387 0.287 1750 VGG 0.214 0.153 570 CNN+LSTM 0.337 0.241 860

[0077] Step 3: Data Collection

[0078] Take the arrival signal as a mixture of QPSK, GFSK and SSB-AM as an example, where QPSK is on channel 1, GFSK is on channel 4, and SSB-AM is on channel 13. After receiving the RF signal, the (channelized receiver) outputs a complex baseband signal (to the PC), and its power spectrum is as follows Figure 6 As shown, there are three sub-channels in the entire frequency band with obvious amplitude protrusions, corresponding to QPSK, GFSK and SSB-AM signals respectively.

[0079] Step 4: Perform complex baseband signal processing in Matlab environment

[0080] 4.1. Decimating the complex baseband signal by 2 to reduce the signal sampling rate, we get r = [r1, r2, ···, r n ],r k =I k +jQ k , k = 1, 2, ..., n, complete the preliminary processing of the complex baseband signal, where n is the number of sampling points, I k is the real part of the complex baseband signal, Q k is the imaginary part of the complex baseband signal, and j is the imaginary unit.

[0081] 4.2. For r = [r1, r2, ···, r n ] to perform serial-to-parallel conversion, the results are as follows:

[0082] r i,s =r sD+i-D ,i=1,2,…,D; s=1,2,…,M

[0083] Then we get the matrix R:

[0084]

[0085] in D is the number of channels.

[0086] Step 5: Perform channelized filtering in Matlab environment

[0087] 5.1. Expand the matrix R to get R′:

[0088]

[0089] where r i, ' l The definition is as follows:

[0090]

[0091] Then we get the output y′ of the i-th row i =(y′ i,1 ,y′ i,2 ,,…,y′ i,m ), i=1,2,…,D; m=1,2,…,M+N-1, where y i,m The expression is as follows:

[0092]

[0093] Where j represents an imaginary number.

[0094] 5.2. For y i ′ performs discrete Fourier transform to obtain D channel signals y ρ (m),ρ=1,2,…,D; m=1,2,…,M+N-1:

[0095]

[0096] After this step is completed Figure 1 The mid-channelization filtering function of the mid-channelization filtering module separates the time-domain aliased signal into D-channel complex signals. The channel division corresponding to each output signal is as follows: Figure 7 As shown: According to the digital frequency from 0 to 2π, there are D channels in sequence, and the bandwidth of a single sub-channel is The center frequency of each subchannel is Figure 7 The vertical axis is the filter frequency response.

[0097] Step 6: Perform spectrum sensing in Matlab environment

[0098] Perform spectrum sensing on each channel and use energy detection to calculate the total energy E of the ρth channel signal ρ :

[0099]

[0100] where conj(y ρ (m)) means taking y ρ In this embodiment, the energy threshold λ is set to 0.02 to determine the channel status:

[0101] like

[0102] E ρ <λ

[0103] Then the channel ρ is a hole channel, and no signal exists;

[0104] like

[0105] E ρ ≥λ

[0106] Then the channel ρ is a non-hole channel and there is a signal.

[0107] like Figure 8 As shown in the figure, it is the total energy distribution diagram of each channel in this embodiment. According to the set energy threshold, it is determined that there are signals in channels 1, 4 and 13, and there are no signals in the other channels. After this step, Figure 1 The spectrum sensing function can distinguish channels into hole channels and non-hole channels.

[0108] Step 7: Energy Normalization

[0109] For non-hole channels, intercept each y ρ (m),ρ=1,2,…,n0 consecutive points in the middle of D, get y′ ρ (i), i = 1, 2, ..., n0, where n0 matches the input data length of the BDVNN network model (i.e., the number is equal), and in this embodiment n0 = 128; then y′ ρ (i), i=1,2,…,n0 perform energy normalization processing to obtain the normalized signal o ρ (i)

[0110]

[0111] Step 8: Identify and classify complex baseband signals

[0112] Since the receiving device for data collection relies on the Matlab environment, and the BDVNN network model relies on the Python environment, in order to achieve cross-platform real-time recognition function, the two platforms are connected through Figure 1 The data interaction module is used to exchange data. The data interaction module is based on Matlab's script call support for the Python environment to achieve data format conversion and transmission functions between Matlab and Python.

[0113] Matlab calls the model recognition script in Python to convert the normalized signal o ρ (i) The data is transmitted to the Python side. The Python side data exchange module performs a simple dimension conversion on the data format to adapt to the data format requirements of different platforms, and sends the converted data to the BDVNN model trained in step 2 for recognition and classification. The recognition result is further converted into a data format through the data exchange module and transmitted back to Matlab. Matlab uses the normalized signal o ρ (i) Classification and recognition results The corresponding complex baseband signal is classified.

[0114] In this embodiment, the complex baseband signal output by Matlab and its identification results in each channel are as follows: Fig. 9 As shown, channel 1 is determined as QPSK, channel 4 is determined as GFSK, channel 13 is determined as SSB-AM, and the remaining channels are determined as Zeros (i.e., hole channels), indicating that no signal exists, and all are determined correctly.

[0115] By repeating steps 3 to 8, real-time and continuous monitoring of channel conditions can be achieved.

Claims

1. A channelized intelligent receiving method based on BDVNN, characterized in that: the specific steps are as follows: Step 1: Construct the polyphase filter coefficient matrix 1.

1. Design a FIR low-pass filter and obtain the FIR low-pass filter coefficient vector h = [h 1 ,h 2 ,···,h t+1 ]; 1.

2. Perform polyphase decomposition on the designed FIR low-pass filter and obtain: h i,n =h nD+i-D ,i=1,2,…,D;n=1,2,…,N in D is the number of channels, t is the filter order of the FIR low-pass filter, an even number of channels is taken and N≥1 is satisfied. It means rounding down, and finally the decomposed polyphase filter coefficient matrix H is as follows: Step 2: Build and train the BDVNN network model. The structure of the BDVNN network model is as follows: the input data is extracted through parallel structures A and B respectively. The features of parallel structures A and B are merged and then passed through serial structures C and D. Finally, a fully connected layer and a Softmax layer are passed to obtain the classification probability value. The class with the largest classification probability value is taken as the judgment category. Step 3, data acquisition: after receiving the RF signal, the channelized receiver outputs a complex baseband signal; Step 4: Perform complex baseband signal processing 4.

1. Decimating the complex baseband signal by 2 times, we get r = [r 1 ,r 2 ,···,r n ],r k =I k +jQ k ,k=1,2,…,n, where n is the number of sampling points, I k is the real part of the complex baseband signal, Q k is the imaginary part of the complex baseband signal, j is the imaginary unit; 4.

2. For r = [r 1 ,r 2 ,···,r n ] to perform serial-to-parallel conversion, the results are as follows: r i,s =r sD+i-D ,i=1,2,…,D;s=1,2,…,M Then we get the matrix R: in D is the number of channels, Indicates rounding down; Step 5: Channelize and filter the complex baseband signal processed in step 4 5.

1. Expand the matrix R to get R′: where r′ i,l The definition is as follows: Then we get the output y′ of the i-th row i =(y′ i,1 ,y′ i,2, ,…,y′ i,m ), i=1,2,…,D; m=1,2,…,M+N-1, where y i,m The expression is as follows: Where j represents an imaginary number; 5.

2. For y i ′ Perform discrete Fourier transform to obtain D channel signals y ρ (m),ρ=1,2,…,D; m=1,2,…,M+N-1: Step 6: Perform spectrum sensing Perform spectrum sensing on each channel and calculate the total energy E of the ρth channel signal ρ : where conj(y ρ (m)) means taking y ρ The complex conjugate of (m); Set the energy threshold λ to determine the channel condition: if E ρ <λ, then the channel ρ is a hole channel; if E ρ ≥λ, then the channel ρ is a non-hole channel; Step 7: Energy Normalization For non-hole channels, intercept each y ρ (m),ρ=1,2,…,D continuous n 0 points, and get y′ ρ (i),i=1,2,…,n 0 , where n 0 Match the input data length of the BDVNN network model; then y′ ρ (i),i=1,2,…,n 0 Perform energy normalization to obtain the normalized signal o ρ (i) Step 8: Use the trained BDVNN model to normalize the signal o ρ (i) Perform recognition and classification based on the normalized signal o ρ (i) Classification and recognition results The corresponding complex baseband signals are classified.

2. According to the channelized intelligent receiving method based on BDVNN according to claim 1, Features: Call the remezord function of the Matlab environment to design a FIR low-pass filter.

3. According to the channelized intelligent receiving method based on BDVNN according to claim 1 or 2, Features: Design the passband cutoff frequency ω of the FIR low-pass filter ρ =Fs / (2r), stopband cutoff frequency ω st =Fs / 2, where the sampling frequency Fs = 500KHz, the low-pass filter rectangular coefficient r = 1.56; the passband ripple δ of the FIR low-pass filter 1 =0.001, stopband ripple δ 2 =0.001, filter order t=207.

4. According to the channelized intelligent receiving method based on BDVNN as claimed in claim 1, Features: In the BDVNN network model, the parallel structure A is composed of a convolution layer with a convolution kernel size of 1×7, a convolution layer with a convolution kernel size of 1×3, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel size of 1×1 and an expansion rate of 2, and a convolution layer with a convolution kernel size of 1×3 and an expansion rate of 2, which are connected in sequence to extract coarse features; The parallel structure B is composed of a convolution layer with a convolution kernel size of 1×19, a convolution layer with a convolution kernel size of 1×11, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel size of 1×5 and a dilation rate of 2, and a convolution layer with a convolution kernel size of 1×9 and a dilation rate of 2, which are connected in sequence to extract fine features; The serial structure C is composed of a convolution layer with a convolution kernel size of 1×3 and a 2×2 maximum pooling layer connected in sequence, which is used to fuse feature information; the serial structure D is composed of 2 LSTM layers, which is used to extract time series features.

5. According to the channelized intelligent receiving method based on BDVNN as claimed in claim 1, Features: The dataset used to train the BDVNN network model contains 11 signal modulation types: WBFM, DSB-AM, SSB-AM, BPSK, CPFSK, GFSK, PAM, 16-QAM, 64-QAM, QPSK, and 8PSK. Each signal sample contains 128 sampling points, expressed as X = [X 1 ,X 2 ,…,X 128 ], and the energy is normalized by the following formula Among them, X i is the i-th sampling point, conj(X i ) means taking X i The conjugate complex number of ; the energy-normalized sequence XN is used as the input of the BDVNN network model to train the BDVNN network model.

6. According to the channelized intelligent receiving method based on BDVNN as claimed in claim 1, Features: After training the BDVNN network model, build a model recognition calling script in the Python environment.

7. According to claim 6, a channelized intelligent receiving method based on BDVNN, Features: Matlab calls the model recognition script in Python to convert the normalized signal o ρ (i) The data is transmitted to the Python end. Python converts the dimension of the data format and sends the converted data to the BDVNN model trained in step 2 for recognition and classification. The recognition result is converted into a data format and sent back to Matlab. Matlab converts the normalized signal into ρ (i) Classification and recognition results The corresponding complex baseband signals are classified.