Multi-path interference detection and suppression method based on multi-modal deep learning

Abstract

The invention relates to a multipath interference detection and suppression method based on multi-modal deep learning, and belongs to the technical field of wireless communication and artificial intelligence crossing. According to the method, interference detection, intensity grading and path separation are realized through multi-task deep network combination, and signal reconstruction is optimized by using a conditional generative adversarial network in combination with time domain sparse constraint. The method is suitable for a radio communication scene, the communication reliability in a complex time-varying environment can be remarkably improved, the bit error rate is reduced, and the spectrum efficiency is optimized. Simulation experiments show that the method can reduce the bit error rate by one order of magnitude in scenes of ground radio communication, satellite communication and the like, and is obviously superior to a traditional scheme.

Description

Technical Field

[0001] The present invention belongs to the technical field of intersection of wireless communications and artificial intelligence, and specifically relates to a multipath interference detection and suppression method based on multimodal deep learning. Background Art

[0002] 1. Challenges of multipath interference

[0003] Multipath interference occurs when a radio signal travels multiple paths from the transmitting antenna to the receiving antenna. Due to the varying lengths and propagation conditions of these paths, the signals arrive at the receiving end at different times, amplitudes, and phases. The resulting interference, caused by the superposition of their phases, can distort the original signal or produce errors. The main causes of multipath interference include three main factors: the diversity of signal propagation paths, channel characteristics, and the relative position of the transmitter and receiver. Signal propagation paths refer to the multiple propagation paths created by reflection, refraction, and scattering from buildings, terrain, vegetation, water, or moving objects during signal propagation. Channel characteristics refer to the frequency-selective fading and time-varying nature of wireless channels. Signals attenuate differently at different frequencies, and channel characteristics vary over time. Multipath interference between the transmitter and receiver is related to the distance between them. When the distance between the transmitter and receiver is close, multipath interference is relatively minimal, while at the opposite distance, multipath interference is more severe.

[0004] The impact of multipath on signal transmission is mainly manifested in the following four aspects:

[0005] (1) Delay extension

[0006] The time difference between signals from different paths arriving at the receiving end causes inter-symbol interference (ISI), which makes it impossible for the receiving end to correctly decode and identify the signal.

[0007] (2) Signal attenuation

[0008] The multipath effect causes the signal energy to be dispersed over multiple paths, resulting in signal attenuation and reduced signal strength, making it impossible for the receiving end to receive the signal correctly.

[0009] (3) Signal distortion

[0010] The signal amplitude fluctuates in different frequency bands due to multipath superposition. In mobile scenarios, Doppler frequency shift aggravates phase noise, causing signal distortion at the receiving end.

[0011] (4) Increased error rate

[0012] Due to the interference and overlap of multiple signal paths, the bit error rate of the received signal at the receiving end increases.

[0013] Traditional methods for mitigating signal multipath effects, such as minimum mean square error (MMSE) equalization, orthogonal frequency division multiplexing (OFDM) cyclic prefix, and space-time coding (STC), have the following limitations:

[0014] (1) Poor adaptability to high dynamic environments

[0015] The equalizer converges slowly in time-varying channels and is not suitable for highly dynamic environments.

[0016] (2) Weak resolution of dense multipath

[0017] Traditional filtering algorithms have difficulty separating overlapping path signals and identifying and distinguishing each signal.

[0018] (3) High computational complexity

[0019] The calculation is complex, requires high computing power, and has large delays, making it difficult to meet the low-latency requirements of real-time systems.

[0020] 2. Application of Deep Learning in Signal Processing

[0021] Deep learning is a machine learning method based on neural networks. It realizes automatic learning and representation of complex data through multi-layer nonlinear transformations, providing a new means for multipath interference suppression.

[0022] (1) Signal modulation recognition

[0023] By learning the characteristics of signals under different modulation modes, it is possible to automatically identify the modulation mode of communication signals. For example, by using convolutional neural networks (CNNs) to extract and classify the time-frequency characteristics of communication signals, it can accurately identify modulation modes such as AM, FM, and PSK, which has important applications in spectrum monitoring and communication reconnaissance.

[0024] (2) Channel estimation and equalization

[0025] Used for channel estimation and equalization to improve the quality and reliability of signal transmission. For example, deep learning-based channel estimation methods can leverage the statistical characteristics of received signals to quickly and accurately estimate channel parameters, achieving signal equalization and recovery.

[0026] (3) Signal detection and decoding

[0027] Used for signal detection and decoding to improve the performance of communication systems. For example, deep learning-based signal detection algorithms can accurately detect the presence and parameters of signals in complex channel environments and achieve correct signal decoding. Summary of the Invention

[0028] (1) Technical issues to be resolved

[0029] The technical problem to be solved by the present invention is how to provide a multipath interference detection and suppression method based on multimodal deep learning to achieve real-time detection and classification of multipath interference in a highly dynamic environment, suppress interference and restore high signal-to-noise ratio signals through adaptive signal reconstruction technology.

[0030] (2) Technical solution

[0031] In order to solve the above technical problems, the present invention proposes a multipath interference detection and suppression method based on multimodal deep learning, which includes the following steps: a multimode signal acquisition step, a joint time-frequency preprocessing step, a multipath interference dynamic detection step, and an adaptive signal reconstruction step;

[0032] The multi-mode signal acquisition step is used to enable the receiving end to obtain radio signals through the antenna array, including: radio frequency signals, Doppler frequency shift data and environmental noise signals;

[0033] A joint time-frequency preprocessing step is used to denoise and align the signal using wavelet packet transform (WPT) and adaptive Kalman filtering to achieve signal preprocessing;

[0034] The dynamic detection step of multipath interference is used to detect the presence of interference, grade the intensity, and separate the paths of the signal after joint time-frequency preprocessing based on the multi-task deep neural network MT-DNN.

[0035] The adaptive signal reconstruction step is used to jointly optimize signal reconstruction through conditional generative adversarial network (cGAN) and time-domain sparse coding.

[0036] (3) Beneficial effects

[0037] This paper proposes a multipath interference detection and mitigation method based on multimodal deep learning. This method uses a multi-task deep network to jointly implement interference detection, intensity classification, and path separation. Furthermore, it utilizes a conditional generative adversarial network combined with time-domain sparsity constraints to optimize signal reconstruction. Simulation experiments demonstrate that this method can reduce the bit error rate by an order of magnitude in scenarios such as terrestrial radio communications and satellite communications, significantly outperforming traditional approaches. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The overall architecture and signal processing flow of the present invention;

[0039] Figure 2 This is the flow chart of wavelet packet transform;

[0040] Figure 3 This is the diagram of the wavelet packet filter coefficient decomposition process;

[0041] Figure 4 is the signal time domain diagram;

[0042] Figure 5 is the signal spectrum diagram;

[0043] Figure 6 It is a three-layer wavelet packet decomposition tree diagram;

[0044] Figure 7 It is the wavelet coefficient graph of 8 nodes in the third layer;

[0045] Figure 8 It is the signal frequency distribution diagram of each node in the third layer;

[0046] Figure 9 It is the signal spectrum diagram of each node in the third layer;

[0047] Figure 10 The time domain diagram of each frequency band signal is decomposed and reconstructed by the three-layer wavelet packet of Haar wavelet basis;

[0048] Figure 11 This is a diagram of the multi-task deep network structure;

[0049] Figure 12 This is a time domain diagram of a signal affected by multipath for a simulation example;

[0050] Figure 13 This is the time-frequency analysis diagram of the signal affected by multipath through wavelet packet;

[0051] Figure 14 This is the short-time Fourier transform (STFT) analysis diagram of the signal affected by multipath;

[0052] Figure 15 This is a comparison chart of detection performance of different analysis methods for the same signal affected by multipath;

[0053] Figure 16 Schematic diagram of the cGAN signal reconstruction process;

[0054] Figure 17 Figure 2 is a BER performance curve at different mobile speeds. DETAILED DESCRIPTION

[0055] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings and examples.

[0056] This invention, at the intersection of wireless communications and artificial intelligence, relates to a method for dynamic detection and adaptive suppression of multipath interference that combines multimodal deep learning, adaptive signal reconstruction, and time-frequency analysis. This method, applicable to radio communication scenarios, can significantly improve communication reliability, reduce bit error rates, and optimize spectrum efficiency in complex, time-varying environments.

[0057] 1. Core steps and processes of the present invention

[0058] The core steps and processes of this method include: multi-mode signal acquisition step, joint time-frequency preprocessing step, multipath interference dynamic detection step and adaptive signal reconstruction step, such as Figure 1 shown.

[0059] in,

[0060] The multi-mode signal acquisition step is used to enable the receiving end to obtain radio signals such as radio frequency signals, Doppler frequency shift data and environmental noise signals through the antenna array;

[0061] The joint time-frequency preprocessing step is mainly used to denoise and align the signal using wavelet packet transform (WPT) and adaptive Kalman filtering to achieve signal preprocessing;

[0062] The dynamic detection step of multipath interference is used to detect the presence of interference, grade the intensity, and separate the paths of the signal after joint time-frequency preprocessing based on the multi-task deep neural network MT-DNN.

[0063] Adaptive signal reconstruction step, used to jointly optimize signal reconstruction through conditional generative adversarial network (cGAN) and time-domain sparse coding;

[0064] 2. Key technologies and models for each step

[0065] 2.1 Multimode Signal Acquisition Model

[0066] The radio signals (including RF signals, Doppler frequency shift data, and environmental noise signals) received by the receiving antenna array at time t from k paths are expressed as x(t):

[0067]

[0068] In formula (1),

[0069] α k (t): time-varying attenuation coefficient of the kth path;

[0070] s k (t): radio signal of the kth path at time t;

[0071] τ k : time-varying path delay;

[0072] Doppler shift of the kth path;

[0073] n(t): additive white Gaussian noise;

[0074] j: imaginary unit, satisfying j 2 =-1.

[0075] 2.2 Joint time-frequency preprocessing

[0076] Multipath propagation can cause signal shifts in time and frequency. The combined time-frequency preprocessing step uses the Wavelet Packet Transform (WPT) to perform multi-level segmentation of the signal's frequency band. The low- and high-frequency components of the decomposed signal are then further decomposed. Based on the characteristics of the signal being analyzed, the corresponding frequency band is adaptively selected to match the signal's spectral terms, thereby improving time-frequency resolution. For signals x(t) containing multipath propagation, the Wavelet Packet Transform allows for better time-frequency localization analysis, providing high-quality signals for multipath interference detection and mitigation.

[0077] (1) Wavelet Packet Transform (WPT) denoising

[0078] After the continuous radio signal x(t) is discretized, it becomes a discrete signal x(n), where n=0, 1, ..., N-1, n represents the number of sampling cycles, and N is the number of signal samples. The wavelet packet transform continuously decomposes c(n) step by step through the low-pass filter h(n) and the high-pass filter g(n) to obtain the wavelet packet coefficients. The principle is as follows Figure 2 shown.

[0079] The filter coefficient decomposition process of wavelet packet transform is as follows: Figure 3 As shown in the figure, S represents the low-frequency signal of the input signal x(n) after being processed by the low-pass filter h(n), and D represents the high-frequency signal of the input signal x(n) after being processed by the high-pass filter g(n). SS represents the low-frequency signal of the S signal after being processed by h(n) in the second-layer wavelet packet decomposition process, and SD represents the high-frequency signal of the signal after being processed by g(n) in the second-layer wavelet packet decomposition process. DS represents the low-frequency signal of the D signal after being processed by h(n) in the second-layer wavelet packet decomposition process, and DD represents the high-frequency signal of the signal after being processed by g(n) in the second-layer wavelet packet decomposition process. Following this rule step by step, the corresponding SSS, SSD, SDS, SDD, DSS, DSD, DDS, DDD, ... are represented.

[0080] Depend on Figure 3 It can be seen that starting from the first layer, the filtering of the discrete signal x(n) is distributed according to the frequency, and its wavelet packet transform can be realized by a set of filter banks. Let h(n) be a low-pass filter and g(n) be a high-pass filter, j represents the number of decomposition layers, k represents the different frequency channels on the jth layer, n represents the discrete time point, which is the discrete time point index variable, and m represents the traversal of all possible discrete time points, which is also the summation index variable. The wavelet packet coefficient d on the jth layer and the kth frequency channel is j,k (n) can be calculated by the following recursive formula:

[0081] Initial condition: d 0,0 (n,n)=x(n)

[0082] The decomposition process is shown in formula (2):

[0083]

[0084] Each wavelet packet coefficient d j,k (n) corresponds to the signal's information within a specific time-frequency region. As the number of decomposition layers j increases, frequency resolution gradually increases while time resolution decreases. Within the same layer, different values ​​of k correspond to different frequency channels, enabling detailed analysis of the signal across different frequency bands. The coefficients of the low-frequency portion primarily reflect the overall trend and slowly varying components of the signal, while the coefficients of the high-frequency portion contain more information about signal details, sudden changes, and noise. Analyzing the high-frequency portion is particularly valuable in communication signals containing multipath.

[0085] After wavelet packet transform, a discrete signal is decomposed into a series of wavelet packet coefficients of varying frequencies and time resolutions. These coefficients form a time-frequency matrix, where rows correspond to different time points or sampling points, and columns correspond to different frequency channels or scales. Each element in the matrix represents the value of a wavelet packet coefficient at a specific time and frequency position. This value reflects the signal strength at that time-frequency point and is often used for signal feature extraction.

[0086] (2) Wavelet packet denoising threshold processing

[0087] There are two thresholding methods for wavelet packet coefficients in equation (2): hard thresholding and soft thresholding. These two thresholding methods have different characteristics. Hard thresholding compares the signal value with a set threshold. The portion below the threshold is set to zero, while the portion above the threshold remains unchanged. This is suitable for situations with high signal-to-noise ratios. Soft thresholding, on the other hand, subtracts the set threshold from the signal value and thresholds the result to achieve signal sparsification. This is suitable for situations where the signal is subject to strong interference or high noise. In the preprocessing process of radio wavelet packet denoising for multipath signals, it is recommended to use soft thresholding in the first layer and hard thresholding in the second layer and above.

[0088] Hard threshold: Let λ be the threshold, wavelet packet coefficient d, and the coefficient dth after hard threshold processing is

[0089]

[0090] Soft threshold: The coefficient dth after soft threshold processing is

[0091]

[0092] Where sgn(d) is the sign function, that is

[0093]

[0094] (3) Wavelet packet coefficient reconstruction

[0095] After the signal is processed by wavelet packet denoising threshold, the wavelet packet coefficients are reconstructed to obtain the denoised signal. The wavelet packet coefficient reconstruction process is the inverse operation of the decomposition process and is achieved through a set of reconstruction filters. and are the coefficients of the reconstructed low-pass filter and high-pass filter corresponding to the decomposition filters h(n) and g(n), respectively. Starting from the bottom layer, the reconstruction is gradually performed upwards. The reconstruction model from the jth layer to the j-1th layer is as shown in formula (6):

[0096]

[0097] The definitions of the variables are the same as those in the wavelet packet transform denoising part, that is, n represents a discrete time point and m is a summation index variable. j,k (n) is the wavelet packet coefficient of different layer j and different frequency channel k.

[0098] On Reconstructing Low-Pass Filter Coefficients High-pass filter coefficients The decomposition low-pass filter coefficient h(n) and the high-pass filter coefficient g(n) are mutually dual, satisfying the biorthogonal relationship and conjugate mirror relationship. The reconstructed filter coefficient can be calculated by decomposing the filter coefficient. The common method is to express the decomposition and reconstruction process with matrices. Let H and G be the matrices composed of the decomposition low-pass filter h(n) and the high-pass filter g(n), respectively. and is reconstructed by the low-pass filter and high-pass filter The matrix formed by

[0099]

[0100] Where I is the identity matrix. By solving equation (7), we can get the reconstruction filter matrix and Then we get the reconstruction filter and

[0101] (4) Adaptive Kalman filtering for signal alignment

[0102] After the signal is reconstructed through wavelet packet denoising, it is necessary to align the signal and apply adaptive Kalman filtering to adjust the process noise covariance matrix Q in real time according to the changes in the signal. k and the measurement noise covariance matrix R k . The new information sequence through Kalman filtering The statistical characteristics of z are adjusted, where z k is the signal measurement value at time k, H kis the observation matrix of the signal system, is the predicted value of the signal at time k, ∈ k is the new information of Kalman filter at time k (that is, the difference between the observed value and the predicted observation value). Use the maximum likelihood estimation method to minimize the negative log-likelihood function of the new information sequence to estimate Q k and R k Achieve signal alignment.

[0103] Through the above joint time-frequency preprocessing method, the noise in the signal can be effectively removed, and the signal can be aligned in time and frequency, providing high-quality signal input for subsequent multipath interference detection and suppression.

[0104] Take an example to illustrate the signal decomposition process of wavelet packet transform. Suppose a signal consists of three sinusoidal signals with amplitude of 1 at 350Hz, 400Hz and 750Hz, and contains Gaussian white noise with mean of 0 and variance of 1. The sampling frequency is 2048Hz and the sampling time is 2 seconds. The Haar wavelet basis is used to perform three-layer wavelet packet denoising on the signal.

[0105] According to the above principles and methods, the results of the signal processing process are as follows, where the time domain distribution of the signal is as follows Figure 4 As shown, the frequency characteristics are as follows Figure 5 As shown, the three-layer wavelet packet decomposition tree is as follows Figure 6 As shown, the coefficients corresponding to each node in the third layer of wavelet packet decomposition are as follows Figure 7 As shown, the frequency distribution of the signal sampling points corresponding to each node in the third layer of wavelet packet decomposition is as follows: Figure 8 As shown, the spectrum characteristics of the signals corresponding to each node in the third layer of wavelet packet decomposition are as follows Figure 9 As shown in the figure, the reconstructed signals corresponding to different frequency bands after wavelet packet denoising are as follows: Figure 10 shown.

[0106] 2.3 Multi-Task Deep Network (MT-DNN) Design

[0107] The multi-task deep network is a multi-task learning network, which is divided into four layers, from top to bottom: input layer, shared feature extraction layer, specific task layer and output layer, such as Figure 11 shown.

[0108] Figure 11 In , the input layer refers to the signal after receiving joint time-frequency preprocessing, that is, the time-frequency matrix of the signal after wavelet packet transform (WPT);

[0109] The shared feature extraction layer refers to the application of convolutional neural networks (CNN) to extract the time-frequency features of the time-frequency matrix. This layer consists of three layers of convolutional structural blocks. Each convolutional structural block contains three core elements: one-dimensional convolution Conv1D for extracting time-frequency features, batch normalization processing layer (BatchNorm) for accelerating training, and ReLu activation function for introducing nonlinearity.

[0110] The specific task layer is a specific task analysis set around multipath signal analysis. The present invention designs three specific task networks, namely a multipath interference signal detection network, an interference signal strength classification network, and a multipath signal path delay distribution network, which respectively realize multipath signal interference detection, multipath signal interference strength classification, and multipath path delay distribution prediction of the signal processed by the shared feature extraction layer;

[0111] The output layer corresponds to the prediction results of the three specific task networks in the specific task layer, namely, the interference signal existence rate output by the multipath signal interference detection network, the interference intensity level of the multipath interference signal intensity classification network, and the delay distribution of the multipath signal path delay distribution network.

[0112] The key steps of multi-task deep network processing time-frequency signals are as follows:

[0113] (1) Shared feature extraction layer

[0114] Input: time-frequency matrix decomposed by wavelet packet decomposition Where X represents the time-frequency matrix of the signal after wavelet packet decomposition, Indicates a matrix whose elements are real numbers, T represents the number of rows of the time-frequency matrix X, corresponding to different time points, and F represents the number of columns of the time-frequency matrix X, corresponding to different frequency channels.

[0115] Network structure: 3 layers of convolution blocks (each layer contains Conv1D+BatchNorm+LeakyReLU),

[0116] Output: Feature map Among them F c It is the feature map output by the convolution block, which is a three-dimensional real number matrix. T' represents the height of the feature map in space, corresponding to the time information, F' represents the width of the feature map in space, corresponding to the frequency information, and C represents the number of channels of the feature map, corresponding to a feature extraction result of the signal.

[0117] (2) Specific task layer

[0118] Input: Feature map F c Features weighted by spatial attention;

[0119] Network structure: Bidirectional long short-term memory network (LSMT) (hidden layer dimension 256)

[0120] Output: Time series features It includes interference signal existence matrix, interference signal strength distribution matrix and multipath signal delay distribution matrix.

[0121] (3) Multi-task output

[0122] (3.1) Interference detection output

[0123] The output function of the multipath interference signal detection network usually uses a binary cross entropy loss function to measure the difference between the actual state of whether the output signal has interference in a multipath interference environment and the signal whether there is interference based on the multi-task deep network model. For a real signal, there is either no interference or interference. Let the label y be the label of whether the signal has interference. When y = 0, it means that the signal has no interference, and when y = 1, it means that the signal has interference. Let the probability of signal interference based on the multi-task deep network model be p exist , then the binary cross entropy loss function is defined as:

[0124]

[0125] This loss function measures the predicted probability p exist The difference between the actual label y, when p exist The closer it is to y, the smaller the loss value.

[0126] In the specific task layer, after time-frequency shared feature extraction and specific task layer processing, a signal value is obtained, set as z. The last layer uses the Sigmoid activation function to map the output to the (0,1) interval, indicating the probability of interference existence, which is output to the output layer. The Sigmoid function is (3.2) Intensity regression output

[0127] The interference signal strength classification network uses the mean square error loss function to measure the difference between the predicted interference strength and the actual interference strength. Let the actual interference strength be y i , the interference intensity predicted by the interference signal strength classification network in the specific task layer is For n signal samples, the mean square error loss function is defined as

[0128]

[0129] This loss function effectively reflects the accuracy of the signal strength grading network model. A smaller loss value indicates that the prediction result is closer to the true value. In practical applications, the fully connected layer of the network is used to predict the interference strength feature. The fully connected layer maps the previously extracted interference strength feature to the interference strength numerical space [0,1]. The stronger the strength, the strength value is approximately 1, and the weaker the strength, the strength value is closer to 0.

[0130] (3.3) Path delay distribution

[0131] The non-negative matrix factorization loss of the multipath signal path delay distribution network is usually defined based on the objective of non-negative matrix factorization (NMF). Suppose the time-frequency matrix decomposed by wavelet packet is is decomposed into two non-negative matrices and That is, X≈WH, where T represents the rows of the time-frequency matrix, i.e., the number of sampled signals, F represents the number of features or dimensions of the sampled signals, and k represents the dimension of the latent variable after decomposition, i.e., the number of potential features, which must be pre-set to k≤min(T,F). Each column of W represents a potential path feature, and each row corresponds to a path; each column of H corresponds to a time point, and each column corresponds to a potential feature. The loss function is defined in the form based on Euclidean distance, i.e.

[0132]

[0133] Where i and j represent the rows and columns of the matrix, corresponding to the path number and time point respectively. (WH) i,j Reflects the estimated delay of the i-th path at the j-th time point. W ≥ 0 and H ≥ 0 must be guaranteed. This loss function measures the difference between the decomposed matrix WH and the time-frequency matrix X. By minimizing this loss function, appropriate W and H are found to generate the path delay distribution.

[0134] The update of W and H uses the following multiplication update rule:

[0135]

[0136] In formula (11), (W T X) j,k represents the “correlation” between the jth time-frequency matrix X signal feature and the kth potential feature, (W T WH) j,k Indicates the "reconstruction error correlation term" between the jth time-frequency matrix X signal feature and the kth potential feature in the current WH matrix. (XH T ) j,k Represents the “correlation” between the jth time-frequency matrix X signal feature and the kth potential feature, (WHH T ) j,k It represents the "reconstruction error correlation term" between the jth time-frequency matrix X signal feature and the kth potential feature in the current WH matrix. In each iteration, H is updated first, then W is updated. That is, W is fixed first to update H, and then the updated H is fixed to update W until the loss function converges to a sufficiently small value or reaches the preset maximum number of iterations. Finally, the path delay distribution matrix WH is obtained. Let D = WH, then we have Where K represents the number of paths and T represents the delay.

[0137] In summary, the loss function of the multi-task deep network output is designed to be the weighted sum of the multi-task outputs, that is,

[0138]

[0139] in:

[0140] and λ1 are the output entropy loss of the interference detection task and its corresponding weight, and λ2 are the mean square error of the intensity regression output and its corresponding weight, and λ3 are the non-negative matrix decomposition losses of the path delay distribution output and their corresponding weights, respectively. Through multi-task deep network training, λ1, λ2, and λ3 are adjusted according to different signal characteristics to minimize the loss function.

[0141] 2.4 Adaptive Signal Reconstruction Algorithm

[0142] Because the signal output by a multi-task deep network contains noise characteristics (interference signal type, interference signal strength, and path delay), to achieve high-quality adaptive signal reconstruction, the advantages of a Conditional Generative Adversarial Network (CGAN) and time-domain sparse coding are fully exploited to reconstruct the signal. The CGN learns the data distribution through an adversarial game between a generator and a discriminator. The generator receives random noise and conditional information and generates false signals in an attempt to deceive the discriminator. The discriminator distinguishes between real signals and false signals. The two are continuously optimized during the adversarial process, enabling the generator to produce more realistic signals. Time-domain sparse coding, on the other hand, is based on the assumption that signals are sparse in the time domain or a specific transform domain. By using sparsity constraints, it can remove redundant information from the signal, extract key features, and improve the signal's interpretability and robustness. Joint optimization combines cGAN with time-domain sparse coding. cGAN leverages its generative capabilities to capture the complex distribution of signals and generate near-realistic signals. TSC, through sparsity constraints, guides the generator to produce structured and sparse signals. Joint optimization considers both the adversarial loss of CGAN and the sparsity loss of TSC, ensuring that the signal generated by the generator meets both adversarial training requirements and sparsity requirements, achieving a complementary effect.

[0143] The core elements and models of the adaptive signal reconstruction algorithm using a joint optimization method of conditional generative adversarial networks and time-domain sparse coding are as follows:

[0144] (1) Generator G

[0145] The generator receives random noise and conditional information to generate a false signal. The random signal is consistent with the output attributes of the multi-task deep neural network (MT-DNN) (interference signal type, interference signal strength, and path delay), while the conditional information is various prior knowledge or specific requirements about the signal, such as signal type, frequency, and amplitude range. It learns to map the low-dimensional noise space to the high-dimensional signal space to generate samples that are as close to the real signal as possible. This can be expressed as a nonlinear mapping function:

[0146] G:(z,y)→G(z,y) (13)

[0147] in It is a random noise that is consistent with the signal noise characteristics output by the multi-task deep network (MT-DNN). is the conditional information, that is, the prior knowledge of the signal. is the generated reconstructed signal, d z , d y and d x are the dimensions of random noise, conditional information, and true signal, respectively.

[0148] (2) Discriminator D

[0149] It is used to distinguish between real signals and false signals generated by the generator. Through continuous learning, the ability to distinguish between real and false signals is improved, thereby prompting the generator to generate more realistic signals. The discriminator is also a nonlinear mapping function, and the output is a probability value:

[0150] D:(x,y)→D(x,y) (14)

[0151] Here, x represents the true signal, and y, defined as above, is the conditional information. The output is a scalar value between [0, 1], representing the probability that the input signal is the true signal. The closer D(x, y) is to 1, the closer the input signal is to the true signal. When the input is the generated signal G(z, y), it indicates that the input signal is the generated signal, and D(G(z, y), y) approaches 0.

[0152] (3) Time Domain Sparse Coding

[0153] Based on the sparse nature of signals in the time domain, the signals are sparsely represented and a suitable sparse coding algorithm is designed to represent the signals as a linear combination of a few non-zero coefficients, which helps to extract the main features of the signals and reduce redundancy.

[0154] Assume that the signal x is represented by the signal dictionary matrix Dic and the sparse coefficient α vector:

[0155] x≈Dicα (15)

[0156] in Is a signal dictionary matrix composed of multiple basis vectors, d x represents the dimension of the signal (i.e. how many types of signals there are), and k represents the number of each signal. is a sparse coefficient vector, and ||α||0 (zero norm, indicating the number of non-zero elements) or ||α||1 (one norm) is as small as possible.

[0157] (4) Loss function

[0158] The adversarial loss of the conditional generative adversarial network consists of the generator loss and the discriminator loss, where the generator loss is defined as

[0159]

[0160] represents the loss of the conditional generative adversarial network generator, G represents the generator, adv represents the adversarial network, y represents the input signal condition information, E[] represents the mathematical expectation, z~p z (z) indicates that the noise z generated complies with p z (z) distribution, G(z,y) represents the generator input condition information y and the p z (z) is the generated signal generated after the distributed noise, D(G(z,y),y) represents the probability of whether the output signal of the discriminator is a real signal after receiving the generated signal G(z,y) of the generator and the conditional information y, and log is the logarithm operation.

[0161] Similarly, the discriminator loss is defined as

[0162]

[0163] represents the loss of the conditional generative adversarial network generator, D() represents the discriminator output probability, x~p data (x) to express obedience to p data (x) is the true signal x of the distribution. D(x,y) is the probability of the discriminator outputting the true signal x and the conditional information y. E[] represents the mathematical expectation.

[0164] The sparsity loss is used to constrain the sparse representation of the generated signal. For the generated signal G(z,y), the sparsity loss is defined as:

[0165]

[0166] α is defined as the definition of simultaneous domain sparse coding, and the optimal solution can be obtained by solving the following formula.

[0167]

[0168] Among them, argmin means to find the minimum value, G(z,y) and Dic are defined as above. represents the second norm, and ||α||1 represents the first norm of α, i.e., its maximum value. λ is a trade-off parameter used to balance reconstruction error and sparsity. It should be adjusted according to the actual situation in the application.

[0169] (5) Joint losses

[0170] In summary, we have analyzed the adversarial loss and sparsity loss including generation loss and discrimination loss. In the adaptive signal reconstruction process, we need to consider the joint loss. The joint loss function is defined as the weighted sum of the adversarial loss and the sparsity loss:

[0171]

[0172] Where β is a weight parameter used to adjust the sparsity loss in the joint loss.

[0173] The goal of joint optimization of conditional generative adversarial networks and temporal sparse coding includes two parts: one is to maximize the discriminator loss to achieve discriminator optimization, and the other is to optimize the generator and sparse coding to minimize the joint loss.

[0174] In practical applications, an alternating optimization method is usually adopted, that is, first fix the generator and sparse coding module to maximize the discriminator loss function; then fix the discriminator, optimize the generator and sparse coding module, and continue to iterate until convergence.

[0175] Example 1:

[0176] Simulation test: 5G base station uplink interference suppression, such as Figure 12-17 shown.

[0177] Scenario: Urban high-rise environment, 100MHz bandwidth, 30kHz subcarrier spacing, 1024-QAM modulation. Vehicle terminal moving speed 0-80km / h;

[0178] Dataset: 100,000 sets of multipath signals (including labels: interference intensity, number of paths, 80% for training, 20% for testing);

[0179] Training parameters: Adam optimizer (learning rate 3e-4), batch size 128, training epoch 200s.

[0180] The following results are obtained through simulation analysis:

[0181] Bit error rate (BER): from 1.2×10 -3 Reduced to 3.5×10 -5 .

[0182] Signal-to-noise ratio improvement (SNR Gain): average improvement of 8.2dB;

[0183] Processing delay: <2ms.

[0184] Example 2:

[0185] A dynamic detection and adaptive suppression method for multipath interference based on multimodal deep learning includes multimodal signal acquisition, joint time-frequency preprocessing, multi-task deep network interference detection, conditional generative adversarial network signal reconstruction, and meta-learning online optimization steps.

[0186] The multi-task deep network (MT-DNN) simultaneously outputs interference existence probability, intensity classification and path delay distribution.

[0187] The signal reconstruction process introduces time domain sparse regularization constraints, and the optimization expression is:

[0188]

[0189] The specific meaning of each part of the expression can be found in formulas (16), (17), (18), (19) and (20) in the main text.

[0190] The key innovations of the present invention are:

[0191] Multimodal data fusion: Utilizing data from multiple modalities, such as signal strength, phase information, and spectral characteristics, and applying wavelet packet analysis, we can more comprehensively describe the characteristics of multipath interference and provide richer information for accurate detection and suppression of interference.

[0192] Application of deep learning models: The adversarial generative network model is used to automatically extract features from multimodal data, avoiding the complex manual feature engineering in traditional methods. It can learn more advanced and representative features, and improve the accuracy and robustness of detection and suppression.

[0193] Dynamic detection mechanism: It can monitor the changes of multipath interference in real time, adaptively adjust the detection strategy according to the dynamic changes of the environment and signals, and promptly detect multipath interference at different times and in different scenarios. Compared with static detection methods, it is more flexible and adaptable.

[0194] Adaptive suppression algorithm: Based on the detection results, it can automatically adjust the parameters and strategies of the suppression algorithm to adapt to multipath interference of different strengths and characteristics, effectively suppressing multipath interference while minimizing the impact on useful signals, thereby improving signal quality and system performance.

[0195] Model generalization capability: Through extensive data training and optimization, the model has strong generalization capabilities and can adapt to different communication environments, signal types, and interference scenarios. It has wider applicability and promotion value in practical applications.

[0196] This paper discloses a method for dynamic multipath interference detection and adaptive suppression based on multimodal deep learning. This method uses a multi-task deep network to jointly implement interference detection, intensity classification, and path separation. Furthermore, it utilizes a conditional generative adversarial network combined with time-domain sparsity constraints to optimize signal reconstruction. Simulation experiments demonstrate that this method can reduce the bit error rate by an order of magnitude in scenarios such as terrestrial radio communications and satellite communications, significantly outperforming traditional solutions.

[0197] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multipath interference detection and suppression method based on multimodal deep learning, characterized in that: The method comprises the following steps: a multi-mode signal acquisition step, a joint time-frequency preprocessing step, a multipath interference dynamic detection step and an adaptive signal reconstruction step; The multi-mode signal acquisition step is used to enable the receiving end to obtain radio signals through the antenna array, including: radio frequency signals, Doppler frequency shift data and environmental noise signals; A joint time-frequency preprocessing step is used to denoise and align the signal using wavelet packet transform (WPT) and adaptive Kalman filtering to achieve signal preprocessing; The dynamic detection step of multipath interference is used to detect the presence of interference, grade the intensity, and separate the paths of the signal after joint time-frequency preprocessing based on the multi-task deep neural network MT-DNN. The adaptive signal reconstruction step is used to jointly optimize signal reconstruction through conditional generative adversarial network (cGAN) and time-domain sparse coding.

2. The multipath interference detection and suppression method based on multimodal deep learning according to claim 1, characterized in that: The multimode signal acquisition step specifically includes: The radio signals received by the receiving antenna array from k paths at time t are expressed as x(t): In formula (1), α k (t): time-varying attenuation coefficient of the kth path; s k (t): radio signal of the kth path at time t; τ k : time-varying path delay; Doppler shift of the kth path; n(t): additive white Gaussian noise; j: imaginary unit, satisfying j 2 =-1.

3. The multipath interference detection and suppression method based on multimodal deep learning according to claim 1, characterized in that: The joint time-frequency preprocessing step includes: wavelet packet transform (WPT) denoising, wavelet packet denoising threshold processing, wavelet packet coefficient reconstruction and adaptive Kalman filtering for signal alignment; Wavelet packet transform (WPT) denoising involves discretizing a continuous radio signal x(t) into a discrete signal x(n), where n = 0, 1, ..., N-1, where n represents the number of sampling periods and N is the number of signal samples. Wavelet packet transform (WPT) decomposes x(n) step by step through a low-pass filter h(n) and a high-pass filter g(n) to obtain wavelet packet coefficients. Wavelet packet denoising threshold processing includes: using hard threshold or soft threshold to process wavelet packet coefficients; The wavelet packet coefficient reconstruction includes: after the wavelet packet denoising threshold processing, reconstructing the wavelet packet coefficient to obtain the denoised signal; Adaptive Kalman filtering for signal alignment includes: after the signal is reconstructed through wavelet packet denoising, adaptive Kalman filtering is applied to adjust the process noise covariance matrix Q in real time according to the changes in the signal. k and the measurement noise covariance matrix R k ;New information sequence through Kalman filter The statistical characteristics of z are adjusted, where z k is the signal measurement value at time k, H k is the observation matrix of the signal system, is the predicted value of the signal at time k, ∈ k is the new information of Kalman filter at time k; the maximum likelihood estimation method is used to minimize the negative log-likelihood function of the new information sequence to estimate Q k and R k Achieve signal alignment.

4. The multipath interference detection and suppression method based on multimodal deep learning according to claim 3, characterized in that: Wavelet packet transform WPT denoising specifically includes: S represents the low-frequency signal of the input signal x(n) after being processed by the low-pass filter h(n), and D represents the high-frequency signal of the input signal x(n) after being processed by the high-pass filter g(n); SS represents the low-frequency signal of the S signal after being processed by h(n) in the second-layer wavelet packet decomposition process, and SD represents the high-frequency signal of the signal after being processed by g(n) in the second-layer wavelet packet decomposition process; DS represents the low-frequency signal of the D signal after being processed by h(n) in the second-layer wavelet packet decomposition process, and DD represents the high-frequency signal of the signal after being processed by g(n) in the second-layer wavelet packet decomposition process, and the processing is carried out step by step according to this rule; Starting from the first layer, the filtering of the discrete signal x(n) is distributed according to the frequency, and its wavelet packet transform is realized by a set of filter banks; let h(n) be a low-pass filter and g(n) be a high-pass filter, j represents the number of decomposition layers, k represents the different frequency channels on the jth layer, n represents the discrete time point, which is the discrete time point index variable, m represents the traversal of all possible discrete time points, which is also the summation index variable; the wavelet packet coefficient d on the jth layer and the kth frequency channel j,k (n) is calculated by the following recursive formula: Initial condition: d 0,0 (n,n)=x(n) The decomposition process is shown in formula (2): Each wavelet packet coefficient d j,k (n) corresponds to the information of the signal in a specific time-frequency region; after wavelet packet transform, the discrete signal is decomposed into a series of wavelet packet coefficients with different frequencies and time resolutions. These coefficients form a time-frequency matrix, in which rows correspond to different time points or sampling points, and columns correspond to different frequency channels or scales; each element in the matrix represents the value of the wavelet packet coefficient at a specific time and frequency position, which reflects the strength of the signal at that time-frequency point.

5. The multipath interference detection and suppression method based on multimodal deep learning according to claim 4, characterized in that: Wavelet packet denoising threshold processing specifically includes: Hard threshold: Let λ be the threshold, wavelet packet coefficient d, and the coefficient dth after hard threshold processing is Soft threshold: The coefficient dth after soft threshold processing is Where sgn(d) is the sign function, that is 6. The multipath interference detection and suppression method based on multimodal deep learning according to claim 4, characterized in that: Wavelet packet coefficient reconstruction includes: set up and are the coefficients of the reconstructed low-pass filter and high-pass filter corresponding to the decomposition filters h(n) and g(n), respectively. Starting from the bottom layer, the reconstruction is gradually performed upwards. The reconstruction model from the jth layer to the j-1th layer is as shown in formula (6): Where n represents a discrete time point, m is the summation index variable; d j,k (n) is the wavelet packet coefficient of different layer j and different frequency channel k; Reconstruct low-pass filter coefficients High-pass filter coefficients The decomposition low-pass filter coefficient h(n) and the high-pass filter coefficient g(n) are in a dual relationship, satisfying the biorthogonal relationship and conjugate mirror relationship. The reconstructed filter coefficient is calculated by decomposing the filter coefficient; the decomposition and reconstruction process is represented by a matrix, and H and G are matrices composed of the decomposition low-pass filter h(n) and the high-pass filter g(n), respectively. and is reconstructed by the low-pass filter and high-pass filter The matrix formed by Where I is the identity matrix. By solving equation (7), we can get the reconstruction filter matrix and Then we get the reconstruction filter and 7. The multipath interference detection and suppression method based on multimodal deep learning according to any one of claims 1 to 6, characterized in that: In the multipath interference dynamic detection step, the multi-task deep network is a multi-task learning network, which is divided into four layers, from top to bottom: input layer, shared feature extraction layer, specific task layer and output layer; The input layer is used to receive the signal after joint time-frequency preprocessing, that is, the time-frequency matrix of the signal after wavelet packet transform (WPT); The shared feature extraction layer is used to apply the convolutional neural network (CNN) to extract the time-frequency features of the time-frequency matrix. This layer consists of three layers of convolutional structural blocks. Each convolutional structural block contains three core elements: a one-dimensional convolution Conv1D to extract time-frequency features, a batch normalization layer to accelerate training, and a ReLu activation function to introduce nonlinearity. The specific task layer is used to analyze specific tasks set around multipath signal analysis. The specific task networks designed include: multipath interference signal detection network, interference signal strength classification network and multipath signal path delay distribution network, which respectively realize multipath signal interference detection, multipath signal interference strength classification and multipath path delay distribution prediction of the signal processed by the shared feature extraction layer; The output layer outputs the prediction results of the three specific task networks corresponding to the specific task layer, namely, the interference signal existence rate output by the multipath signal interference detection network, the interference intensity level of the multipath interference signal intensity classification network, and the delay distribution of the multipath signal path delay distribution network.

8. The multipath interference detection and suppression method based on multimodal deep learning according to claim 7, characterized in that: Input of the shared feature extraction layer: time-frequency matrix decomposed by wavelet packet decomposition Where X represents the time-frequency matrix of the signal after wavelet packet decomposition, Indicates a matrix whose elements are real numbers, T represents the number of rows of the time-frequency matrix X, corresponding to different time points, and F represents the number of columns of the time-frequency matrix X, corresponding to different frequency channels; Network structure: 3 layers of convolution blocks, each layer contains Conv1D+BatchNorm+LeakyReLU, Output: Feature map Among them F c It is the feature map output by the convolution block, which is a three-dimensional real number matrix. T' represents the height of the feature map in space, corresponding to the time information, F' represents the width of the feature map in space, corresponding to the frequency information, and C represents the number of channels of the feature map, corresponding to a feature extraction result of the signal.

9. The multipath interference detection and suppression method based on multimodal deep learning according to claim 7, wherein: Input to the task-specific layer: feature map F c Features weighted by spatial attention; Network structure: bidirectional long short-term memory network LSMT, hidden layer dimension is 256; Output: Time series features Including interference signal existence matrix, interference signal strength distribution matrix and multipath signal delay distribution matrix; The output function of the multipath interference signal detection network uses a binary cross entropy loss function to measure the difference between the actual state of whether the output signal has interference in a multipath interference environment and the signal whether there is interference based on the multi-task deep network model prediction. For the real signal, either there is interference or there is interference. Let the label y be the label of whether the signal has interference. When y = 0, it means that the signal does not have interference, and when y = 1, it means that the signal has interference. Let the probability of signal interference predicted by the multi-task deep network model be p exist , then the binary cross entropy loss function is defined as: This loss function measures the predicted probability p exist The difference between the actual label y, when p exist When it is close to y, the loss value is smaller; in the specific task layer, after time-frequency shared feature extraction and specific task layer processing, a signal value is obtained, set as z, z is mapped to the (0,1) interval using the Sigmoid activation function in the last layer, indicating the probability of interference existence is output to the output layer; where the Sigmoid function is The interference signal strength classification network uses the mean square error loss function to measure the difference between the predicted interference strength and the actual interference strength; let the actual interference strength be y i , the interference intensity predicted by the interference signal strength classification network in the specific task layer is For n signal samples, the mean square error loss function is defined as The interference intensity feature is predicted using the fully connected layer of the network. The fully connected layer maps the previously extracted interference intensity feature to the numerical space of interference intensity [0,1]. The stronger the intensity, the closer the intensity value is to 1, and the weaker the intensity, the closer the intensity value is to 0. The non-negative matrix factorization loss of the multipath signal path delay distribution network is defined based on the objective of non-negative matrix factorization NMF; suppose the time-frequency matrix decomposed after wavelet packet decomposition is is decomposed into two non-negative matrices and That is, X≈WH, where T represents the rows of the time-frequency matrix, that is, the number of sampled signals, F represents the number of features or dimensions of the sampled signals, and k represents the dimension of the latent variable after decomposition, that is, the potential number of features, which is pre-set to k≤min(T,F); each column of W represents a potential path feature, and each row corresponds to a path; each column of H corresponds to a time point, and each column corresponds to a potential feature; the loss function is defined in the form based on Euclidean distance, that is, Where i and j represent the rows and columns of the matrix, corresponding to the path number and time point respectively; (WH) i,j Reflects the estimated delay of the i-th path at the j-th time point; ensure that W ≥ 0 and H ≥ 0; this loss function measures the difference between the decomposed matrix WH and the time-frequency matrix X. By minimizing this loss function, appropriate W and H are found to achieve the generation of path delay distribution; The update of W and H uses the following multiplication update rule: In formula (11), (W T X) j,k represents the "correlation" between the jth time-frequency matrix X signal feature and the kth potential feature, (W T WH) j,k Indicates the "reconstruction error correlation term" between the jth time-frequency matrix X signal feature and the kth potential feature in the current WH matrix; (XH T ) j,k Represents the "correlation" between the jth time-frequency matrix X signal feature and the kth potential feature, (WHH T ) j,k represents the "reconstruction error correlation term" between the jth time-frequency matrix X signal feature and the kth potential feature in the current WH matrix. In each iteration, H is updated first, then W is updated. That is, W is fixed first to update H, and then the updated H is fixed to update W until the loss function converges to a sufficiently small value or reaches the preset maximum number of iterations. Finally, the path delay distribution matrix WH is obtained. Let D = WH, then Where K represents the number of paths and T represents the delay; The loss function of the multi-task deep network output is designed as the weighted sum of the multi-task outputs, that is, in: and λ1 are the output entropy loss of the interference detection task and its corresponding weight, and λ2 are the mean square error of the intensity regression output and its corresponding weight, and λ3 are the non-negative matrix decomposition losses of the path delay distribution output and their corresponding weights, respectively. Through multi-task deep network training, λ1, λ2, and λ3 are adjusted according to different signal characteristics to minimize the loss function.

10. The multipath interference detection and suppression method based on multimodal deep learning according to claim 7, wherein: The adaptive signal reconstruction step includes: fully exploiting the advantages of conditional generative adversarial networks (cGANs) and time-domain sparse coding to reconstruct signals; wherein, the conditional generative adversarial network learns data distribution through an adversarial game between a generator and a discriminator, and the generator receives random noise and conditional information to generate a false signal in an attempt to deceive the discriminator; the discriminator distinguishes between real signals and generated false signals, and the two are continuously optimized during the adversarial process, so that the generator can generate more realistic signals; and time-domain sparse coding is based on the assumption that the signal is sparse in the time domain or a specific transform domain. Through sparsity constraints, it can remove redundant information in the signal, extract key features, and improve the interpretability and robustness of the signal; joint optimization is to combine cGAN with time-domain sparse coding, and cGAN uses its generation ability to capture the complex distribution of the signal and generate a signal close to the real one; time-domain sparse coding guides the generator to generate structured and sparse signals through sparsity constraints. During joint optimization, the adversarial loss of cGAN and the sparsity loss of time-domain sparse coding are simultaneously considered.

11. The multipath interference detection and suppression method based on multimodal deep learning according to claim 10, wherein: The adaptive signal reconstruction step specifically includes: (1) Generator G The generator accepts random noise and conditional information to generate a false signal. The random signal is random noise that is consistent with the output attributes of the multi-task deep network MT-DNN, including the type of interference signal, the strength of the interference signal, and the path delay. The conditional information is various prior knowledge or specific requirements about the signal. The generator learns to map the low-dimensional noise space to the high-dimensional signal space to generate samples that are as close to the real signal as possible, expressed as a nonlinear mapping function: G:(z,y)→G(z,y) (13) in It is a random noise that is consistent with the signal noise characteristics output by the multi-task deep network MT-DNN. is the conditional information, i.e., the prior knowledge of the signal; is the generated reconstructed signal, d z , d y and d x are the dimensions of random noise, conditional information, and true signal, respectively; (2) Discriminator D It is used to distinguish between real signals and false signals generated by the generator. Through continuous learning, the ability to distinguish between real and false signals is improved, thereby prompting the generator to generate more realistic signals. The discriminator is also a nonlinear mapping function, and the output is a probability value: D:(x,y)→D(x,y) (14) Where x represents the real signal, and y is defined as before, i.e., conditional information. The output is a scalar value between [0, 1], which represents the probability that the input signal is the real signal. The closer D(x, y) is to 1, the closer the input signal is to the real signal. When the input is the generated signal G(z, y), it indicates that the input signal is the generated signal, and D(G(z, y), y) approaches 0. (3) Time Domain Sparse Coding Based on the sparse characteristics of the signal in the time domain, the signal is sparsely represented and a suitable sparse coding algorithm is designed to represent the signal as a linear combination of a few non-zero coefficients; Assume that the signal x is represented by the signal dictionary matrix Dic and the sparse coefficient α vector: x≈Dicα (15) in Is a signal dictionary matrix composed of multiple basis vectors, d x represents the dimension of the signal, and k represents the number of each signal; is a sparse coefficient vector, and the zero norm ||α||0 or the one norm ||α||1 is as small as possible; (4) Loss function The adversarial loss of the conditional generative adversarial network consists of the generator loss and the discriminator loss, where the generator loss is defined as represents the loss of the conditional generative adversarial network generator, G represents the generator, adv represents the adversarial network, y represents the input signal condition information, E[] represents the mathematical expectation, z~p z (z) indicates that the noise z generated complies with p z (z) distribution, G(z,y) represents the generator input condition information y and the p z (z) The generated signal after the distributed noise, D(G(z,y),y) represents the probability of whether the output signal of the discriminator is a true signal after receiving the generated signal G(z,y) and the conditional information y. Log is the logarithm operation; Similarly, the discriminator loss is defined as represents the loss of the conditional generative adversarial network generator, D() represents the discriminator output probability, x~p data (x) to express obedience to p data (x) is the real signal x of the distribution; D(x,y) is the probability of whether the discriminator outputs the real signal after receiving the real signal x and the conditional information y; E[] represents the mathematical expectation; The sparsity loss is used to constrain the sparse representation of the generated signal. For the generated signal G(z,y), the sparsity loss is defined as: The optimal solution is obtained by solving the following equation: Among them, argmin means to find the minimum value, G(z,y) and Dic are defined as above. represents the second norm, ||α||1 represents the first norm of α, i.e. the maximum value; λ is a trade-off parameter used to balance the reconstruction error and sparsity; (5) Joint losses In the adaptive signal reconstruction process, the joint loss is considered, and the joint loss function is defined as the weighted sum of the adversarial loss and the sparsity loss: Where β is a weight parameter used to adjust the sparsity loss in the joint loss; The goal of the joint optimization of conditional generative adversarial networks and time-domain sparse coding includes two parts. The first is to maximize the discriminator loss to achieve discriminator optimization, and the second is to optimize the generator and sparse coding to minimize the joint loss. An alternating optimization method is adopted, that is, the generator and sparse coding modules are first fixed to maximize the discriminator loss function; then the discriminator is fixed, the generator and sparse coding modules are optimized, and the iteration is continued until convergence.

Images