Radio signal arrival time estimation method based on deep learning

Abstract

The present invention relates to a method for estimating the arrival time of radio signals based on deep learning, belonging to the field of radio signal processing. The method solves the problem of low accuracy in estimating the arrival time of radio signals in complex electromagnetic environments. The method includes preprocessing simulated complex baseband signals and pure noise data to obtain sample data in the form of an IQ matrix, which is then combined with sample labels to form a signal sample set; using the signal sample set to train an arrival time estimation model, achieving a maximum number of iterations to preserve model parameters with a minimum loss function, and obtaining a trained arrival time estimation model; acquiring radio signals in real time and processing them to obtain corresponding complex baseband signals; intercepting multiple signal segments based on a sliding window of a predetermined length and normalizing them; extracting the real and imaginary parts, stacking them into corresponding multiple IQ matrices, and inputting them into the trained model to obtain an index estimate of the signal starting sampling point; and calculating the signal arrival time based on the signal starting sampling point index estimate. This method improves the accuracy of signal arrival time estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal processing, and in particular to a radio signal arrival time estimation method based on deep learning. Background Art

[0002] Radio signal arrival time estimation (TOA) determines the time it takes for a radio signal to reach a receiver. It's a key technology commonly used in wireless communications and a core enabler for key functions such as positioning, navigation, and wireless ranging. With the rapid development of wireless communications, increasingly complex and diverse application scenarios are placing higher demands on the accuracy and reliability of TOA estimation.

[0003] In traditional wireless communication scenarios, commonly used methods for time-of-arrival (TOA) estimation include peak detection, correlation, and time-frequency analysis. Peak detection determines the signal's arrival time by finding the peak amplitude of the received signal. Its principle is simple and computationally low, but in practice, due to multipath effects, signals propagate along multiple reflection paths, resulting in multiple amplitude peaks in the received signal. This makes it difficult for the peak detection method to accurately determine which peak corresponds to the true signal arrival time, thus reducing estimation accuracy. Furthermore, environmental noise and interference can cause fluctuations in the amplitude of the received signal, further interfering with peak detection.

[0004] The correlation method uses a known signal waveform to correlate with the received waveform, determining the signal arrival time by finding the maximum value of the correlation function. This method places extremely high demands on signal synchronization, requiring precise knowledge of the signal waveform structure to ensure the effectiveness of the correlation operation. In non-cooperative scenarios, such as certain wireless communication interference scenarios or monitoring of unknown signal sources, where prior waveform information of the signal is unavailable, the correlation method struggles.

[0005] Time-frequency analysis transforms signals from the time domain to the time-frequency domain, allowing for clearer observation of the signal's time-frequency characteristics and subsequent time-of-arrival (TOA) estimation. However, it is also susceptible to complex electromagnetic environments. Multipath effects can create complex signal structures in the time-frequency domain, and interference can affect the extraction of time-frequency characteristics, leading to inaccuracies in TOA estimates.

[0006] The increasing complexity of the electromagnetic spectrum environment, such as the intensive use of spectrum resources and the intertwined coexistence of various wireless communication systems and interference sources, poses unprecedented challenges to traditional estimation methods. Methods such as peak detection and time-frequency analysis are susceptible to multipath effects, environmental noise, and interference. Correlation methods, on the other hand, require high signal synchronization and require correlation between the known signal waveform and the received waveform, making them difficult to use in non-cooperative scenarios. Summary of the Invention

[0007] In view of the above analysis, an embodiment of the present invention aims to provide a radio signal arrival time estimation method based on deep learning, so as to solve the technical problems of low radio signal arrival time estimation accuracy and poor generalization ability caused by multipath channels and noise injection in existing methods.

[0008] The purpose of the present invention is mainly achieved through the following technical solutions:

[0009] The present invention provides a method for estimating the arrival time of radio signals based on deep learning, comprising the following steps:

[0010] Preprocessing the complex baseband signal and pure noise data generated by simulation to obtain sample data in the form of an IQ matrix, and forming a signal sample set with the corresponding sample label; wherein the sample label corresponding to the complex baseband signal is the index value of the signal starting sampling point;

[0011] The arrival time estimation model is trained using the signal sample set to obtain a trained arrival time estimation model by saving the model parameters with the minimum loss function after a maximum number of iterations.

[0012] The radio signal is collected in real time and down-converted to obtain a corresponding complex baseband signal. The signal is cut into multiple signal segments based on a sliding window of predetermined length and normalized. The real and imaginary parts are extracted and stacked into corresponding multiple IQ matrices. The matrices are then input into the trained arrival time estimation model in sequence until the index estimate of the signal starting sampling point is found. The signal arrival time is calculated based on the index estimate of the signal starting sampling point and the sequence number of the signal segment.

[0013] Furthermore, the complex baseband signal includes a communication complex baseband signal and a radar complex baseband signal; and the preprocessing of the simulated complex baseband signal and the pure noise data to obtain sample data includes:

[0014] Cutting the complex baseband signal into a plurality of signal segments based on a predetermined length, constructing signal samples containing different arrival time information based on the signal segments to obtain a first sample signal; performing multipath channel simulation and / or noise injection on the first sample signals to obtain a signal set corresponding to the first sample signal;

[0015] The pure noise data has a length of sampling points;

[0016] The signal set corresponding to the first sample signal has a length of The pure noise data of the sampling points are merged into the second sample signal;

[0017] The second sample signal is normalized, and the real and imaginary parts are extracted and stacked into an IQ matrix as sample data.

[0018] Furthermore, constructing signal samples containing different arrival time information based on the complex baseband signal to obtain a first sample signal includes:

[0019] The complex baseband signal is transmitted according to a predetermined length Perform signal truncation to obtain multiple lengths of For each sub-complex baseband signal, a sub-complex baseband signal of length An all-zero sequence of sampling points; where ;

[0020] The intercept length is The all-zero subsequence of the sampling points is spliced ​​to the starting point of the complex baseband signal; the remaining After the all-zero subsequence of sampling points is spliced ​​to the end point of the complex baseband signal, the length of the complex baseband signal is obtained. The first sample signal of the sampling point; wherein, is the number of sampling points of the complex baseband signal, ;

[0021] The sample label corresponding to the complex baseband signal is the index value of the signal starting sampling point ; The sample label corresponding to the pure noise data is .

[0022] Furthermore, the arrival time estimation model includes an input layer, an adaptive correction module, a feature extraction module, a feature focusing module and an arrival time prediction module in sequence;

[0023] The input layer is configured to receive the IQ matrix;

[0024] The adaptive correction module is used to eliminate the offset of the IQ matrix caused by multipath and noise, and output the corrected time domain IQ matrix;

[0025] The feature extraction module is used to capture the fusion features of different scales of the corrected time-domain IQ matrix;

[0026] The feature focusing module is used to select signal features in the time channel based on the self-attention mechanism and output the focused signal features;

[0027] The arrival time prediction module includes neurons for output The probability of the estimated value of the starting sampling point index of the signal and the probability of 1 pure noise data.

[0028] Furthermore, the adaptive correction module sequentially includes a Fourier transform layer, a fully connected frequency domain weight learning layer, and an inverse Fourier transform reconstruction layer;

[0029] The Fourier transform layer is used to transform the IQ matrix from the time domain to the frequency domain to obtain a frequency domain signal;

[0030] The fully connected frequency domain weight learning layer is used to learn the weight of the frequency domain signal, suppress the phase offset component caused by multipath, and attenuate the high-frequency or low-frequency region dominated by noise, so as to obtain a corrected frequency domain signal;

[0031] The inverse Fourier transform reconstruction layer is used to perform inverse fast Fourier transform on the corrected frequency domain signal to restore it to a time domain signal and output an adaptively corrected time domain IQ matrix.

[0032] Furthermore, the feature extraction module includes parallel first, second and third convolutional layers with different receptive fields, and a Concat layer;

[0033] The first convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract local features of the time-domain IQ matrix after the adaptive correction;

[0034] The second convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract the medium-scale features of the time-domain IQ matrix after the adaptive correction;

[0035] The third convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract the global features of the time-domain IQ matrix after the adaptive correction; wherein, ;

[0036] The Concat splicing layer splices the local features, medium-scale features and global features output by the first, second and third convolutional layers along the channel dimension to obtain fused features.

[0037] Furthermore, the feature focusing module includes a self-attention feature focusing layer; wherein the output of the Concat splicing layer is residually connected with the output of the self-attention feature focusing layer;

[0038] The self-attention feature focusing layer includes a self-attention layer, a focused residual connection and a layer normalization in sequence;

[0039] The self-attention layer is used to map the fused features into Q, K, and V matrices through linear transformation; calculate attention weights based on the Q, K, and V matrices, and output weighted fused features;

[0040] The focused residual connection and layer normalization are used to connect and normalize the fused features and the weighted fused features; the output focused features are transferred to the arrival time prediction module.

[0041] Furthermore, the loss function is as follows:

[0042] ;

[0043] in, is the second sample signal; is the normalized second sample signal; 、 are respectively the real part and imaginary part of the normalized second sample signal; is the parameterized function of the arrival time estimation model, which is used to obtain the probability distribution of the model prediction; is a learnable parameter, is the true label corresponding to the input sample, The number of categories output by the model.

[0044] Furthermore, the signal arrival time is calculated based on the estimated value of the signal starting sampling point index and the sequence number of the signal segment. ,as follows:

[0045] ;

[0046] in, is the moment when the receiver collects the signal. is the sampling rate of the receiver; The estimated value of the signal starting sampling point index; For the first time The sequence number of the signal segment with the estimated value of the signal starting sampling point index within .

[0047] Furthermore, the multipath channel simulation includes Rice channel and Rayleigh channel simulation;

[0048] The noise injection includes additive white Gaussian noise, generalized white Gaussian noise and colored noise injection.

[0049] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0050] 1. This invention trains a time-of-arrival estimation model based on the original radio signal and directly learns the signal starting point features without manual feature extraction. This lowers the threshold for use and is applicable to unknown signal types and non-cooperative scenarios. It also solves the problems of traditional methods that rely on manually designed features, require complex signal processing experience, and have poor generalization capabilities.

[0051] 2. This invention introduces Ricean and Rayleigh channel conversions and multiple noise injections into the signal sample sets used for training, generating signal samples that are closer to the real environment. The sample data is more consistent with the actual electromagnetic environment, effectively improving the adaptability and accuracy of signal arrival time estimation in complex electromagnetic spectra. It also solves the problem that the multipath effect and noise injection in existing methods cause the accuracy of signal arrival time estimation to drop significantly.

[0052] 3. The adaptive correction module in the present invention suppresses multipath and noise injection, dynamically corrects the signal through frequency domain weight learning, and suppresses multipath phase offset and noise frequency band. This solves the problem of signal distortion caused by multipath in existing methods, making it difficult to distinguish between the main path and the delayed path.

[0053] 4. The feature extraction module and feature focusing module in this invention are based on multi-scale feature fusion and self-attention mechanism. Multi-scale convolution extracts local, medium and global features in parallel to cover signal mutations. Self-attention focuses dynamically enhances the time channel features near the starting point and suppresses irrelevant areas. This solves the problem of single-scale features missing key information in existing methods.

[0054] 5. The sample data in the present invention covers communication signals and radar signals. Based on multiple modulation types, the time arrival estimation model adaptively learns the characteristics of heterogeneous signals through the feature extraction module, which solves the problem of poor adaptability to specific modulation types in traditional methods.

[0055] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0057] Figure 1 This is a flow chart of a method for estimating the arrival time of a radio signal based on deep learning according to an embodiment of the present invention;

[0058] Figure 2Schematic diagram of the network structure of the time arrival estimation model in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0060] With the rapid development of artificial intelligence (AI), deep learning has been successfully applied to problems such as signal detection, recognition, and direction finding. Without loss of generality, deep learning also provides an effective solution for radio signal arrival time estimation, enabling efficient estimation in complex scenarios. This paper discloses a deep learning-based signal arrival time estimation method, providing important support for the development of modern wireless communications and positioning systems.

[0061] A specific embodiment of the present invention discloses a deep learning signal detection method that adapts to changes in the number of receiving antennas, such as Figure 1 As shown, the following steps are included:

[0062] Step S1: Preprocess the complex baseband signal and pure noise data generated by simulation to obtain sample data in the form of an IQ matrix, and form a signal sample set with the corresponding sample label; wherein the sample label corresponding to the complex baseband signal is the index value of the signal starting sampling point;

[0063] Step S2: using the signal sample set to train the arrival time estimation model, and saving the model parameters with the minimum loss function after reaching the maximum number of iterations, thereby obtaining a trained arrival time estimation model;

[0064] Step S3: Real-time radio signal acquisition is performed, and a corresponding complex baseband signal is obtained after down-conversion processing. The signal is cut into multiple signal segments based on a sliding window of a predetermined length and normalized. The real and imaginary parts are respectively extracted and stacked into corresponding multiple IQ (In-Phase, in-phase; Quadrature, orthogonal) matrices, and the matrices are sequentially input into the trained arrival time estimation model until the signal starting sampling point index estimate is found; the signal arrival time is calculated based on the signal starting sampling point index estimate and the sequence number of the signal segment.

[0065] The IQ matrix is ​​used to represent the real part (in-phase component) and imaginary part (quadrature component) of a complex baseband signal and is often used in radio signal processing.

[0066] Step S1 includes steps S11-S14.

[0067] Based on the knowledge of radio signals, the simulation generates different types of complex baseband signals, including communication complex baseband signals and radar complex baseband signals; used to simulate actual communication scenarios and radar scenarios; and generates pure noise data to simulate noise interference in actual environments.

[0068] The complex baseband signal includes a communication complex baseband signal and a radar complex baseband signal; the simulated complex baseband signal and pure noise data are preprocessed to obtain sample data, including:

[0069] Cutting the complex baseband signal into a plurality of signal segments based on a predetermined length, constructing signal samples containing different arrival time information based on the signal segments, and obtaining a first sample signal;

[0070] performing multipath channel simulation and / or noise injection on the first sample signals respectively to obtain a signal set corresponding to the first sample signals;

[0071] The pure noise data has a length of sampling points;

[0072] The signal set corresponding to the first sample signal has a length of The pure noise data of the sampling points are merged into the second sample signal;

[0073] Specifically, it includes performing multipath channel simulation and noise injection; performing multipath channel simulation and noise injection in sequence (first performing multipath channel simulation, then performing noise injection).

[0074] The second sample signal is normalized, and the real and imaginary parts are extracted and stacked into an IQ matrix as sample data.

[0075] The multipath channel simulation includes Rice channel and Rayleigh channel simulation;

[0076] The noise injection includes additive white Gaussian noise, generalized white Gaussian noise and colored noise injection.

[0077] Step S11: Simulate and generate complex baseband signals of communication signals and radar signals, as well as pure noise data.

[0078] Communication signals are used for information transmission (such as voice and data), representing typical narrowband, high data rate scenarios with diverse modulation methods. Communication signals emphasize spectral efficiency and noise immunity. They are characterized by high signal stability and are suitable for verifying model performance in common scenarios.

[0079] Radar signals are used for target detection and parameter estimation (such as ranging and velocity measurement). They represent broadband, high-time-bandwidth product scenarios and have complex waveforms. Radar signals emphasize resolution and immunity to the Doppler effect. Their wide dynamic range makes them suitable for verifying model robustness under high dynamics and multipath interference.

[0080] The simulation generates different types of complex baseband signals, illustratively including at least communication signals and radar signals. In practical applications, these signals may be increased according to specific needs.

[0081] The joint training arrival time estimation model of two types of signals, communication signals and radar signals, can adapt to a wider range of signal characteristics, avoid overfitting, and enhance adaptability.

[0082] (1) Generate a complex baseband signal corresponding to the communication signal.

[0083] The complex baseband signal corresponding to the communication signal of predetermined length that is continuous in time is simulated based on the randomly generated binary bit stream, as follows:

[0084] Randomly generate binary bit stream ,as follows:

[0085] Formula (1)

[0086] in, is the length of the binary bit stream.

[0087] Complete the mapping of bit stream to symbol according to different modulation types, Bits are grouped together, and if each symbol contains bits, then:

[0088] Formula (2)

[0089] in, is the number of modulation constellation points.

[0090] Different modulation types include QPSK (Quadrature Phase Shift Keying), PSK (Phase Shift Keying), FSK (Frequency Shift Keying), PAM (Pulse Amplitude Modulation), ASK (Amplitude Shift Keying), QAM (Quadrature Amplitude Modulation), and CPM (Continuous Phase Modulation).

[0091] First, the bit stream is grouped into groups based on different modulation types. Consecutive bits are grouped into a group;

[0092] Next, symbol mapping is performed, and each group of bits is mapped to a complex symbol using the Map(.) function. ;

[0093] Finally, the complex baseband signal is generated, and the symbol sequence is pulse-shaped and filtered to generate the baseband signal.

[0094] Exemplarily, for QPSK modulation, ,but , each symbol consists of two bits; for 16-QAM ( =16), , each symbol consists of 4 bits.

[0095] The bit stream is mapped to symbols based on multiple modulation types to obtain corresponding symbol sequences.

[0096] The mapping from bit stream to symbols is as follows:

[0097] Formula (3)

[0098] in, For the general bits are mapped to symbols on the complex plane , For index symbols.

[0099] For example, for QPSK modulation, , then the bit combinations {0,0}, {0,1}, {1,0}, {1,1} are mapped to

[0100] For QPSK modulation, , bitstream

[0101] The symbol mapping is as follows:

[0102] The first symbol {0,1}→ ;

[0103] The second symbol {1,0} → ;

[0104] The third symbol {1,1}→ .

[0105] Each symbol is usually subjected to a pulse shaping filter (such as a raised cosine filter) to generate a time-continuous complex baseband signal.

[0106] Each symbol in the symbol sequence is pulse-shaped filtered to obtain a time-continuous complex baseband signal of a predetermined length.

[0107] After pulse shaping, a time-continuous complex baseband signal is generated ,as follows:

[0108] Formula (4)

[0109] in, is the symbol period, i.e. the duration of each symbol, in seconds; is a rectangular pulse, which is the impulse response of a pulse shaping filter (such as a raised cosine filter), used to limit the signal bandwidth and reduce intersymbol interference; Binary bit stream The mapped symbols.

[0110] For different modulation types, the binary bit stream to symbol mapping method is different; for example, for the K-order PSK modulation type, the bit stream to symbol The mapping is as follows:

[0111] Formula (5)

[0112] in, is the symbol number. Each symbol corresponds to a point on the unit circle on the complex plane, and the rotation angle is ; is an imaginary unit, satisfying , Represents the imaginary part of a complex number, allowing the signal to be represented on the complex plane, facilitating modulation and demodulation operations.

[0113] for FSK modulation type, bit stream to symbol The mapping is as follows:

[0114] Formula (6)

[0115] in, is the carrier frequency, is the frequency interval; is a time variable, which represents the duration of the signal at that frequency.

[0116] for PAM modulation, each symbol is directly mapped to a real number of different amplitudes, bit stream to symbol The mapping is as follows:

[0117] Formula (7)

[0118] For K-order ASK modulation, bit stream to symbol The mapping is as follows:

[0119] Formula (8)

[0120] in, is the carrier amplitude; each symbol sends a carrier with a different amplitude.

[0121] For K-order QAM modulation type, bit stream to symbol The mapping is as follows:

[0122] Formula (9)

[0123] in, is the row and column number; each symbol point is distributed on a two-dimensional rectangular grid, and both amplitude and phase are modulated.

[0124] For CPM modulation type, bit stream to symbol The mapping is as follows:

[0125] Formula (10)

[0126] in, is the modulation index; is a bit sequence; is the phase impulse response; n is a discrete time index, representing the nth symbol position in the symbol sequence. Each phase changes continuously, and the modulation information is hidden in the phase change rate.

[0127] (2) Generate the complex baseband signal corresponding to the radar signal.

[0128] Generate a time-continuous complex baseband signal based on the linear frequency modulation signal (LFM), as follows:

[0129] Formula (11)

[0130] in, is the initial center frequency (e.g. 10 MHz), is the signal frequency modulation slope (e.g. 100MHz / s), is the initial phase (for example, 0).

[0131] Generate a time-continuous complex baseband signal based on a two-phase coded signal (Barker, Barker Code), as follows:

[0132] Formula (12)

[0133] in, is the Barker code phase, which can be 0 or .

[0134] Based on the pulse repetition interval (PRI) jitter signal, a time-continuous complex baseband signal is generated as follows:

[0135] The time interval between the arrival of two consecutive pulses from the same radar emitter is called the pulse repetition interval, or PRI. For a PRI jitter signal, the PRI value jumps randomly or follows a distribution around a central value, as follows:

[0136] Formula (13)

[0137] in, is the number of arrivals, Indicates the PRI jump center value, is the reference pulse interval (e.g. 1ms), The first PRI hop Secondary value.

[0138] PRI staggered signal: There are 2 to 7 fixed PRI values ​​repeated in a certain order. The mathematical model can be expressed as

[0139] Formula (14)

[0140] in, yes One of the fixed PRI values, mod represents the modulo operation.

[0141] Generate a time-continuous complex baseband signal as follows:

[0142] Formula (15)

[0143] The frequency agile signal generates a time-continuous complex baseband signal as follows:

[0144] Formula (16)

[0145] in, is the initial carrier frequency (e.g. 1GHz), is the agile frequency offset (e.g. ).

[0146] (3) Generate pure noise data

[0147] Based on additive white Gaussian noise, generalized white Gaussian noise and colored noise models, pure noise data is generated. The length of pure noise data is time sampling points.

[0148] The function of step S11 is to generate corresponding complex baseband signals from the communication signal and the radar signal, and to generate pure noise data for generating joint sample data to train the arrival time estimation model.

[0149] Step S12: construct signal sample data containing different arrival time information based on the complex baseband signal to obtain first sample signal data.

[0150] Constructing signal samples containing different arrival time information based on the complex baseband signal to obtain a first sample signal includes:

[0151] The complex baseband signal is transmitted according to a predetermined length Perform signal truncation to obtain multiple lengths of For each sub-complex baseband signal, a sub-complex baseband signal of length An all-zero sequence of sampling points; where ;

[0152] The cut length is The all-zero subsequence of the sampling points is spliced ​​to the starting point of the complex baseband signal; the remaining After the all-zero subsequence of sampling points is spliced ​​to the end point of the complex baseband signal, the length of the complex baseband signal is obtained. The signal samples of sampling points; among them, ;

[0153] The sample label corresponding to the complex baseband signal is the index value of the signal starting sampling point ; The sample label corresponding to the pure noise data is .

[0154] The first sample signal is established as , By a predetermined length of time continuum The complex baseband signal and all-zero sequences, For each radio signal, the corresponding time-continuous complex baseband signal is , generating a length of An all-zero sequence; illustratively, The length of is 100 sampling time points.

[0155] For complex baseband signals (including communication signals and radar signals, pure noise data does not require this processing step), the interception length is The all-zero subsequence is spliced ​​to the starting point of the radio signal, and the rest After splicing the all-zero subsequences of the sampling time points to the end point of the radio signal, the signal sample is obtained. , the length is ;

[0156] The value is from 1 to Select Sample Label Index of the signal starting sampling point ; For length of , the sampling point sequence is .

[0157] Spliced ​​before the starting point of the radio signal All-zero subsequences, indexed by 0 ; Therefore, the starting sampling point of the radio signal is .

[0158] For pure noise data, directly take Noise data with a length of sampling points; there is no signal data in pure noise data; there is no corresponding starting sampling point; to distinguish, the starting sampling point, that is, the sample label is set to .

[0159] In real-world wireless communication environments, radio signals may arrive at the receiver at different times. By appending all-zero sequences of varying lengths before and after the signal, we can simulate signals arriving at the receiver at different times. This helps the time-of-arrival estimation model learn the signal characteristics at different arrival times, thereby improving the accuracy of the arrival time estimation model.

[0160] The function of step S12 is to construct signal sample data containing different arrival time information based on the complex baseband signal, and to simulate the situation in actual applications where pure noise data may be collected, and generate sample data corresponding to the pure noise data; this is used to train the arrival time estimation model to improve its learning ability and estimation accuracy of the signal characteristics of different arrival times.

[0161] Step S13: perform multipath channel simulation and / or noise injection on the first sample signal in sequence to obtain a signal set corresponding to the first sample signal.

[0162] The first sample signal is subjected to multipath channel simulation and / or injection in three ways:

[0163] (1) Perform multipath channel simulation on the first sample signal to obtain the corresponding signal sample set ;

[0164] (2) Noise injection is performed on the first sample signal to obtain the corresponding signal sample set ;

[0165] (3) Perform multipath channel simulation and noise injection on the first sample signal in turn to obtain the corresponding signal sample set .

[0166] Will With the original complex baseband signal , and merged with pure noise data to obtain sample data .

[0167] Multipath channel simulation includes Ricean channel and Rayleigh channel simulation;

[0168] Noise injection includes additive white Gaussian noise, generalized white Gaussian noise and colored noise injection;

[0169] The purpose of multipath channel simulation and noise injection is to simulate the multipath effect and noise injection in a real electromagnetic environment, enhance the diversity and complexity of sample data, and thus improve the robustness and estimation accuracy of the arrival time estimation model in complex scenarios.

[0170] Use multipath channel and noise models to transform signal samples to obtain signal samples with different signal-to-noise ratios that are closer to the real electromagnetic spectrum environment. .

[0171] The multipath channel model includes at least a Rice channel and a Rayleigh channel; the noise model includes at least additive white Gaussian noise, generalized white Gaussian noise, and colored noise. In practical applications, additions and deletions can be made based on specific needs.

[0172] a) Ricean channel transform

[0173] The Ricean channel is a channel model used in mobile and wireless communication systems. It is typically used to describe a situation where a signal has one Line of Sight (LOS) and multiple Non-Line of Sight (NLOS) paths during transmission. The signal amplitude distribution in this channel model is a Ricean distribution.

[0174] The Rice distribution is a probability distribution used to describe the amplitude distribution of the received signal in the case of a dominant LOS and multiple NLOS, and is applicable to the signal propagation characteristics in a scattering environment.

[0175] The Rice channel transform is as follows:

[0176] Formula (17)

[0177] in, is the signal after Rice channel transformation, is Gaussian white noise (usually complex Gaussian noise), is the Rician Factor; is the complex baseband signal.

[0178] Rice factor The Rice factor is the ratio of the signal power of the direct line of sight path to the signal power of the scattered path.

[0179] Formula (18)

[0180] in, is the amplitude of the direct-view path, is the power of the multipath scattering path (i.e., non-line-of-sight path).

[0181] The larger the value of the Rice factor, the stronger the direct-line path signal is and the more concentrated the signal amplitude distribution is; conversely, the smaller the value of the Rice factor, the more dispersed the signal amplitude distribution is. As it approaches zero, the Rice distribution approaches the Rayleigh distribution.

[0182] b) Rayleigh channel transform

[0183] The Rayleigh channel model is used to describe situations where there is no direct line of sight during wireless signal propagation, that is, the signal is completely composed of multipath reflection or scattering paths. The amplitude distribution of the Rayleigh fading channel is a Rayleigh distribution. The channel model formula is as follows:

[0184] Formula (19)

[0185] in, is the signal after Rayleigh channel transformation; is the Rayleigh fading coefficient, which is usually a complex Gaussian random variable. Its amplitude conforms to the Rayleigh distribution and describes the time-varying characteristics of the channel, including the changes in amplitude and phase.

[0186] In formula (19) Amplitude is a Rayleigh distributed random variable, and its probability density function is given by formula (20).

[0187] The probability density function of the Rayleigh distribution is as follows:

[0188] Formula (20)

[0189] in, is the amplitude of the received signal, which is a non-negative scalar random variable; is the standard deviation of Rayleigh fading, reflecting the power of multipath signals.

[0190] The transformation of the noise model mainly involves superimposing different types of noise on the complex baseband signal.

[0191] c) Additive White Gaussian Noise Injection

[0192] Additive white Gaussian noise is superimposed on the complex baseband signal as follows:

[0193] Formula (21)

[0194] in, is the signal after additive white Gaussian noise is injected superimposed on the complex baseband signal; w It depends on the noise type. If it is additive white Gaussian noise, it usually has zero mean and variance of Gaussian noise; In order to control the coefficient of the noise power, the noise power can be controlled to complete the transformation of different signal-to-noise ratios.

[0195] d) Generalized Gaussian white noise transform

[0196] For generalized white Gaussian noise, its probability density function (PDF) can be described by an exponential family. It is widely used to simulate noise in real signals, especially when the noise exhibits a heavy tail.

[0197] The probability density function of generalized Gaussian white noise is as follows:

[0198] Formula (22)

[0199] in, is a shape parameter that determines the tail characteristics of the noise distribution; It is a scale parameter that controls the width of the noise distribution and affects the amplitude of the noise. is the gamma function, which is used to normalize the probability density function to ensure that its integral is 1.

[0200] when When is smaller, the distribution has heavier microsteps, indicating that there are more extreme values ​​in the noise; when When , the generalized Gaussian noise degenerates into additive white Gaussian noise.

[0201] The generalized white Gaussian noise injection is superimposed on the complex baseband signal to obtain the corresponding signal sample; the generalized white Gaussian noise injection obeys the generalized white Gaussian noise probability density in formula (22).

[0202] e) Color noise transformation

[0203] Colored noise is a type of noise different from white noise. Its power spectral density is Varies with frequency. Colored noise can usually be generated by filtering additive white Gaussian noise.

[0204] The power spectral density of the colored noise transform is as follows:

[0205] Formula (23)

[0206] in, is the cutoff frequency (or characteristic frequency) of the colored noise; is the noise baseline power spectral density. The power spectral density of colored noise decays as the frequency increases, and it usually has the characteristic of high gain at low frequencies.

[0207] Colored noise injection is superimposed on the complex baseband signal to obtain the corresponding signal sample; the colored noise injection obeys the colored noise injection probability density in formula (23).

[0208] Different channel models (such as Rayleigh channel, Rice channel) and noise models (such as additive white Gaussian noise, generalized Gaussian noise, colored noise) generate corresponding signal sample sets , which can be used to simulate and analyze the performance of communication systems closer to the actual environment.

[0209] The purpose of step S13 is to simulate the multipath channel simulation and noise injection of the first sample signal to simulate the multipath effect simulation and noise injection in a real electromagnetic environment, thereby enhancing the diversity and complexity of the sample data, thereby improving the robustness and estimation accuracy of the arrival time estimation model in complex scenarios.

[0210] Step S14: construct a signal sample set based on the second sample signal.

[0211] (1) For each second sample signal Perform normalization.

[0212] For example, the normalization method may be a z-score method, which calculates the signal x new The mean u and standard deviation σ of , and normalized calculation are performed using the following formula:

[0213] Formula (24)

[0214] (2) Extract the real and imaginary parts of the normalized signal and stack them into an IQ matrix.

[0215] Formula (25)

[0216] The IQ matrix dimension is , P is the total number of signal samples.

[0217] (3) The data were randomly divided into training set and validation set in a ratio of 8:2.

[0218] The function of step S1 is to generate a signal sample set used for arrival time model training.

[0219] Step S2 includes steps S21-S22.

[0220] Step S21: Construct an arrival time estimation model.

[0221] like Figure 2 As shown in the figure, the arrival time estimation model includes the input layer, adaptive correction module, feature extraction module, feature focusing module and arrival time prediction module in sequence, among which the arrival time prediction module is the model output layer, which contains There are N neurons, corresponding to the probability results of the estimated value of the signal starting sampling point index, and 1 probability result of pure noise data.

[0222] The arrival time estimation model includes an input layer, an adaptive correction module, a feature extraction module, a feature focusing module and an arrival time prediction module in sequence;

[0223] The input layer is configured to receive the IQ matrix;

[0224] The adaptive correction module is used to eliminate the offset of the IQ matrix caused by multipath and noise, and output the corrected time domain IQ matrix;

[0225] The feature extraction module is used to capture fusion features of different scales of the corrected signal;

[0226] The feature focusing module is used to select signal features in the time channel based on the self-attention mechanism and output the focused signal features;

[0227] The arrival time prediction module includes neurons for output The probability of the estimated value of the starting sampling point index of the signal and the probability of 1 pure noise data.

[0228] (1) Input layer: receives the IQ matrix of signal samples and inputs it into the adaptive correction module.

[0229] (2) Adaptive correction module, as follows:

[0230] The adaptive correction module includes a Fourier transform layer, a fully connected frequency domain weight learning layer and an inverse Fourier transform reconstruction layer in sequence;

[0231] The Fourier transform layer is used to transform the IQ matrix from the time domain to the frequency domain to obtain a frequency domain signal;

[0232] The fully connected frequency domain weight learning layer is used to learn the weight of the frequency domain signal, suppress the phase offset component caused by multipath, and attenuate the high-frequency or low-frequency region dominated by noise, so as to obtain a corrected frequency domain signal;

[0233] The inverse Fourier transform reconstruction layer is used to perform inverse fast Fourier transform on the corrected frequency domain signal to restore it to a time domain signal and output an adaptively corrected time domain IQ matrix.

[0234] Fourier transform layer, for each signal sample data The IQ matrix performs a Fast Fourier Transform (FFT) to convert the time domain signal into a frequency domain signal representation.

[0235] Input time domain IQ matrix ,in, Represents a complex field.

[0236] Formula (26)

[0237] in, is the frequency domain signal representation of the sample data.

[0238] The fully connected frequency domain weight learning layer is a fully connected layer with input , learning the frequency domain weight matrix , dynamically adjust the gain of each frequency component as follows:

[0239] Formula (27)

[0240] in, is the corrected frequency domain signal. Fully connected frequency domain weight learning is used to suppress the phase offset component caused by multipath.

[0241] The inverse Fourier transform reconstruction layer is An inverse fast Fourier transform (IFFT) is performed to restore the time domain signal. The corrected time domain IQ matrix significantly reduces multipath simulation and noise injection interference.

[0242] The adaptive correction module corrects the signal by learning a set of weights in the frequency domain. It mainly performs Fourier transform on the signal, passes it through a fully connected frequency domain weight learning layer, and then reconstructs the signal through inverse Fourier transform. Its main function is to eliminate the offset caused by multipath, noise, etc.

[0243] (3) Feature extraction module, as follows:

[0244] Exemplarily, it includes three parallel convolution layers with different convolution kernels.

[0245] The feature extraction module includes parallel first, second and third convolutional layers with different receptive fields, and a Concat splicing layer;

[0246] The first convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract local features of the time-domain IQ matrix after the adaptive correction;

[0247] The second convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract the medium-scale features of the time-domain IQ matrix after the adaptive correction;

[0248] The third convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract the global features of the time-domain IQ matrix after the adaptive correction; wherein, ;

[0249] The Concat splicing layer splices the local features, medium-scale features and global features output by the first, second and third convolutional layers along the channel dimension to obtain fused features.

[0250] For example, in the first convolutional layer, the convolution kernel size is , the number of channels is 64, the padding is 0, and the stride is 1; in the second convolution layer, the convolution kernel size is , the number of channels is 64, the padding is 1, and the stride is 1; the third convolution layer, the convolution kernel size is , the number of channels is 64, the padding is 2, and the stride is 1.

[0251] The function of the feature extraction module is to effectively capture signal features of different scales and obtain corresponding fusion features.

[0252] (4) Feature aggregation module, specifically.

[0253] The feature focusing module includes a self-attention feature focusing layer, and a splicing residual connection between the output of the Concat splicing layer and the output of the self-attention feature focusing layer;

[0254] The self-attention feature focusing layer includes a self-attention layer, a focused residual connection and a layer normalization in sequence;

[0255] The self-attention layer is used to map the fused features into Q, K, and V matrices through linear transformation; calculate attention weights based on the Q, K, and V matrices, and output weighted fused features;

[0256] The focused residual connection and layer normalization are used to connect and normalize the fused features and the weighted fused features; the output focused features are transferred to the arrival time prediction module.

[0257] The feature aggregation module receives the fused features and passes them to the self-attention layer.

[0258] The self-attention layer maps the input fusion features into query matrix Q, key matrix K, and value matrix V through linear transformation, as follows:

[0259] Formula (28)

[0260] in, is the fusion feature matrix; 、 and Are the conversion matrices respectively. Calculate the attention weights to obtain the weighted fusion features .

[0261] Perform focused residual connection and layer normalization as follows:

[0262] Formula (29)

[0263] in, It is the focused feature output by the feature focusing module.

[0264] The feature focusing module selects features in the time channel based on the self-attention mechanism, allowing the neural network to focus on important features.

[0265] (5) Arrival time prediction module, specifically

[0266] Feature Focusing Module Output Features Input to the arrival time prediction module is the output layer of the network, including neurons, which is a standard fully connected layer.

[0267] Outputs the probability of each possible signal starting sampling point index estimate. This module includes neurons, each neuron corresponds to a possible signal starting position; The estimated sampling point index represents pure noise data. The module predicts the probability distribution of the signal starting point, providing a basis for arrival time estimation.

[0268] Specifically, the focused features output by the feature focusing module are received, linearly transformed through the weights and bias parameters of the fully connected layer, and then the output is converted into a probability distribution form using the softmax function.

[0269] Step S22: Train the arrival time estimation model using the signal sample set, and save the model parameters with the minimum loss function after the maximum number of iterations, to obtain a trained arrival time estimation model.

[0270] During the training process, the model parameters are updated by minimizing the cross-entropy loss function, and the parameter optimization and update are performed using the Adam optimizer. A dynamic learning rate adjustment strategy is set to dynamically adjust the learning rate during the training process. The learning rate constrains the scale of each parameter update. The dynamic adjustment strategy can prevent large deviations in the model parameter update and speed up the training speed. At the same time, the training loss is monitored on the validation set. If the loss continues to increase during the training process, the training is stopped. Otherwise, the training is stopped when the preset maximum number of iterations is reached.

[0271] The loss function is as follows:

[0272] Formula (30)

[0273] in, is the second sample signal; Normalizing the second sample signal; 、 respectively extracting the real part and the imaginary part of the normalized second sample signal; is the parameterized function of the arrival time estimation model, which is used to obtain the probability distribution of the model prediction; is a learnable parameter, is the true label corresponding to the input sample, The number of categories output by the model.

[0274] Exemplarily, the maximum number of iterations is set to 500.

[0275] During the training process, the parameters of the arrival time estimation model of each training cycle are saved. When the maximum number of iterations is reached, the arrival time estimation model with the smallest loss on the validation set is selected as the final trained arrival time estimation model.

[0276] The function of step S2 is to construct and train the arrival time estimation model. The signal sample set is learned through a neural network model including an input layer, an adaptive correction module, a feature extraction module, a feature focusing module and an arrival time prediction module. When the maximum number of iterations is reached, the model parameters with the minimum loss function are saved to obtain a trained arrival time estimation model.

[0277] Step S3, specifically.

[0278] The receiver collects the actual radio signal in real time; records the time when the receiver collects the signal ; The actual collected radio signal is intercepted based on a predetermined length sliding window to obtain multiple signal segments; signal segment sequence number Start from 1;

[0279] The sliding window step size can be set according to application requirements. For example, the sliding window step size is , to achieve sliding by sampling point.

[0280] The length of each signal segment is sampling points.

[0281] If the radio signal length is insufficient sampling points, there is only one signal segment, and the deficiency is supplemented by a zero sequence;

[0282] If the radio signal is greater than The length of the sampling points is The sliding window is used to intercept the signal fragments.

[0283] The receiver performs down-conversion processing to obtain a complex baseband signal corresponding to each signal segment; the complex baseband signal corresponding to each signal segment is normalized to extract the IQ matrix, thereby obtaining multiple corresponding IQ matrices;

[0284] Multiple IQ matrices are sequentially input into the trained arrival time estimation model. If the estimated values ​​of the signal starting sampling point index obtained by all signal segments are , then the actual radio signal collected in real time is pure noise data;

[0285] Otherwise, the signal arrival time is calculated based on the estimated value of the signal starting sampling point index and the sequence number of the signal segment. ,as follows:

[0286] Formula (31)

[0287] in, is the moment when the receiver collects the signal. is the sampling rate of the receiver; The estimated value of the signal starting sampling point index; For the first time The sequence number of the signal segment with the estimated value of the signal starting sampling point index within .

[0288] The function of step S3 is to collect radio signals in real time. If the radio signals collected in real time include communication signals or radar signals, the IQ matrix is ​​extracted after down-conversion and normalization, and input into the trained arrival time estimation model to obtain the index estimate of the signal starting sampling point, and then calculate the signal arrival time. Otherwise, the radio signals collected in real time are pure noise data.

[0289] In summary, the deep learning signal detection method adapted to changes in the number of receiving antennas according to an embodiment of the present invention has the following beneficial effects:

[0290] 1. This invention trains a time-of-arrival estimation model based on the original radio signal and directly learns the signal starting point features without manual feature extraction. This lowers the threshold for use and is applicable to unknown signal types and non-cooperative scenarios. It also solves the problems of traditional methods that rely on manually designed features, require complex signal processing experience, and have poor generalization capabilities.

[0291] 2. This invention introduces Ricean and Rayleigh channel conversions and multiple noise injections into the signal sample sets used for training, generating signal samples that are closer to the real environment. The sample data is more consistent with the actual electromagnetic environment, effectively improving the adaptability and accuracy of signal arrival time estimation in complex electromagnetic spectra. It also solves the problem that the multipath effect and noise injection in existing methods cause the accuracy of signal arrival time estimation to drop significantly.

[0292] 3. The adaptive correction module in the present invention suppresses multipath and noise injection, dynamically corrects the signal through frequency domain weight learning, and suppresses multipath phase offset and noise frequency band. This solves the problem of signal distortion caused by multipath in existing methods, making it difficult to distinguish between the main path and the delayed path.

[0293] 4. The feature extraction module and feature focusing module in this invention are based on multi-scale feature fusion and self-attention mechanism. Multi-scale convolution extracts local, medium and global features in parallel to cover signal mutations. Self-attention focuses dynamically enhances the time channel features near the starting point and suppresses irrelevant areas. This solves the problem of single-scale features missing key information in existing methods.

[0294] 5. The sample data in the present invention covers communication signals and radar signals. Based on multiple modulation types, the time arrival estimation model adaptively learns the characteristics of heterogeneous signals through the feature extraction module, which solves the problem of poor adaptability to specific modulation types in traditional methods.

[0295] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0296] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A radio signal arrival time estimation method based on deep learning, characterized in that: include: Preprocessing the complex baseband signal and pure noise data generated by simulation to obtain sample data in the form of an IQ matrix, and forming a signal sample set with the corresponding sample label; wherein the sample label corresponding to the complex baseband signal is the index value of the signal starting sampling point; The arrival time estimation model is trained using the signal sample set to obtain a trained arrival time estimation model by saving the model parameters with the minimum loss function after a maximum number of iterations. The radio signal is collected in real time, down-converted to obtain a corresponding complex baseband signal, cut into multiple signal segments based on a sliding window of a predetermined length, and normalized. The real and imaginary parts are extracted and stacked into corresponding multiple IQ matrices, which are sequentially input into the trained arrival time estimation model until an index estimate of the signal starting sampling point is found; and the signal arrival time is calculated based on the index estimate of the signal starting sampling point and the sequence number of the signal segment; The arrival time estimation model includes an input layer, an adaptive correction module, a feature extraction module, a feature focusing module and an arrival time prediction module in sequence; The input layer is configured to receive the IQ matrix; The adaptive correction module is used to eliminate the offset of the IQ matrix caused by multipath and noise, and output the corrected time domain IQ matrix; The feature extraction module is used to capture the fusion features of different scales of the corrected time-domain IQ matrix; The feature focusing module is used to select signal features in the time channel based on the self-attention mechanism and output the focused signal features; The arrival time prediction module includes neurons for output The probability of the estimated value of the starting sampling point index of the signal and the probability of 1 pure noise data; The adaptive correction module includes a Fourier transform layer, a fully connected frequency domain weight learning layer and an inverse Fourier transform reconstruction layer in sequence; The loss function is as follows: ； in, is the second sample signal; is the normalized second sample signal; 、 are respectively the real part and imaginary part of the normalized second sample signal; is the parameterized function of the arrival time estimation model, which is used to obtain the probability distribution of the model prediction; is a learnable parameter, is the true label corresponding to the input sample, The number of categories output by the model; The signal arrival time is calculated based on the estimated index value of the signal starting sampling point and the sequence number of the signal segment. ,as follows: ； in, is the moment when the receiver collects the signal. is the sampling rate of the receiver; The estimated value of the signal starting sampling point index; For the first time The sequence number of the signal segment with the estimated value of the signal starting sampling point index within .

2. The method for estimating the arrival time of radio signals based on deep learning according to claim 1, wherein: The complex baseband signal includes a communication complex baseband signal and a radar complex baseband signal; The preprocessing of the complex baseband signal and pure noise data generated by simulation to obtain sample data includes: Cutting the complex baseband signal into a plurality of signal segments based on a predetermined length, constructing signal samples containing different arrival time information based on the signal segments, and obtaining a first sample signal; performing multipath channel simulation and / or noise injection on the first sample signals respectively to obtain a signal set corresponding to the first sample signals; The pure noise data has a length of sampling points; The signal set corresponding to the first sample signal has a length of The pure noise data of the sampling points are merged into the second sample signal; The second sample signal is normalized, and the real and imaginary parts are extracted and stacked into an IQ matrix as sample data.

3. The method for estimating the arrival time of radio signals based on deep learning according to claim 2, wherein: Constructing signal samples containing different arrival time information based on the complex baseband signal to obtain a first sample signal includes: The complex baseband signal is transmitted according to a predetermined length Perform signal truncation to obtain multiple lengths of For each sub-complex baseband signal, a sub-complex baseband signal of length An all-zero sequence of sampling points; where ; The intercept length is The all-zero subsequence of the sampling points is spliced ​​to the starting point of the complex baseband signal; the remaining After the all-zero subsequence of sampling points is spliced ​​to the end point of the complex baseband signal, the length of the complex baseband signal is obtained. The first sample signal of the sampling point; wherein, is the number of sampling points of the complex baseband signal, ; The sample label corresponding to the complex baseband signal is the index value of the signal starting sampling point ; The sample label corresponding to the pure noise data is .

4. The method for estimating the arrival time of radio signals based on deep learning according to claim 1, wherein: The Fourier transform layer is used to transform the IQ matrix from the time domain to the frequency domain to obtain a frequency domain signal; The fully connected frequency domain weight learning layer is used to learn the weight of the frequency domain signal, suppress the phase offset component caused by multipath, and attenuate the high-frequency or low-frequency region dominated by noise, so as to obtain a corrected frequency domain signal; The inverse Fourier transform reconstruction layer is used to perform inverse fast Fourier transform on the corrected frequency domain signal to restore it to a time domain signal and output an adaptively corrected time domain IQ matrix.

5. The method for estimating the arrival time of radio signals based on deep learning according to claim 4, wherein: The feature extraction module includes parallel first, second and third convolutional layers with different receptive fields, and a Concat splicing layer; The first convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract local features of the time-domain IQ matrix after the adaptive correction; The second convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract the medium-scale features of the time-domain IQ matrix after the adaptive correction; The third convolutional layer has a convolution kernel size of , the number of channels is , filled with , the stride is , used to extract the global features of the time-domain IQ matrix after the adaptive correction; wherein, ; The Concat splicing layer splices the local features, medium-scale features and global features output by the first, second and third convolutional layers along the channel dimension to obtain fused features.

6. The method for estimating the arrival time of radio signals based on deep learning according to claim 5, characterized in that: The feature focusing module includes a self-attention feature focusing layer; wherein the output of the Concat splicing layer is spliced ​​with the output of the self-attention feature focusing layer through a residual connection; The self-attention feature focusing layer includes a self-attention layer, a focused residual connection and a layer normalization in sequence; The self-attention layer is used to map the fused features into Q, K, and V matrices through linear transformation; calculate attention weights based on the Q, K, and V matrices, and output weighted fused features; The focused residual connection and layer normalization are used to connect and normalize the fused features and the weighted fused features; the output focused features are transferred to the arrival time prediction module.

7. The method for estimating the arrival time of radio signals based on deep learning according to any one of claims 2 to 6, wherein: The multipath channel simulation includes Rice channel and Rayleigh channel simulation; The noise injection includes additive white Gaussian noise, generalized white Gaussian noise and colored noise injection.

Images