Deep learning bandwidth estimation method integrated with spectrum priori knowledge

Abstract

The invention discloses a deep learning bandwidth estimation method integrated with spectrum priori knowledge, and belongs to the field of wireless communication signal parameter estimation. The method specifically comprises the following steps: firstly, collecting simulation signals of a transmitting end and a receiving end as a group; for each group of simulation signals, carrying out frequency band width estimation by using MATLAB (Matrix Laboratory); then, obtaining a two-dimensional real number array of each signal through IQ two paths, and writing the two-dimensional real number array into an h5 file in combination with the estimated frequency band width to form a simulation radio frequency signal data set; thirdly, performing spectrum feature extraction on each receiving end simulation signal in the data set to obtain respective frequency domain amplitude and corresponding frequency, and inputting the frequency domain amplitude and the corresponding frequency as a training set into a Roinformer neural network for regression task learning; and finally, transmitting a signal in real time by a transmitting end, directly inputting the signal into the learned Roinformer neural network, and directly outputting a corresponding bandwidth estimation value. The error rate in the bandwidth estimation task is remarkably reduced, and the method is suitable for the field of communication signal processing and spectrum analysis and has wide application value.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication signal parameter estimation, and specifically is a deep learning bandwidth estimation method incorporating prior knowledge of spectrum. Background Art

[0002] In non-cooperative communications, accurate estimation of signal bandwidth is crucial to obtain enemy communication content by monitoring the electromagnetic spectrum. Bandwidth estimation is a key step in signal processing and a prerequisite for signal analysis and processing. According to Shannon's theorem, the maximum data transmission rate of a channel is related to its bandwidth: the larger the bandwidth, the higher the channel capacity and the higher the data transmission rate.

[0003] By accurately estimating bandwidth, communication systems can optimize their design within limited frequency resources, thereby ensuring that data is transmitted at the best rate. Especially in wireless communications, each system needs to share limited spectrum resources, so the accuracy of bandwidth estimation helps reduce interference between signals, reasonably allocate frequency resources, and improve system stability and efficiency. In addition, bandwidth estimation can effectively improve power and spectrum efficiency, and reasonably adjust modulation and coding methods to optimize transmission efficiency.

[0004] Traditional bandwidth estimation methods, including autocorrelation analysis, spectral entropy method and instantaneous frequency method, rely on the prior model and statistical characteristics of the signal. However, in actual communication environments, the signal types are diverse and are affected by noise, interference and nonlinear distortion, making it difficult for traditional methods to obtain accurate estimates in complex situations.

[0005] In addition, although pure deep learning methods can automatically extract signal features and are suitable for complex signal environments, they are highly dependent on large amounts of data and lack the ability to generalize to different scenarios. Summary of the invention

[0006] In view of the above problems, the present invention proposes a deep learning bandwidth estimation method that incorporates spectrum prior knowledge. First, a simulated RF signal data set is generated. Then, the spectrum prior features are extracted, and a Roinformer network is constructed to perform regression task learning on the extracted spectrum features. By introducing spectrum-related prior knowledge, the present invention can not only automatically learn the characteristics of complex signals, but also maintain a high generalization in a variety of environments, thereby achieving more stable and accurate results in bandwidth estimation.

[0007] The deep learning bandwidth estimation method incorporating spectrum prior knowledge comprises the following specific steps:

[0008] Step 1: Given the signal parameters of the transmitter, a simulation signal of the transmitter is obtained through digital modulation; the receiver samples the transmitter by setting the sampling rate to obtain the simulation signal of the receiver, and the simulation signals of the transmitter and the receiver are taken as a group;

[0009] The parameters of the transmitting signal include bandwidth, roll-off factor and modulation method.

[0010] The digital modulation method selects seven common digital modulation methods, including BASK, QASK, BFSK, QFSK, BPSK, QPSK and 16QAM.

[0011] Step 2: For each group of transmitting end and receiving end simulation signals, use MATLAB to estimate the bandwidth respectively, and save each group of simulation signals with qualified signal bandwidth;

[0012] The bandwidth of a set of simulated signals at the transmitting end and the receiving end is estimated respectively, specifically:

[0013] First, for the simulation signal x[n] at the transmitter or receiver, its length is N and the sampling frequency is f s , the estimate of the power spectral density is expressed as:

[0014]

[0015] Where X(f) is the discrete Fourier transform of the simulation signal x[n]:

[0016]

[0017] In the formula, f n is the frequency value of the nth discrete frequency component in the frequency domain.

[0018] Then, the total power P of the signal is calculated using the power spectral density estimate within the start and end frequency range of the simulated signal spectrum. total :

[0019]

[0020] Where f1 and f2 are the start and end frequencies of the signal spectrum, respectively.

[0021] Finally, when the total signal power reaches 90%, the low-frequency and high-frequency boundary ranges of the simulated signal spectrum [f L ,f H ] within the bandwidth B;

[0022] The following conditions are met:

[0023]

[0024] B=f H -f L

[0025] In the formula, f L and f HIt is the low-frequency and high-frequency boundary when the signal power is accumulated to 90%.

[0026] The bandwidth estimation result is judged: when the relative difference between the bandwidth estimation value of the transmitting end and the bandwidth estimation value of the receiving end (based on the bandwidth estimation value of the transmitting end) exceeds 20%, there is a function estimation error in the generation and calculation process of the group of signals, and the group of data is not saved. Otherwise, the data quality is considered good, the inspection passes and the group of data is saved.

[0027] Step 3: Pass each group of qualified transmitter and receiver simulation signals through the IQ paths respectively to obtain their own two-dimensional real number arrays, and write them into the h5 file in combination with their estimated bandwidths to form a simulated RF signal data set;

[0028] The real part and the imaginary part of the one-dimensional complex array of the transmitter signal data in the current group of simulation signals are respectively split into an I path and a Q path for storage to generate a two-dimensional real array.

[0029] The same operation is performed on the signal data at the receiving end. The real part and the imaginary part of the one-dimensional complex array are respectively split into I path and Q path for storage, thereby generating a two-dimensional real array.

[0030] Step 4, extracting spectrum features of each receiving end simulation signal in the simulated RF signal data set to obtain respective frequency domain amplitudes and corresponding frequencies;

[0031] The specific extraction process is:

[0032] 1) Combining the two-dimensional real number array of the current receiving end simulation signal s into a complex signal x′(n);

[0033] The complex signal x′(n) is represented as follows:

[0034] x′(n)=I(n)+jQ(n)

[0035] In the formula, I(n) is the first dimension data in the two-dimensional real number array, corresponding to the real part of the signal, and Q(n) is the second dimension data in the two-dimensional real number array, corresponding to the imaginary part of the signal.

[0036] 2) Perform a fast Fourier transform on the complex signal x′(n) to convert it from the time domain to the frequency domain to obtain the complex representation of the frequency X(k);

[0037] Its formula is expressed as:

[0038]

[0039] Where x′(n) is the complex sequence of the time domain signal, X(k) is the complex sequence in the frequency domain, and M is the number of FFT points.

[0040] 3) Calculate the amplitude |X(k)| in the complex representation X(k) of the frequency;

[0041] The calculation formula is:

[0042]

[0043] Where Re(X(k)) and Im(X(k)) are the real and imaginary parts of the complex representation of the frequency X(k), respectively.

[0044] 4) Convert the amplitude |X(k)| to a logarithmic scale to obtain the frequency domain amplitude A of the simulated signal s dB ;

[0045] Its formula is expressed as:

[0046] A dB =20·log 10 |X(k)|

[0047] Step 5: The frequency domain amplitude and frequency of each signal are used as training sets and input into the Roinformer neural network for regression task learning. The estimated signal bandwidth of each signal is output and compared with the estimated bandwidth written into the h5 file. When the error is within the threshold range, it means that the Roinformer neural network learning is completed.

[0048] Specifically:

[0049] First, the Roinformer neural network consists of multiple informer modules connected by convolution, pooling and downsampling operations. Each informer module contains the ProbSparse self-attention mechanism and distillation operation.

[0050] The process of ProbSparse self-attention mechanism is expressed as:

[0051]

[0052] In the formula, Indicates that the dimension of the query matrix Q in the real field is L Q ×d; Indicates that the dimension of the key matrix K in the real number field is L K ×d; The dimension of the value matrix V in the real number field is L V ×d; d represents the characteristic dimension (number of columns) of the matrix; d k is the characteristic dimension (number of columns) of the key matrix K.

[0053] Step 6: The transmitter transmits the signal in real time, inputs it into the learned Roinformer neural network, and directly outputs the corresponding bandwidth estimation value.

[0054] The advantages of the present invention are:

[0055] (1) The present invention simulates and generates multi-modulation communication signals with different signal bandwidths through a parameterized method, ensuring that the data set covers a wide range of practical application scenarios while ensuring the richness of signal samples. The baseband simulation signal is generated at the transmitting end, and the signal is uniformly resampled and zero-filled at the receiving end. The efficient storage method of the h5 file format is combined to ensure data consistency, thereby improving the generalization ability of the deep model.

[0056] (2) The present invention utilizes the signal frequency domain power spectrum amplitude information for feature extraction, directly captures the signal energy distribution, effectively compresses the data, and reduces the redundant information in the time domain data. By integrating the spectrum prior knowledge, the frequency domain power spectrum amplitude and frequency are used as the input of the model to perform deep learning bandwidth estimation, so that the model can learn the bandwidth characteristics more efficiently, thereby improving the learning efficiency and accuracy.

[0057] (3) The present invention adopts the Roinformer neural network, through the multi-head probabilistic sparse self-attention mechanism and distillation operation, to effectively extract the global features of long sequences and significantly reduce the computational complexity. High parallelism accelerates model training and improves the processing ability of long sequence data, which has significant advantages in the regression task of estimating continuous numerical attributes such as signal bandwidth. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a general framework diagram of generating a simulated radio frequency signal data set and constructing a neural network for bandwidth estimation in the present invention;

[0059] Figure 2 It is a frequency domain diagram of the simulated signal of the transmitting end generated in the present invention;

[0060] Figure 3 is a frequency domain diagram of a simulated signal at a receiving end generated in the present invention;

[0061] Figure 4 It is a signal frequency domain diagram of the first 4096 points of the receiving end simulation signal generated in the present invention;

[0062] Figure 5 It is a flow chart of a method for constructing a neural network incorporating prior knowledge of spectrum to perform bandwidth estimation according to the present invention;

[0063] Figure 6 It is a loss function and error rate curve diagram of the bandwidth estimation model training process constructed by the present invention;

[0064] Figure 7 It is a scatter plot of the estimated value and the true value of the bandwidth estimation model constructed by the present invention; DETAILED DESCRIPTION

[0065] The content of the present invention is described in detail below in conjunction with implementation steps and drawings.

[0066] Based on the current situation that bandwidth estimation is easily affected and inaccurate in complex environments, the present invention adopts a deep learning bandwidth estimation method that incorporates spectrum prior knowledge, such as Figure 1 As shown, the specific steps are as follows:

[0067] Step 1: Given the signal parameters of the transmitter, a simulation signal of the transmitter is obtained through digital modulation; the receiver samples the transmitter by setting the sampling rate to obtain the simulation signal of the receiver, and the simulation signals of the transmitter and the receiver are taken as a group;

[0068] The parameters of the transmitting signal include bandwidth, roll-off factor and modulation method.

[0069] The digital modulation method selects seven common digital modulation methods, including BASK, QASK, BFSK, QFSK, BPSK, QPSK and 16QAM.

[0070] Step 2: For each group of transmitting end and receiving end simulation signals, use MATLAB to estimate the bandwidth respectively, and save each group of simulation signals with qualified signal bandwidth;

[0071] The bandwidth of a set of simulated signals at the transmitting end and the receiving end is estimated respectively, specifically:

[0072] First, for the simulation signal x[n] at the transmitter or receiver, its length is N and the sampling frequency is f s , the estimate of the power spectral density is expressed as:

[0073]

[0074] Where X(f) is the discrete Fourier transform of the simulation signal x[n]:

[0075]

[0076] In the formula, f n is the frequency value of the nth discrete frequency component in the frequency domain.

[0077] Then, the total power P of the signal is calculated using the power spectral density estimate within the start and end frequency range of the simulated signal spectrum. total :

[0078]

[0079] Where f1 and f2 are the start and end frequencies of the signal spectrum, respectively.

[0080] Finally, when the total signal power reaches 90%, the low-frequency and high-frequency boundary ranges of the simulated signal spectrum [f L ,f H ] within the bandwidth B;

[0081] Bandwidth B refers to the frequency range [f L ,f H ], the following conditions are met:

[0082]

[0083] B=f H -f L

[0084] In the formula, f L and f H It is the low-frequency and high-frequency boundary when the signal power is accumulated to 90%.

[0085] According to the estimation results, if the relative difference between the estimated value of the transmitting end bandwidth and the estimated value of the receiving end bandwidth (based on the estimated value of the transmitting end bandwidth) exceeds 20%, there is a function estimation error in the generation and calculation process of this group of signals, and this group of data is not saved. If the check passes, the data quality is considered good and the group of data is saved.

[0086] Step 3: Pass each group of qualified transmitter and receiver simulation signals through the IQ paths respectively to obtain their own two-dimensional real number arrays, and write them into the h5 file in combination with their estimated bandwidths to form a simulated RF signal data set;

[0087] The signal data of the transmitting and receiving ends are divided into IQ paths for storage, and the real and imaginary parts of the one-dimensional complex array are taken respectively to generate a two-dimensional real array; each digital modulation method generates an h5 file.

[0088] Specifically:

[0089] The real part and the imaginary part of the one-dimensional complex array of the transmitter signal data in the current group of simulation signals are respectively split into an I path and a Q path for storage to generate a two-dimensional real array.

[0090] The same operation is performed on the signal data at the receiving end. The real part and the imaginary part of the one-dimensional complex array are respectively split into I path and Q path for storage, thereby generating a two-dimensional real array.

[0091] Step 4, extracting spectrum features of each receiving end simulation signal in the simulated RF signal data set to obtain respective frequency domain amplitudes and corresponding frequencies;

[0092] Specifically:

[0093] The specific extraction process is:

[0094] 1) Combining the two-dimensional real number array of the current receiving end simulation signal s into a complex signal x′(n);

[0095] The complex signal x′(n) is represented as follows:

[0096] x′(n)=I(n)+jQ(n)

[0097] In the formula, I(n) is the first dimension data in the two-dimensional real number array, corresponding to the real part of the signal, and Q(n) is the second dimension data in the two-dimensional real number array, corresponding to the imaginary part of the signal.

[0098] 2) Perform Fast Fourier Transform (FFT) on the complex signal x′(n) to convert it from the time domain to the frequency domain to obtain the complex representation of the frequency X(k);

[0099] X(k) contains amplitude and phase information, and its formula is expressed as:

[0100]

[0101] Where x′(n) is the complex sequence of the time domain signal, X(k) is the complex sequence in the frequency domain, and M is the number of FFT points.

[0102] 3) Calculate the amplitude |X(k)| in the complex representation X(k) of the frequency;

[0103] The calculation formula is:

[0104]

[0105] Where Re(X(k)) and Im(X(k)) are the real and imaginary parts of the complex representation of the frequency X(k), respectively.

[0106] 4) Convert the amplitude |X(k)| to a logarithmic scale, i.e., use 20log 10 |X(k)| to obtain the frequency domain amplitude A of the simulated signal s dB ;

[0107] Its formula is expressed as:

[0108] A dB =20·log 10 |X(k)|

[0109] Step 5: The frequency domain amplitude and frequency of each signal are used as training sets and input into the Roinformer neural network for regression task learning. The estimated signal bandwidth of each signal is output and compared with the estimated bandwidth written into the h5 file. When the error is within the threshold range, it means that the Roinformer neural network learning is completed.

[0110] Specifically:

[0111] First, the Roinformer network structure adopts an encoder architecture. The encoder is responsible for mapping the input sequence, that is, the signal spectrum features output by the spectrum feature extraction module, to a continuous representation space. It consists of multiple informer modules. Each informer module contains the ProbSparse self-attention mechanism and distillation operation, which is responsible for extracting features from historical time series. Different attention modules perform feature downsampling after convolution and pooling operations, and extract important features through distillation operations. The process of the ProbSparse self-attention mechanism is expressed as:

[0112]

[0113] In the formula, Indicates that the dimension of the query matrix Q in the real field is L Q ×d; Indicates that the dimension of the key matrix K in the real number field is L K ×d; The dimension of the value matrix V in the real number field is L V ×d; d represents the characteristic dimension (number of columns) of the matrix; d k is the characteristic dimension (number of columns) of the key matrix K. T represents the transpose of the matrix.

[0114] The model performance is comprehensively evaluated by using two evaluation indicators: the mean absolute error loss function (MAE) and the mean absolute percentage error (MAPE). The evaluation indicators are defined as follows:

[0115]

[0116] In the formula, Represents the predicted value, y i represents the true value, and n represents the number of prediction samples.

[0117] Step 6: The transmitter transmits the signal in real time, inputs it into the learned Roinformer neural network, and directly outputs the corresponding bandwidth estimation value.

[0118] Finally, through experimental simulation, the bandwidth estimation effect of the Roinformer network model was obtained, and a comprehensive evaluation of the Roinformer network was conducted.

[0119] Example:

[0120] The overall system framework of the present invention for generating simulation signal data sets and estimating RF signal bandwidth parameters is as follows: Figure 1As shown, it includes two parts: generating a simulated RF signal data set and building a neural network to estimate the RF signal bandwidth. The specific composition and steps are as follows:

[0121] Part I: Simulated RF Signal Dataset Generation

[0122] Step 1: Given the transmitter signal parameters, generate a transmitter simulation signal;

[0123] Specifically:

[0124] First, given the bandwidth, roll-off factor, and modulation method of the transmitter signal, the sampling rate can be calculated according to the formula:

[0125]

[0126] In the formula, f s is the sampling rate, B is the signal bandwidth, sps is the number of samples per symbol, and α is the roll-off factor. Bandwidth range (1×10 5 ,3×10 7 )Hz, the roll-off factor is 0.03;

[0127] The modulation method selects seven common digital modulation methods, including BASK, QASK, BFSK, QFSK, BPSK, QPSK and 16QAM, to generate baseband signals.

[0128] Call the function in MATLAB to generate the transmitter signal. The data format is a one-dimensional complex array with a length of 1024. Use MATLAB to estimate the bandwidth of the corresponding signal that contains 90% of the signal power as the label value.

[0129] The bandwidth of a signal is defined as the range of frequencies that contains a certain percentage of the signal's total power.

[0130] The calculation steps are as follows:

[0131] 1) Calculate the power spectral density (PSD) of the signal, that is, the distribution of power in the frequency domain P xx (f) The power spectral density is calculated based on the discrete Fourier transform (DFT).

[0132] Assume the signal is x[n], its length is N, and the sampling frequency is f s , then the estimate of its power spectral density can be expressed as:

[0133]

[0134] Where X(f) is the discrete Fourier transform of the signal x[n]:

[0135]

[0136] 2) Calculate the cumulative power over the frequency range and find the frequency range where the signal power reaches a specified percentage.

[0137] The accumulated power can be expressed in the form of integration or discrete summation; the total power of the signal P total for:

[0138]

[0139] Where f1 and f2 are the start and end frequencies of the signal spectrum, respectively.

[0140] 3) Determine the frequency range that contains 90% of the total power. The difference between the frequency intervals is the bandwidth of the signal.

[0141] Bandwidth B refers to the frequency range [f L ,f H ], the following conditions are met:

[0142]

[0143] B=f H -f L

[0144] In the formula, f L and f H It is the low-frequency and high-frequency boundary when the signal power is accumulated to 90%.

[0145] Digital modulation is to move the spectrum of the digital baseband signal to a high frequency to form a frequency band signal suitable for transmission in the communication channel. The general expression of the modulated signal received by the communication system is:

[0146]

[0147] Where A(t) is the instantaneous amplitude, f c is the carrier frequency, f(t) is the modulation frequency, is the modulation phase, θ is the initial phase, and e(t) is the noise.

[0148] Amplitude keying is a modulation method in which the amplitude of the sinusoidal carrier changes with the digital baseband signal. The time domain expression of M-ary amplitude keying MASK is:

[0149]

[0150] In the formula, a i is an M-ary amplitude sequence, with values ​​of {(2m-M-1), m=1,2,...,M}. i ={1,-1}; a in QASK signal i={-3,-1,1,3}.

[0151] Frequency shift keying is a modulation method that changes the frequency of a sinusoidal carrier with the digital baseband signal. In an M-ary frequency shift keying system, the carrier has M frequency values. The time domain expression of MFSK is:

[0152]

[0153] Where, T s is the code element width, f c is the carrier frequency, f i ∈{f1,f2,f3,...f M}, BFSK modulation signal M = 2, f i There are two values; QFSK signal M = 4, f i There are four possible values.

[0154] Phase shift keying is a modulation method that changes the phase of a sinusoidal carrier with the digital baseband signal. In an M-ary phase shift keying system, the phase of the carrier has M values. In an MPSK signal, each code element consists of log2M digital bits. The time domain expression of an MPSK modulated signal is:

[0155]

[0156] In the formula, The carrier is at t = iT s The phase of the moment, and Where θ is the initial phase. BPSK modulated signal i={0,1}, carrier phase QPSK modulated signal, i = {0, 1, 2, 3}, carrier phase

[0157] The simulated signal spectrum at the transmitter is as follows: Figure 2 shown.

[0158] Step 2, given the receiving end signal sampling rate, generating a receiving end simulation signal;

[0159] Specifically:

[0160] First, the sampling rate of the receiving end is 100MHz, which satisfies the Nyquist theorem. Use MATLAB to resample the transmitting end signal. s Switch to the new sampling rate f′ s .

[0161]

[0162] Where p and q are the upsampling and downsampling ratios, and the sampling rate is changed to the original sampling rate. times; the calculation steps are as follows:

[0163] 1) Upsample the original signal x[n] and interpolate the signal to a higher sampling rate. The interpolation coefficient is p, that is, p-1 samples are inserted between every two adjacent samples. The original signal x[n] is interpolated to x up [n]. The interpolation process is achieved by inserting zeros between the original signal samples and combining it with an interpolation filter for smoothing, namely:

[0164]

[0165] 2) Since the signal may contain high-frequency components (frequencies higher than half of the new sampling rate) before downsampling, aliasing problems may occur. A low-pass filter needs to be applied to suppress the high-frequency components. up [n] is low-pass filtered to eliminate high-frequency components. The filter h[n] is a filter with a cutoff frequency of The ideal low-pass filter. The filtered signal x filt [n] is represented by:

[0166]

[0167] 3) Down-sample the filtered signal and retain one sample every q samples, thereby reducing the sampling rate.

[0168] y[n]=x filt [nq]

[0169] The spectrum of the simulated signal at the receiving end is as follows: Figure 3 As shown. Since the signal lengths of different transmitters are the same and the sampling rates are different, the lengths of the signals at the receiving end are different after being resampled at the same receiving end sampling rate. The first 4096 points are uniformly taken, and zeros are added if the length is less than 4096. The signal data format of the receiving end is a one-dimensional complex array with a length of 4096. The signal frequency domain diagram of the receiving end simulation signal with the first 4096 points is shown as follows Figure 4 shown.

[0170] Step 3, estimating the bandwidth of the generated signal to determine whether the simulated signal is qualified;

[0171] Specifically:

[0172] First, use MATLAB to estimate the bandwidth of the simulated signal at the transmitter and the simulated signal at the receiver, and make a judgment based on the estimated results. If the difference between the estimated values ​​is too large, there is a problem of function estimation error in the generation and calculation process of this group of signals, and this group of data is not saved. If the check passes, the data quality is considered good and the data is saved.

[0173] The code is as follows:

[0174]

[0175] Step 4: Write qualified signal data into h5 file to generate simulation signal data set;

[0176] Specifically:

[0177] First, write the data into the h5 file. In modern communication systems, IQ two-way transmission of modulated signals is a common technology. IQ modulation transmits two signals on one frequency band at the same time, thereby improving spectrum efficiency. The modulated signal s(t) is shown below:

[0178]

[0179] Where I(t) and Q(t) are the in-phase and quadrature components of the signal, respectively. c is the carrier frequency, θ0 is the initial phase, and T is the signal duration.

[0180] The signal data at the receiving end is divided into IQ paths for storage, and the real and imaginary parts of the one-dimensional complex array are taken respectively to generate a two-dimensional real array with a length of 4096 and a format of (4096,2). Each digital modulation method generates an h5 file, and a single file contains 50,000 data.

[0181] Part II: Building a Spectral Feature Extraction Neural Network for Bandwidth Estimation

[0182] Step 1: construct a spectrum feature extraction module to extract spectrum features from the original time domain signal;

[0183] Specifically:

[0184] First, although learning can be performed by directly inputting time-domain IQ signals, the time-domain information may not directly reflect the spectral characteristics of the signal. The bandwidth is closely related to the distribution of spectral energy, and it is difficult for the time-domain IQ signal to directly capture the frequency characteristics of the signal. At the same time, the time-domain signal contains a lot of redundant information, and direct input may increase the complexity of the model and reduce the learning efficiency.

[0185] Frequency domain amplitude information is an effective feature for bandwidth estimation, which can directly reflect the energy distribution of the signal at each frequency, which is crucial for identifying bandwidth. FFT converts the signal into the frequency domain, which can reveal the energy distribution of the signal at different frequencies. For bandwidth estimation tasks, frequency features are more valuable than time domain features. The calculation of frequency domain amplitude can be regarded as a further compression and abstraction of signal features. Compared with complex signals or IQ signals directly input into the time domain, amplitude information reduces data redundancy to a certain extent and focuses on the energy characteristics of the signal, which helps the neural network learn the characteristics of the signal more efficiently.

[0186] The flow chart of the method for constructing a neural network for bandwidth estimation is as follows: Figure 5 As shown, the specific steps are as follows:

[0187] 1) Combine the IQ paths of the training set input signal into a complex signal. The complex signal x(n) is represented as follows:

[0188] x(n)=I(n)+jQ(n)

[0189] 2) Performing Fast Fourier Transform (FFT) on the complex signal in the time domain can convert the signal from the time domain to the frequency domain.

[0190] FFT is an efficient discrete Fourier transform (DFT), which utilizes the symmetry and periodicity of the rotation factors in the DFT algorithm. By continuously decomposing long sequences into shorter sequences, it effectively reduces the number of multiplications and additions and the number of rotation factors, and reduces the amount of algorithm calculations. Its formula is expressed as:

[0191]

[0192] In the formula, x(n) is the complex sequence of the time domain signal, X(k) is the complex sequence in the frequency domain, and M is the number of FFT points.

[0193] X(k) obtained after FFT is a complex representation of frequency, which contains amplitude and phase information. The amplitude is the distribution of signal energy, while the phase is related to the coherence characteristics of the signal.

[0194] 3) Calculate the amplitude of the FFT result. For bandwidth estimation tasks, the frequency distribution and energy characteristics of the signal are very important. The amplitude information in the frequency domain reflects the energy distribution of the signal at different frequencies and can reveal the bandwidth characteristics of the signal.

[0195] The calculation formula for the amplitude |X(k)| is:

[0196]

[0197] Where Re(X(k)) and Im(X(k)) are the real and imaginary parts of X(k), respectively.

[0198] Convert the signal amplitude to a logarithmic scale, i.e., use 20log 10 |X(k)| is used to represent the signal strength change more intuitively.

[0199] This processing method is often used in signal processing, especially when analyzing the magnitude of frequency domain amplitudes. In practical applications, the amplitude of a signal spans a large dynamic range, and the amplitude difference can reach several orders of magnitude. When using a linear scale to represent the amplitude, these changes are very difficult to display and analyze simultaneously. By taking the logarithm of the amplitude, the dynamic range of the signal can be compressed, so that the amplitudes with large differences can be intuitively displayed and processed, and the energy intensity of each frequency in the frequency domain can be extracted.

[0200] 4) Normalize the bandwidth value of the input signal in the training set and divide it by the maximum bandwidth value in the training set.

[0201] To prevent the data range from being too different, model training will become very slow or even difficult to converge. Normalize and scale the data to a smaller range as the input data label. Perform the same operation on the validation set and test set signals.

[0202] Step 2: construct a Roinformer network to perform regression task learning on the input spectral features, namely, frequency domain amplitude and frequency;

[0203] Specifically:

[0204] First, Roinformer is a deep learning network structure based on a multi-head probabilistic sparse self-attention mechanism. The network structure adopts an encoder architecture, which is responsible for mapping the input sequence into a continuous representation space. It consists of multiple Informer modules, each of which contains a self-attention mechanism and a distillation operation, and is responsible for extracting features from historical time series.

[0205] Unlike traditional Q, K, and V matrix calculation methods, Roinformer compares each moment in the sequence with the uniform distribution and calculates the correlation between that moment and other time points in the time period. The more significant the difference, the higher the correlation of that time point in the interval. Different attention modules perform feature downsampling through convolution and pooling operations, and extract important features through distillation operations.

[0206] Roinformer uses the ProbSparse self-attention mechanism. Traditional self-attention is defined as a probabilistic kernel smoothing method:

[0207]

[0208] The difference between the distribution of query vectors and the uniform distribution can be measured by the KL divergence:

[0209]

[0210] By removing the constant term, the sparsity measure of the i-th query vector can be defined as:

[0211]

[0212] The query matrix Q, key matrix K, and value matrix V are three matrices of the same shape obtained after linear transformation of the same input sequence. The attention weight is obtained by calculating the correlation between Q and K, and the weight is weighted and summed with V to obtain the calculation result of the attention mechanism, which indicates the degree of attention paid to different parts of the input sequence. The process of ProbSparse self-attention mechanism is expressed as:

[0213]

[0214] In the formula, Indicates that the dimension of the query matrix Q in the real field is L Q ×d; Indicates that the dimension of the key matrix K in the real number field is L K ×d; The dimension of the value matrix V in the real number field is L V ×d; d represents the characteristic dimension (number of columns) of the matrix; d k is the characteristic dimension (number of columns) of the key matrix K. T represents the transpose of the matrix.

[0215] The multi-head attention mechanism enables the network to focus on multiple different parts of the input sequence at the same time, thereby effectively capturing the global dependencies in long sequence data. Due to the characteristics of the self-attention mechanism, it has a high degree of parallelism, and the output of each position can be calculated independently, which gives it a clear advantage in processing long sequences.

[0216] ROPE (Rotary Position Embedding) is a position encoding method in natural language processing that introduces relative position information into the self-attention mechanism while retaining the ability to model sequence order. Compared with the traditional absolute position encoding method, ROPE uses a rotation method to embed position information directly into the self-attention calculation. This method is particularly suitable for models with multi-head attention mechanisms, such as Informer networks. ROPE models position information by embedding the position code into the vector space in the self-attention calculation in the form of a rotation matrix.

[0217] After feature extraction, regression tasks are performed to establish a relationship model between features and continuous numerical targets, which is used to predict the numerical targets of unknown samples. Regression tasks are a type of supervised learning task. Unlike classification tasks, the goal of regression tasks is to predict continuous values ​​rather than discrete categories. Regression tasks are often used to estimate continuous numerical attributes of RF signals, such as signal bandwidth, center frequency, and wave direction.

[0218] The difference between regression tasks and classification tasks is that the output dimension is generally a scalar or a one-dimensional vector, and the values ​​are continuous within a certain range, without the need for probability distribution mapping through the Softmax function.

[0219] Step 3: Experimental simulation to obtain the bandwidth estimation effect under different network models and conduct a comprehensive evaluation of each model;

[0220] Specifically:

[0221] 1) Experimental simulation platform

[0222] The experimental simulation platform is the Windows 10 64-bit operating system, the programming environment is Python 3.9, the integrated development environment is PyCharm 2020.3.5, and the model construction is mainly implemented by deep learning frameworks such as Pytorch.

[0223] 2) Model evaluation indicators

[0224] In order to comprehensively evaluate the performance of the model, the mean absolute error loss function (MAE) and the mean absolute percentage error (MAPE), that is, the error rate, are introduced. The smaller the MAE and the error rate, the better the prediction effect of the model.

[0225] MAE and error rate are defined as follows.

[0226]

[0227] In the formula, Represents the predicted value, y i represents the true value, and N represents the number of predicted samples.

[0228] 3) Simulation experiment design

[0229] The experimental simulation proves that the model is effective. The training set contains 50,000 data, the validation set contains 1,000 data, and the test set contains 1,000 data. The batch size is set to 64. The training round is set to 200. The learning rate is learning rate = 1×10 -3 The loss function is the mean absolute error loss function (MAE). The criterion for model convergence is that when the MAE index of the validation set does not decrease for ten consecutive times, the training is stopped and the model is saved.

[0230] 4) Experimental results

[0231] The loss function curve during training and verification is as follows Figure 6 As shown in (a), the error rates in the training set and the validation set are Figure 6(b). The Roinformer and Transformer networks were used respectively, and the experimental results are shown in the following table. The error rate of the Roinformer prediction network's best training batch validation set is 7.09%. The experimental results of different networks are shown in Table 1:

[0232] Table 1

[0233]

[0234] The scatter plot of estimated values ​​and true values ​​in the test set is as follows Figure 7 As shown in the figure, the scatter plot can intuitively show the relationship between the true value and the estimated value, helping to identify the prediction accuracy of the model, potential systematic deviations, nonlinear relationships and outliers. Ideally, the relationship between the estimated value and the true value should be close to linear, that is, the data points should be close to the diagonal of y=x. It can be seen that the data points (blue scatter points) in the picture are closely surrounded by the red dotted line, and the distribution of the data points shows a good linear trend, which means that the error between the estimated value and the true value of the model is very small, and the estimation effect of the model is good. At larger true values ​​(upper right corner) and smaller true values ​​(lower left corner), the degree of agreement between the estimated value and the true value is also consistent, indicating that the model's estimation ability is balanced over the entire data interval.

Claims

1. A deep learning bandwidth estimation method incorporating prior knowledge of spectrum, characterized in that: This includes parameterizing the generation of high-quality RF signal datasets covering a variety of modulation modes and bandwidth ranges, and achieving efficient and accurate bandwidth estimation through spectrum feature extraction and Roinformer network. The specific steps are as follows: Step 1, using a parameterized method to generate diversified transmitter and receiver signal data; Given the signal parameters of the transmitter, the simulation signal of the transmitter is obtained through digital modulation; the receiver samples the transmitter by setting the sampling rate to obtain the simulation signal of the receiver, and the simulation signals of the transmitter and the receiver are taken as a group; Step 2: For each group of transmitter and receiver simulation signals, use MATLAB to estimate the bandwidth respectively, and save each group of simulation signals with qualified signal bandwidth; The bandwidth of a set of simulated signals at the transmitting end and the receiving end is estimated respectively, specifically: First, for the simulated signal x[n] at the transmitter or receiver, its length is N and the sampling frequency is f s , the estimate of the power spectral density is expressed as: Where X(f) is the discrete Fourier transform of the simulation signal x[n]; Then, the total power P of the signal is calculated using the power spectral density estimate within the start and end frequency range of the simulated signal spectrum. total : Where f1 and f2 are the starting and ending frequencies of the signal spectrum respectively; Finally, when the total signal power reaches 90%, the low-frequency and high-frequency boundary ranges of the simulated signal spectrum [f L ,f H ] within the bandwidth B; The following conditions are met: B=f H -f L In the formula, f L and f H It is the low-frequency and high-frequency boundary when the signal power is accumulated to 90%; Step 3: Pass each group of qualified transmitter and receiver simulation signals through the IQ paths respectively to obtain their respective two-dimensional real number arrays, and write them into the h5 file in combination with their respective estimated bandwidths to form a simulated RF signal data set; Step 4: Using fast Fourier transform, extract the spectrum features of each receiving end simulation signal in the simulated RF signal data set to obtain the respective frequency domain amplitude and corresponding frequency; The specific extraction process is: 1) Combining the two-dimensional real number array of the current receiving end simulation signal s into a complex signal x′(n); The complex signal x′(n) is represented as follows: x′(n)=I(n)+jQ(n) In the formula, I(n) is the first dimension data in the two-dimensional real number array, corresponding to the real part of the signal, Q(n) is the second dimension data in the two-dimensional real number array, corresponding to the imaginary part of the signal; 2) Perform a fast Fourier transform on the complex signal x′(n) to convert it from the time domain to the frequency domain to obtain the complex representation of the frequency X(k); Its formula is expressed as: Where M is the number of FFT points; 3) Calculate the amplitude |X(k)| in the complex representation X(j) of the frequency; The calculation formula is: Where Re(X(k)) and Im(X(k)) are the real and imaginary parts of the complex representation of the frequency X(k), respectively; 4) Convert the amplitude |X(k)| to a logarithmic scale to obtain the frequency domain amplitude A of the simulated signal s dB ; Its formula is expressed as: A dB =20·log 10 |X(k)| Step 5: The frequency domain amplitude and frequency of each signal are used as training sets and input into the Roinformer neural network for regression task learning. The estimated signal bandwidth of each signal is output and compared with the estimated bandwidth written into the h5 file. When the error is within the threshold range, it means that the Roinformer neural network learning is completed. Specifically: First, the Roinformer neural network consists of multiple informer modules connected by convolution, pooling and downsampling operations. Each informer module contains the ProbSparse self-attention mechanism and distillation operation. The process of ProbSparse self-attention mechanism is expressed as: In the formula, Indicates that the dimension of the query matrix Q in the real field is L Q ×d; Indicates that the dimension of the key matrix K in the real number field is L K ×d; The dimension of the value matrix V in the real number field is L V ×d; d represents the characteristic dimension of the matrix; d k is the characteristic dimension of the key matrix K; Step 6: The transmitter transmits the signal in real time, inputs it into the learned Roinformer neural network, and directly outputs the corresponding bandwidth estimation value.

2. A deep learning bandwidth estimation method incorporating spectrum prior knowledge as claimed in claim 1, characterized in that: In the step 1, the parameters of the transmitting end signal include bandwidth, roll-off factor and modulation mode; The digital modulation method selects seven common digital modulation methods, including BASK, QASK, BFSK, QFSK, BPSK, QPSK and 16QAM.

3. The deep learning bandwidth estimation method incorporating spectrum prior knowledge as claimed in claim 1, characterized in that: In step 2, the calculation formula of X(f) is as follows: In the formula, f n is the frequency value of the nth discrete frequency component in the frequency domain.

4. The deep learning bandwidth estimation method incorporating spectrum prior knowledge as claimed in claim 1, characterized in that: In the step 2, the bandwidth estimation result is judged: taking the bandwidth estimation value of the transmitting end as a reference, when the relative difference between the bandwidth estimation value of the transmitting end and the bandwidth estimation value of the receiving end exceeds 20%, there is a function estimation error in the generation and calculation process of the group of signals, and the group of data is not saved; Otherwise, the data quality is considered good, the check passes and the group of data is saved.

5. The deep learning bandwidth estimation method incorporating spectrum prior knowledge as claimed in claim 1, characterized in that: The step three is specifically as follows: the real part and the imaginary part of the one-dimensional complex array of the transmitter signal data in the current group of simulation signals are respectively split into an I path and a Q path for storage, thereby generating a two-dimensional real array; The same operation is performed on the signal data at the receiving end. The real part and the imaginary part of the one-dimensional complex array are respectively split into I path and Q path for storage, thereby generating a two-dimensional real array.

Images