Unmanned vehicle high-reliability communication system

Abstract

The application discloses an unmanned vehicle high-reliability communication system, and relates to the technical fields of data processing and artificial intelligence, and comprises a data acquisition module, a data preprocessing module, a data time-frequency feature extraction module, a data enhancement model training module and a signal guarantee module.The data acquisition module is used for generating signal samples based on original communication signals; the data preprocessing module is used for generating preprocessed signals by processing the signal samples; the data time-frequency feature extraction module is used for generating multi-scale time-frequency features; the data enhancement model training module is used for constructing and training a data enhancement model with a three-level structure of an encoder-bottleneck layer-decoder; and the signal guarantee module is used for constructing an unmanned vehicle communication signal guarantee mechanism based on the communication signal data enhancement model.The application has a communication enhancement system with self-adaptive denoising capability, dynamic spectrum feature extraction capability and high-fidelity reconstruction capability, and can provide a stable and reliable signal guarantee solution in scenarios such as unmanned vehicles which have extremely high requirements for communication quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and artificial intelligence, in particular to an unmanned vehicle high-reliability communication system. BACKGROUND

[0002] In modern intelligent transportation systems, unmanned vehicles rely on wireless communication systems for path planning, environmental perception and cooperative control. The quality of the communication signals directly determines the real-time performance and safety of the system. However, the actual application environment is complex and variable, and the communication signals often face various interference factors, such as multipath effect, Gaussian noise, impulse interference and frequency selective fading, etc., which leads to serious degradation of signal quality, reduction of signal-to-noise ratio and increase of bit error rate, thereby affecting the instruction reception and sensor data synchronization of unmanned vehicles. Traditional communication signal processing methods, such as fixed threshold filtering, Wiener filtering and wavelet denoising, usually use static parameters or linear models, which are difficult to cope with the non-stationary characteristics of signals and complex noise patterns, and have problems such as poor denoising effect, serious detail loss and weak adaptability, which cannot meet the demand of high-reliability communication. At the same time, existing feature extraction and enhancement methods are mostly based on fixed frequency band division and standard activation function design, ignoring the law of time-varying time-frequency characteristics of communication signals, resulting in the enhanced results failing to fully retain key features. In the aspect of deep learning application, traditional signal enhancement models generally have problems such as unreasonable initialization, rigid compression mechanism and insufficient reconstruction capability, which affect the model performance and deployment efficiency.

[0003] Therefore, there is an urgent need for a communication enhancement system with adaptive denoising capability, dynamic spectral feature extraction capability and high-fidelity reconstruction capability to provide stable and reliable signal guarantee solutions in scenarios such as unmanned vehicles that have extremely high requirements for communication quality. SUMMARY

[0004] The purpose of the present application is to provide an unmanned vehicle high-reliability communication system, which can have a communication enhancement system with adaptive denoising capability, dynamic spectral feature extraction capability and high-fidelity reconstruction capability, and can provide stable and reliable signal guarantee solutions in scenarios such as unmanned vehicles that have extremely high requirements for communication quality.

[0005] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0006] In a first aspect, the present application provides an unmanned vehicle high-reliability communication system, comprising:

[0007] A data acquisition module is configured to capture original communication signals in real time in combination with a multi-channel radio frequency front end and a distributed antenna array, and to generate signal samples based on the original communication signals;

[0008] A data preprocessing module is configured to process the signal samples in three stages in combination with local statistical characteristics and an adaptive weight function to generate preprocessed signals;

[0009] The data time-frequency feature extraction module is used to process the preprocessed signal through dynamic frequency band clustering and adaptive wavelet basis function to generate multi-scale time-frequency features.

[0010] The data augmentation model training module is used to build and train a data augmentation model with a three-level structure of "encoder-bottleneck layer-decoder";

[0011] The signal assurance module is used to build a communication signal assurance mechanism for unmanned vehicles based on the communication signal data enhancement model.

[0012] Optionally, the data preprocessing module is further configured to:

[0013] (1) For each signal sample and the corresponding time point, calculate the local mean and variance of the signal sample by using a sliding time window;

[0014] (2) Based on the local mean of the signal sample, the variance of the signal sample and the original communication signal, a denoised signal is generated by an adaptive exponential decay function.

[0015] (3) Calculate the preprocessed signal for each sample based on the global mean and global standard deviation of the denoised signal to generate the preprocessed signal.

[0016] Optionally, generating the denoised signal using an adaptive exponential decay function further includes:

[0017] ;

[0018] In the formula, For the first The denoised signal of a sample at time t is a scalar; For the first The original communication signal of a sample at time t is a scalar; This is the noise sensitivity coefficient, which controls the noise suppression strength. A larger value indicates stronger suppression of outliers. For example, ; It is a natural exponential function; This is to smooth out the terms and prevent the denominator from being 0.

[0019] Optionally, the data time-frequency feature extraction module is further configured to:

[0020] (1) Based on the Fourier transform amplitude of the preprocessed signal, the total number of dynamic frequency bands and the center frequency are calculated by clustering algorithm;

[0021] (2) Calculate wavelet basis functions for extracting local time-frequency features based on the center frequency and time center point;

[0022] (3) Calculate adaptive weights based on Fourier transform amplitude and center frequency;

[0023] (4) Combining adaptive weights, wavelet basis functions and preprocessed signals, a multi-scale time-frequency feature is calculated by weighted aggregation.

[0024] Optionally, the wavelet basis function for extracting the local time-frequency feature further comprises:

[0025] The wavelet basis function for extracting the local time-frequency feature is calculated based on a center frequency and a time center point, and a time window width of the wavelet basis function is dynamically adjusted according to the center frequency, and is expressed as:

[0026] ;

[0027] In the formula, is the wavelet basis function of the i th frequency band, used for extracting the local time-frequency feature; is the wavelet basis function of the i th frequency band, used for extracting the local time-frequency feature; is the circular constant; is the time center point of the i th frequency band, and the initial value is the time midpoint of the signal; is the time window width, ; is the central angle frequency; is the imaginary unit; is the central angle frequency; is the central frequency of the i th frequency band, representing the energy concentration area of the frequency band; is the central frequency of the i th frequency band, representing the energy concentration area of the frequency band; is the frequency band clustering index.

[0028] Optionally, the data enhancement model training module is further used for:

[0029] (1) A deep coding and decoding architecture is used to construct a communication signal data enhancement model composed of a three-level structure of symmetric "encoder-bottleneck layer-decoder";

[0030] (2) An initialization weight matrix of the encoder is generated through a double mechanism of feature statistical mean and frequency band clustering correction;

[0031] (3) A deep encoder is constructed through a gated recurrent unit and an improved adaptive activation function;

[0032] (4) An actual compression rate of the encoded feature vector is generated through a dynamic sparse mask and an energy adaptive threshold mechanism;

[0033] (5) A reconstructed time-frequency feature matrix is generated through a gated recurrent unit-deconvolution hybrid decoder and a residual connection mechanism;

[0034] (6) A total loss function is calculated through a joint reconstruction loss and a signal-to-noise ratio quality loss ;

[0035] (7) The total loss function is based on An adaptive moment estimator optimizer is used for backpropagation to update the parameters of the data augmentation model.

[0036] Optionally, the construction of the communication signal data enhancement model consisting of a symmetrical three-level structure of "encoder-bottleneck layer-decoder" further includes:

[0037] A deep codec architecture is used to enhance end-to-end communication signals, forming a data enhancement model that includes an encoder, a bottleneck layer, and a decoder.

[0038] The encoder is used to receive the time-frequency feature matrix and extract deep temporal features step by step through stacked gated recurrent unit layers;

[0039] Bottleneck layer, used to dynamically select key feature dimensions through adaptive sparse compression;

[0040] The decoder employs a hybrid design of gated recurrent units and deconvolutional layers;

[0041] Among them, the first-level gated recurrent unit reconstructs the temporal dependency relationship, and the second-level deconvolution layer performs feature upsampling to restore the original dimension;

[0042] The residual connection inputs the high-frequency components of the original time-frequency characteristics to the decoder output, forming a high-frequency compensation path;

[0043] The model output layer reconstructs the enhanced time signal through inverse time-frequency transformation.

[0044] Optionally, the calculation of the total loss function It also includes:

[0045] The joint reconstruction loss is calculated based on the reconstructed time-frequency characteristics, the reconstructed time signal, and the original communication signal. ;

[0046] The enhanced signal-to-noise ratio (SNR) is calculated based on the original signal spectrum and the reconstructed signal spectrum, and the quality loss is calculated using the enhanced SNR. ;

[0047] Based on joint reconstruction losses and quality loss Calculate the total loss function ;

[0048] ;

[0049] In the formula, The dynamic quality weight is calculated as follows: ;

[0050] Calculated for a moving average window; This is the total loss function.

[0051] Optionally, the signal guaranteeing module is further configured to:

[0052] output a final enhanced signal through a triple anti-interference design, and transmit the enhanced signal to a decision system;

[0053] the triple anti-interference design comprises:

[0054] at a signal level, the adaptive de-noising eliminates the multipath effect and dynamic amplitude fluctuation;

[0055] at a feature level, the dynamic sparse compression preserves key frequency band information;

[0056] at a reconstruction level, the residual connection restores high-frequency details.

[0057] According to the specific embodiments provided in the present application, the present application has the following technical effects:

[0058] 1. The present application proposes an exponential decay de-noising method based on local mean and variance, which effectively suppresses non-stationary impulse noise and multipath interference, while preserving the detailed features of the communication signal.

[0059] 2. The present application combines spectral clustering with multi-scale wavelet analysis to achieve dynamic division of signal frequency bands and high-resolution time-frequency feature extraction, adapting to the time-varying characteristics of the communication channel.

[0060] 3. The present application constructs an end-to-end deep model, adopts a gated recurrent unit and a residual high-frequency compensation mechanism, effectively solving the problem of high-frequency information loss in traditional methods.

[0061] 4. The present application enhances the signal-to-noise ratio as a quality indicator, and jointly constitutes a loss function with the reconstruction error to realize direct optimization of signal quality and improve the performance of the enhanced communication. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor under the premise of the drawings.

[0063] Figure 1 a functional module schematic diagram of a high-reliability communication system of an unmanned vehicle provided by the present application;

[0064] Figure 2 a performance comparison diagram of different de-noising methods in a low signal-to-noise ratio environment in the present application;

[0065] Figure 3Time-frequency features of fixed frequency band division provided for an embodiment in the present application;

[0066] Figure 4 Time-frequency features of dynamic frequency band clustering of the present technology provided for an embodiment in the present application;

[0067] Figure 5 Comparison diagram of bit error rate of different enhancement methods in complex noise environment provided for an embodiment in the present application;

[0068] Figure 6 Comparison diagram of communication success rate in different interference environments provided for an embodiment in the present application;

[0069] Figure 7 Comparison of different signal waveforms provided for an embodiment of the present application to visually display reconstruction effects, wherein, Figure 7 a is an original communication signal effect diagram, Figure 7 b is a diagram showing a noisy signal effect, Figure 7 c is a diagram showing a traditional method effect, Figure 7 d is a diagram showing a reconstructed signal effect of the present method. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0071] The above purposes, features and advantages of the present application can be more obvious and easy to understand. The present application will be described in further detail below with reference to the drawings and specific embodiments.

[0072] The video tag processing method provided by the embodiments of the present application can be applied to, for example, Figure 1 As shown in a high-reliability communication system for unmanned vehicles, characterized in that it comprises:

[0073] The data acquisition module is used to capture the original communication signal in real time in combination with the multi-channel radio frequency front end and the distributed antenna array, and generate signal samples based on the original communication signal;

[0074] The data preprocessing module is used to process the signal samples in three stages in combination with the local statistical characteristics and the adaptive weight function to generate preprocessed signals;

[0075] The data time-frequency feature extraction module is used to process the preprocessed signals through dynamic frequency band clustering and adaptive wavelet basis function to generate multi-scale time-frequency features;

[0076] The data enhancement model training module is configured to construct and train a data enhancement model with a three-level structure of an encoder, a bottleneck layer and a decoder.

[0077] The signal guarantee module is configured to construct an unmanned vehicle communication signal guarantee mechanism based on the communication signal data enhancement model.

[0078] Further, the unmanned vehicle captures original communication signals in real time through a multi-channel radio frequency front end and a distributed antenna array, and in a complex dynamic environment, the system adopts a wideband receiver to cover the working frequency band and synchronously collects mixed signals from direct paths and multi-path reflections.

[0079] In each sampling period, the analog-to-digital converter digitizes the received signal at a preset sampling rate, while recording the time stamp and channel state information, and the collection process strictly follows the signal integrity principle:

[0080] Firstly, the low-noise amplifier suppresses thermal noise;

[0081] Then, the band-pass filter eliminates out-of-band interference;

[0082] Finally, the discrete-time sequence containing amplitude and phase information is output.

[0083] The communication signal data collection module continuously monitors signal quality indicators such as received signal strength indicator (RSSI) and bit error rate (BER), and automatically switches to a backup antenna or triggers a retransmission mechanism when signal attenuation is detected, ensuring the continuity and completeness of the collected data.

[0084] The original signal is encapsulated and stored in units of samples, and each sample contains a continuous time-domain waveform of a fixed time length.

[0085] Optionally, the data preprocessing module, specifically, the unmanned vehicle communication signal has non-stationary characteristics caused by Gaussian noise, multipath effect noise and dynamic amplitude fluctuation, and conventional fixed threshold denoising methods will cause excessive smoothing in low signal-to-noise ratio areas and cannot effectively suppress impulse noise in high signal-to-noise ratio areas, and global normalization is difficult to adapt to the dynamic amplitude change of the signal.

[0086] The present application combines local statistical properties and adaptive weight functions to achieve noise-robust denoising and normalization through three-stage processing, and the specific steps are as follows:

[0087] (1) For each signal sample and corresponding time point, the local mean and variance of the signal sample are calculated through a sliding time window;

[0088] For each sample and each time point, a local mean value representing the average amplitude of the signal within the window is calculated using a sliding time window to offset the dynamic fluctuations of the signal, while a local variance reflecting the noise intensity is calculated to quantify the noise intensity and dynamic fluctuations, represented as:

[0089] ;

[0090] ;

[0091] wherein, is the local mean value of the i-th sample at t time, is a scalar, representing the average amplitude of the signal within the window, used to offset the dynamic fluctuations;

[0092] is the local variance of the i-th sample at t time, is a scalar, reflecting the noise intensity;

[0093] is the original communication signal of the i-th sample at t time, is a scalar;

[0094] is the time window size, dynamically set according to the sampling rate;

[0095] is the time index within the window, ;

[0096] is the current time point index.

[0097] It should be noted that a sample refers to a complete communication signal sequence, containing all time points, i.e., consecutive time points from to , and is the signal time length.

[0098] (2) Based on the local mean value of the signal sample, the variance of the signal sample, and the original communication signal, a denoised signal is calculated by an adaptive exponential decay function;

[0099] Based on the local mean value, the local variance, and the original communication signal, a denoised signal is calculated by an adaptive exponential decay function, which dynamically suppresses outlier noise points while preserving effective signal components, avoiding excessive smoothing or noise residue, represented as:

[0100] ;

[0101] wherein, is the i-th sample at t time, ​​​The denoised signal of a sample at time t is a scalar;

[0102] For the first The local mean of a sample at time t;

[0103] For the first The original communication signal of a sample at time t is a scalar;

[0104] This is the noise sensitivity coefficient, which controls the noise suppression strength. A larger value indicates stronger suppression of outliers. For example, ;

[0105] It is a natural exponential function;

[0106] For the first The local variance of a sample at time t;

[0107] To smooth out terms and prevent the denominator from being 0, for example, .

[0108] It should be noted that The term achieves nonlinear attenuation of noise intensity through local variance. The noise intensity is dynamically quantized and adaptively adjusted using an exponential decay function. When it is large, that is, the signal Deviation from local mean When the value is large, the exponential term rapidly approaches 0, thus effectively suppressing impulse noise and outliers; conversely, when the value is small, the exponential term rapidly approaches 0. When the signal is small, Deviation from local mean The exponential term is relatively small, rapidly approaching 1, preserving signal details. By considering the local characteristics of non-stationary signals, it avoids the problems of overly smoothing in low SNR regions or noise residue in high SNR regions caused by fixed threshold methods. In addition, the nonlinear characteristics of exponential decay provide a "soft threshold" effect in locally high deviation regions, avoiding signal distortion caused by hard truncation. It can not only effectively remove Gaussian noise, but also adaptively handle sudden interference caused by multipath effects, while preserving transient features such as steep edges of communication signals. This is something that conventional linear filtering or fixed threshold denoising cannot achieve in similar scenarios.

[0109] (3) Calculate the preprocessed signal for each sample based on the global mean and global standard deviation of the denoised signal to generate the preprocessed signal.

[0110] Based on the global mean and global standard deviation of the denoised signal, a preprocessed signal is calculated for each sample to eliminate the influence of amplitude fluctuations and achieve standardization, as shown below:

[0111] ;

[0112] In the formula, For the first The preprocessed signal of a sample at time t is a scalar;

[0113] For the first The global mean of a sample is a scalar, and its calculation method is expressed as:

[0114] ;

[0115] For the first The global standard deviation of a sample is calculated as follows:

[0116] ;

[0117] For smoothing terms, such as, ;

[0118] This represents the signal duration.

[0119] like Figure 2 As shown, in one embodiment, the performance of different denoising methods under low signal-to-noise ratio (SNR) conditions is compared to verify the performance advantage of the adaptive denoising method proposed in this invention in low SNR environments. The experiment compared four techniques: fixed threshold denoising, wavelet denoising, median filtering, and the method of this invention. Performance was evaluated by measuring the relationship between the input SNR (horizontal axis, decibels) and the output SNR (vertical axis, decibels). The figure shows that as the input SNR decreases (horizontal axis shifts to the left), the output SNR of the method of this invention is consistently higher than that of other methods, demonstrating that the adaptive exponential denoising mechanism of this invention can dynamically adjust the denoising intensity according to the local SNR: avoiding over-smoothing in low SNR regions and effectively suppressing impulse noise in high SNR regions. In the line graph, the curve of the method of this invention is not only at the highest position but also has the largest slope, indicating that its SNR improvement efficiency is optimal, which aligns with the design goal of dynamically quantifying noise intensity through local variance in the technical principle.

[0120] Optionally, the data time-frequency feature extraction module addresses the time-varying spectral characteristics of unmanned vehicle communication signals, such as multipath fading and frequency-selective fading. Traditional short-time Fourier transform uses a fixed window, which cannot balance time and frequency resolution, resulting in the loss of high-frequency transient features and the blurring of sudden pulse signals caused by insufficient time resolution when processing low-frequency bands. At the same time, static frequency band division is difficult to adapt to dynamic channel changes, reducing feature robustness.

[0121] The application realizes multi-scale time-frequency feature extraction through dynamic frequency band clustering and adaptive wavelet base function, and the specific steps are as follows:

[0122] (1) Based on the Fourier transform amplitude value of the preprocessed signal, the total number of dynamic frequency bands and the center frequency are calculated through a clustering algorithm;

[0123] Based on the Fourier transform amplitude value of the preprocessed signal, the total number of dynamic frequency bands and the center frequency are calculated through a clustering algorithm, the frequency bands are adaptively divided, the number and position of the frequency bands are optimized, and it is expressed as:

[0124] ;

[0125] In the formula, is the total number of dynamic frequency bands, the elbow method or the contour coefficient can be used to determine the initial value, and then the clustering algorithm is executed to optimize the value of ;

[0126] represents the minimum optimization mode of the total number of dynamic frequency bands and the center frequency , that is, the objective function is minimized to find the optimal total number of dynamic frequency bands and the center frequency ;

[0127] is the center frequency of the first frequency band, representing the energy concentration area of the frequency band;

[0128] is the Fourier transform amplitude value of the preprocessed signal at the frequency;

[0129] is the frequency set belonging to the first cluster;

[0130] is the frequency band clustering index;

[0131] is the frequency index.

[0132] (2) Based on the center frequency and the time center point, a wavelet base function for extracting local time-frequency features is calculated;

[0133] Based on the center frequency and the time center point, a wavelet base function for extracting local time-frequency features is calculated, and the time window width is dynamically adjusted according to the center frequency, and it is expressed as:

[0134] ;

[0135] ​wherein, is the wavelet basis function of the kth frequency band, used to extract local time-frequency features;

[0136] is the ratio of the circumference of a circle to its diameter;

[0137] is the time center point of the kth frequency band, the initial value is the time center point of the signal;

[0138] is the time window width, the window of high frequency band is narrower, and the calculation method is represented as ;

[0139] is the central angle frequency, such as, ;

[0140] is the imaginary unit.

[0141] (3) Adaptive weight is calculated based on the Fourier transform amplitude and the center frequency;

[0142] Adaptive weight is calculated based on the Fourier transform amplitude and the center frequency, reflecting the matching degree of the signal spectrum and the cluster center, and is represented as:

[0143] ;

[0144] wherein, is the adaptive weight of the kth frequency band at time t, the value range is ;

[0145] is different from the frequency band cluster index k;

[0146] is the center frequency of the kth frequency band, representing the energy concentration area of the frequency band;

[0147] is the weight attenuation coefficient, controlling the sensitivity of the frequency band selection, such as, ;

[0148] is the complex amplitude operation;

[0149] is the Fourier transform result of the signal .

[0150] (4) Combined with adaptive weight, wavelet basis function and preprocessed signal, multi-scale time-frequency features are calculated by weighted aggregation.

[0151] ​​​​Based on adaptive weights, wavelet basis functions and preprocessed signals, a time-frequency feature vector is calculated by weighted aggregation, which is expressed as:

[0152] ;

[0153] In the formula, is the time-frequency feature vector of the ith sample at time t, and is a column vector, and each element corresponds to a frequency band;

[0154] is the adaptive weight of the first frequency band at time t;

[0155] is the value of the wavelet basis function of the first frequency band at time t;

[0156] is the adaptive weight of the Kth frequency band at time t;

[0157] is the value of the wavelet basis function of the Kth frequency band at time t.

[0158] As Figures 3-4 shown, in one embodiment, the dynamic frequency band clustering and the fixed frequency band division are compared in time-frequency features, and the differences between the traditional fixed frequency band division and the dynamic frequency band clustering method of the application in time-frequency feature extraction are compared through a heat map. The horizontal coordinates of the two heat maps are both time (unit: second), and the vertical coordinates are both frequency (unit: hertz). The color intensity represents the feature intensity. The heat map of the left fixed frequency band division shows that the feature energy is uniformly distributed in the fixed frequency band, there is obvious strip-shaped segmentation, and the feature intensity does not change significantly in the time dimension. The heat map of the right application presents completely different features: the frequency band boundary is adaptively adjusted according to the signal energy, the feature energy is highly concentrated in a specific time and frequency region, and dynamically changes with time, verifying the effect of optimizing the number and position of frequency bands through the clustering algorithm in the technical principle. The dynamic frequency band clustering can accurately capture the time-varying spectral characteristics such as multipath fading and frequency-selective fading, and avoid the problem of loss of high-frequency transient features caused by fixed windows in traditional methods.

[0159] Optionally, the data enhancement model training module is further used for:

[0160] (1) adopting a deep coding and decoding architecture to construct a communication signal data enhancement model composed of a three-level structure of symmetric "encoder-bottleneck layer-decoder";

[0161] (2) generating an initialization weight matrix of the encoder through a double mechanism of feature statistical mean and frequency band clustering correction;

[0162] Further, the conventional weight initialization method has the problems of random initialization bias and noise sensitivity, the conventional Xavier or He initialization does not consider the time-frequency distribution characteristics of the communication signal, so that the initial feature extraction deviates from the real signal distribution, and at the same time, the random initialization is easily disturbed by high-frequency noise, which easily causes unstable early training of the model. The present application realizes data-driven weight initialization through the double mechanism of feature statistical mean and frequency band clustering correction, and the specific steps are as follows:

[0163] 1) Global statistics of feature matrix:

[0164] Based on the feature matrix of all samples, the global mean value calculation method of feature mapping is used to obtain the basic weight matrix, which is represented as:

[0165] ;

[0166] In the formula, is the basic weight matrix;

[0167] is the feature matrix of the i-th sample, and the dimension is is the transpose of the time-frequency feature vector ;

[0168] is the total number of training samples;

[0169] is the feature vector at t time.

[0170] 2) Calculate the frequency band correction increment:

[0171] Based on the feature matrix and the cluster center, the frequency band correction increment is calculated through the frequency band weighting factor matrix to adjust the basic weight, which is represented as:

[0172] ;

[0173] In the formula, is the frequency band correction increment;

[0174] is the global correction rate, such as 0.1;

[0175] is the Hadamard product;

[0176] is the frequency band weighting factor matrix, The calculation method of

[0177] ; ​

[0178] For The cluster center is calculated in advance by a k-means clustering algorithm;

[0179] The clustering influence coefficient is, for example, =2;

[0180] The Frobenius norm is F.

[0181] 3) Synthesis weight matrix:

[0182] The initialization weight is calculated based on the basic weight matrix and the frequency band correction increment, and the initialization weight of the encoder is generated by superimposing the two, which is represented as:

[0183] ;

[0184] In the formula, The initialization weight of the encoder is.

[0185] (3) Construct a deep encoder through a gated recurrent unit and an improved adaptive activation function;

[0186] Further, the communication signal enhancement faces two core challenges of insufficient modeling of temporal dependence and limited nonlinear feature extraction. When the traditional autoencoder uses a fully connected layer to process the time-frequency features, it cannot capture the long-term and short-term temporal correlation of the signal, and the standard activation function has poor adaptability to the time-varying spectrum, resulting in loss of features of sudden signals. The present application constructs a deep encoder through a gated recurrent unit and an improved adaptive activation function, and the specific steps are as follows:

[0187] 1) Define the adaptive activation function:

[0188] The adaptive activation function is calculated based on the input value and the slope adjustment coefficient and the threshold adjustment coefficient. The adaptive activation function can enhance the nonlinear expression ability in the internal transformation of the gated recurrent unit, and is represented as:

[0189] ;

[0190] In the formula, The adaptive activation function is used for nonlinear transformation inside the gated recurrent unit;

[0191] The input value of the adaptive activation function is;

[0192] The slope adjustment coefficient controls the nonlinear enhancement amplitude, for example, =0.2;

[0193] is the threshold adjustment coefficient, which determines the sensitive interval of the activation function, such as, = 0.3;

[0194] is the natural logarithm function.

[0195] 2) Constructing a gated recurrent unit time series encoder:

[0196] Based on the gating weight matrix, the preprocessed signal and the adaptive activation function, the encoding feature vector is calculated through the gated recurrent unit, and then the time series features are captured, which is expressed as:

[0197]

[0198] In the formula, is the forward propagation function of the gated recurrent unit;

[0199] is the internal parameter set of the encoder gated recurrent unit, including the weight matrix of the update gate and the reset gate;

[0200] is the encoding feature vector;

[0201] is the gating weight matrix, which is a trainable parameter, and its initial value is ;

[0202] is the gating bias vector, which is a trainable parameter, and its initial value is a full 0 vector.

[0203] It should be noted that the adaptive activation function is used for the internal nonlinear transformation of the gated recurrent unit, instead of the standard hyperbolic tangent activation function. In the calculation of the update gate and the reset gate of the gated recurrent unit, the input is transformed through the adaptive activation function, which can enhance the nonlinear expression ability, so that the gated recurrent unit can better adapt to the time-varying spectral characteristics of the communication signal, such as amplitude mutation and frequency selective fading. The slope adjustment coefficient and the threshold adjustment coefficient control the slope and sensitive interval of the activation function, and improve the capture ability of the model to burst signals.

[0204] (4) Through the dynamic sparse mask and energy adaptive threshold mechanism, the actual compression rate of the encoding feature vector is generated;

[0205] ​Further, the communication signal feature compression faces the key challenges of fixed sparsity defect and non-dynamic threshold problem, the traditional bottleneck layer adopts fixed compression rate, leading to high energy features being truncated or low energy noise being reserved, and the conventional method uses preset threshold, which cannot adapt to the time-varying characteristics of signal energy. The present application realizes intelligent compression of features through a dynamic sparse mask and an energy adaptive threshold mechanism, and the specific steps are as follows:

[0206] 1) Perform feature energy statistical calculation:

[0207] Calculate the feature energy median based on the encoded feature vector, which reflects the typical energy level of the feature vector, and is expressed as:

[0208] ;

[0209] In the formula, is the feature energy median, which reflects the typical energy level of the feature vector; is the median function.

[0210] 2) Dynamic sparse threshold generation:

[0211] Based on the overall energy of the encoded feature vector and the sparse sensitive coefficient, the dynamic sparse threshold coefficient is calculated through the Sigmoid function, which is used to adaptively adjust the threshold, and is expressed as:

[0212] ;

[0213] In the formula, is the dynamic sparse threshold coefficient, with a value range of (0, 1); is the sparse sensitive coefficient, which controls the threshold change rate, such as =0.5; is the L2 norm, which calculates the overall energy of the feature vector.

[0214] 3) Binary mask generation and sparse compression:

[0215] Based on the feature energy median, the dynamic sparse threshold coefficient and the encoded feature vector, a binary mask vector is generated, and a sparse coding vector is calculated, realizing feature selection, and is expressed as:

[0216] ;

[0217] In the formula, is the sparse coding vector; is the binary mask vector, and the element of the binary mask vector is ; is the element of the binary mask vector, and the calculation method is expressed as:

[0218] ;​

[0219] is the feature dimension index; is the i-th element of the encoded feature vector; is the dynamic sparse threshold coefficient.

[0220] 4) Perform actual compression rate calculation:

[0221] Based on the binary mask vector and the encoded feature vector dimension, the actual compression rate is calculated, which dynamically reflects the sparsity after feature selection, represented as:

[0222] ;

[0223] In the formula, is the actual compression rate; is the dimension of the encoded feature vector .

[0224] Further, the actual compression rate dynamically calculates the sparsity after feature selection, which is used for system monitoring and evaluation, close to 1, indicating high compression rate, that is, most features are discarded, reducing computational complexity and storage overhead, close to 0, indicating low compression rate, that is, more features are retained, which helps to retain detailed information.

[0225] (5) Generate the reconstructed time-frequency feature matrix through the residual connection mechanism of the gated recurrent unit-deconvolution hybrid decoder;

[0226] Further, signal reconstruction faces the core problems of loss of low-frequency components and insufficient detail recovery. Direct reconstruction by traditional decoders will weaken the low-frequency components of the signal, easily leading to time-domain waveform distortion, and ordinary upsampling operations are difficult to recover high-frequency transient features, affecting the time delay resolution of multipath signals. The present application realizes high-fidelity reconstruction of full-band signals through the residual connection mechanism of the gated recurrent unit-deconvolution hybrid decoder, the specific steps are as follows:

[0227] 1) Build a gated recurrent unit time sequence decoder:

[0228] Based on the first-level decoding weight, sparse encoding vector and gated recurrent unit, calculate the decoding feature vector to process the time sequence features, represented as:

[0229] ;

[0230] In the formula, is the decoding feature vector; is the first-level decoding weight, which is a trainable parameter; is the first-level decoding bias, which is a trainable parameter;​ To decode the internal parameter set of the decoder gating cycle unit, the structure of the decoder gating cycle unit is symmetrical to the encoder structure.

[0231] 2) Perform deconvolution feature upsampling:

[0232] Based on the decoding feature vector and the second-level decoding weight, the up-sampling feature matrix is calculated by deconvolution operation to restore the feature dimension, represented as:

[0233] ;

[0234] In the formula, is the up-sampling feature matrix; is a one-dimensional deconvolution operation; is the second-level decoding weight, representing the deconvolution kernel weight.

[0235] 3) Perform residual high-frequency compensation:

[0236] Based on the up-sampling feature matrix, the residual function, and the residual term weight coefficient, the reconstructed time-frequency feature matrix is calculated to enhance detail recovery through residual connection, represented as:

[0237] ;

[0238] In the formula, is the reconstructed time-frequency feature matrix of the th sample, which is the output of the decoder; is the residual function, which extracts the high-frequency component of the original signal, and the calculation method is represented as ; represents the average pooling operation;

[0239] is the residual term weight coefficient, which is used to adjust the contribution proportion of the residual signal in reconstruction, and the calculation method is represented as .

[0240] It should be noted that, The signal energy distribution is used as the weight adjustment basis to replace the empirical setting, which is dynamically set according to the proportion of high-frequency energy, ensuring that the detail recovery is enhanced when the high-frequency component is strong, and the noise introduction is reduced when the high-frequency is weak. The residual term weight coefficient is proportional to , which not only avoids the excessive amplification of noise when the fixed weight is weak in high frequency, but also enhances the detail recovery when the high frequency is strong.

[0241] (6) Calculate the total loss function by combining the reconstruction loss and the signal-to-noise ratio quality loss;

[0242] Further, the signal enhancement training faces the key problems of single loss limitation and lack of quality evaluation. The traditional mean square error loss only focuses on the time domain amplitude error, ignores the frequency domain features and signal-to-noise ratio improvement, and the conventional method lacks direct quantification of the communication quality of the enhanced signal, resulting in poor actual application effect. The present application constructs an end-to-end signal enhancement optimization target by combining the reconstruction loss and the signal-to-noise ratio quality loss, and the specific steps are as follows:

[0243] 1) Joint reconstruction loss calculation:

[0244] Based on the reconstructed time-frequency feature, time-frequency feature, reconstructed time signal and original communication signal, the joint reconstruction loss is calculated to quantify the time-frequency domain error, which is represented as:

[0245] ;

[0246] In the formula, is the joint reconstruction loss; is the batch sample number; is the th sample t time reconstruction time-frequency feature; is the th sample t time frequency feature; is the th sample t time reconstruction time signal, which is generated by inverse time-frequency transformation of the reconstructed time-frequency feature matrix output by the decoder, such as inverse wavelet transformation method based on the reconstructed time-frequency feature matrix to obtain the reconstructed time signal ; represents the short-time Fourier transform; is the frequency domain weight, such as =0.5.

[0247] It should be noted that the term represents a normalization factor, which combines the batch sample number with the time length to average the loss of each sample at each time step.

[0248] 2) Quality loss calculation:

[0249] Based on the original signal spectrum and the reconstructed signal spectrum, the enhanced signal-to-noise ratio is calculated, and the quality loss is calculated through the enhanced signal-to-noise ratio to optimize the signal quality, which is represented as:

[0250] ;

[0251] In the formula, is the enhanced signal-to-noise ratio;

[0252] is the original signal spectrum, which is obtained by performing short-time Fourier transform on the original communication signal The fast Fourier transform is performed to obtain;

[0253] To reconstruct the signal spectrum, the reconstructed time signal The fast Fourier transform is performed to obtain;

[0254] is the frequency index.

[0255] Further, the quality loss is defined as The calculation method is represented as .

[0256] It should be noted that the quality loss By minimizing the negative signal-to-noise ratio to maximize the enhanced signal-to-noise ratio Directly optimize the signal quality such as clarity and noise resistance, not just amplitude error, which can significantly improve signal clarity and improve the communication robustness of unmanned vehicles in complex environments.

[0257] 3) Calculate the total loss function:

[0258] Based on the joint reconstruction loss and the quality loss, the dynamic quality weight is calculated by the sliding average, and the total loss function is calculated based on the two, and then the optimization target is dynamically adjusted, which is represented as:

[0259] ;

[0260] In the formula, The calculation method is represented as ;

[0261] is the sliding average, and the window size is specifically set to an empirical value of about one training period, for example, the window size is set to 100 steps;

[0262] is the total loss function.

[0263] (7) Based on the total loss function , an adaptive moment estimation optimizer is used to update the data enhancement model parameters by backpropagation.

[0264] Further, based on the total loss function , an adaptive moment estimation optimizer is used to update the parameters by backpropagation:

[0265] In each batch training, first calculate the weighted sum of the joint reconstruction loss and the quality loss , and obtain the gradient of each layer by automatic differentiation;

[0266] Then, gradient clipping is implemented to prevent gradient explosion, and the learning rate is dynamically adjusted to balance the convergence speed and stability.

[0267] Finally, the sliding average strategy is applied to update the model weights, and the actual compression ratio of the validation set is monitored synchronously and the signal-to-noise ratio is enhanced .

[0268] When the early stopping mechanism detects that the validation loss has not improved for multiple consecutive periods, the current optimal model is saved and the training is terminated.

[0269] At the same time, according to the actual compression ratio , the storage resources are flexibly adjusted, when is close to 1, indicating a high compression ratio, i.e. most features are discarded, and the storage resources are reduced, when is close to 0, the storage resources are increased.

[0270] As shown in Figure 5 , in one embodiment, a comparative analysis of the bit error rate of different enhancement methods in a complex noise environment is performed, and the bit error rate performance of the communication enhancement technology is tested in four types of typical noise environments (Gaussian noise, impulse noise, multipath interference, and mixed noise) (vertical coordinate, unit: percentage). The bar chart shows that the method of the present application maintains the lowest bit error rate in all noise types, especially in the pulse noise and mixed noise environment, its bit error rate is less than 1 / 3 of other methods. The bit error rate of the traditional least mean square filter is close to 30% in the pulse noise, and the bit error rate of the Wiener filter and the spectral subtraction method is still more than 20% in the mixed noise, while the method of the present application always maintains below 10% under the same conditions, proving the effectiveness of the triple anti-interference design: the adaptive denoising of the signal layer eliminates the multipath effect, the dynamic sparse compression of the feature layer preserves the key frequency band information, and the residual connection of the reconstruction layer restores the high-frequency details.

[0271] As shown in Figure 6 , in this embodiment, the communication success rate in different interference scenarios is further verified, and five typical unmanned vehicle communication scenarios are simulated: urban environment (building obstruction), tunnel environment (closed space reflection), multipath fading (signal multi-path propagation), strong electromagnetic interference (industrial zone interference), and high-speed movement (Doppler effect), the horizontal coordinate is the interference type, and the vertical coordinate is the communication success rate (unit: percentage). The stacked bar chart shows that the method of the present application (red) has a success rate of more than 83% in all environments, especially in the tunnel environment and the strong electromagnetic interference scenario, the improvement is significant compared to the optimal comparative method, and the present application still maintains a success rate of 83% in the strong electromagnetic interference environment, while the fixed threshold denoising method has fallen below 53%, indicating that the triple anti-interference design has a significant effect, the adaptive denoising of the signal layer eliminates the pulse interference, the dynamic compression of the feature layer preserves the key frequency band, and the residual connection of the reconstruction layer restores the high-frequency details.

[0272] Optionally, the signal guaranteeing module is further configured to:

[0273] output a final enhanced signal through a triple anti-interference design, and transmit the enhanced signal to a decision system;

[0274] The triple anti-interference design comprises:

[0275] At a signal level, adaptive de-noising eliminates multipath effects and dynamic amplitude fluctuations;

[0276] At a feature level, dynamic sparse compression preserves key frequency band information;

[0277] At a reconstruction level, residual connection restores high-frequency details.

[0278] Specifically, the trained communication signal data enhancement model is deployed to perform model inference and unmanned vehicle communication signal guaranteeing:

[0279] After the original signal is acquired by the S1 acquisition module, it is sequentially input into the S2 preprocessing module (adaptive de-noising and standardization), the S3 time-frequency feature extraction module (dynamic frequency band clustering and wavelet transform) to generate a time-frequency feature vector for input into the enhancement model;

[0280] The model inference outputs a reconstructed time-frequency feature matrix, which is inversely transformed to generate a high signal-to-noise ratio time signal.

[0281] The unmanned vehicle communication signal guaranteeing mechanism achieves high reliability through a triple anti-interference design:

[0282] At a signal level, adaptive de-noising eliminates multipath effects and dynamic amplitude fluctuations;

[0283] At a feature level, dynamic sparse compression preserves key frequency band information;

[0284] At a reconstruction level, residual connection restores high-frequency details.

[0285] The final enhanced signal is transmitted to the decision system, significantly improving channel capacity and bit error rate performance, and ensuring reliable transmission of control instructions and synchronization of sensor data for unmanned vehicles in a strong interference environment.

[0286] As shown in FIG. 1, Figure 7 In one embodiment, communication signal reconstruction quality comparative analysis is performed, and the reconstruction effect is intuitively displayed through four groups of signal waveforms, Figure 7 wherein a is an original communication signal (the vertical coordinate is the signal amplitude) with clear double-frequency composite characteristics, Figure 7 wherein b shows that the noisy signal (signal-to-noise ratio -2.5 decibels) has been severely contaminated by noise, Figure 7The reconstruction results of traditional methods (Wiener filtering and wavelet denoising) still have obvious noise residues and waveform distortion in the middle c, Figure 7 The middle d proves that the reconstructed signal (SNR 12.8 decibels) of the method of the application is almost completely coincident with the original signal, especially the key feature points such as the wave peak and the wave trough are kept intact. The SNR values marked in the figure prove that the signal quality is improved by more than 15 decibels, and the fitting degree of the red reconstruction curve and the blue original signal reference line is much higher than that of the traditional method, which verifies the double advantages of the residual connection mechanism to restore high-frequency details and the joint reconstruction loss function to optimize time-frequency domain errors in the technical principle, and realizes the core goal of high-fidelity reconstruction of communication signals from noise.

[0287] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0288] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetic variable memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

[0289] The database involved in each of the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, and the like, without being limited thereto. The processor involved in each of the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, and the like, without being limited thereto.

[0290] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but it should be considered that any combination of the technical features is within the scope of the present disclosure, as long as there is no contradiction.

[0291] The principles and implementation modes of the present application are described by applying specific examples herein, and the above embodiments are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation modes and application ranges will be changed according to the idea of the present application. In summary, the content of the present description should not be understood as a limitation of the present application.

Claims

1. An unmanned vehicle high-reliability communication system, characterized by, Comprise: a data acquisition module for capturing raw communication signals in real time in combination with a multi-channel radio frequency front end and a distributed antenna array, and generating signal samples based on the raw communication signals; a data preprocessing module for generating preprocessed signals by processing the signal samples in three stages in combination with local statistical characteristics and an adaptive weight function; a data time-frequency feature extraction module for generating multi-scale time-frequency features by processing the preprocessed signals through dynamic frequency band clustering and an adaptive wavelet basis function; a data enhancement model training module for constructing and training a data enhancement model with a three-level structure of "encoder-bottleneck layer-decoder"; a signal guaranteeing module for constructing an unmanned vehicle communication signal guaranteeing mechanism based on the communication signal data enhancement model.

2. The high-reliability communication system for unmanned vehicles of claim 1, wherein, The data preprocessing module is further configured to: (1) for each signal sample and corresponding time point, calculate the local mean and variance of the signal sample through a sliding time window; (2) calculate a denoised signal through an adaptive exponential decay function based on the local mean of the signal sample, the variance of the signal sample, and the raw communication signal; (3) calculate a preprocessed signal for each sample based on the global mean and global standard deviation of the denoised signal to generate the preprocessed signal.

3. The high-reliability communication system for unmanned vehicles of claim 2, wherein The calculation of the denoised signal through the adaptive exponential decay function further comprises: ； wherein, is the de-noised signal of the i-th sample at time t, is a scalar; is the local mean of the i-th sample at time t; is the de-noised signal of the i-th sample at time t, is a scalar; is the local mean of the i-th sample at time t; is the original communication signal of the i-th sample at time t, is a scalar; is the de-noised signal of the i-th sample at time t, is a scalar; is the de-noised signal of the i-th sample at time t, is a scalar; is the local mean of the i-th sample at time t; is the local variance of the i-th sample at time t, is a smoothing term to prevent the denominator from being zero.

4. The high-reliability communication system for unmanned vehicles of claim 3, wherein, The data time-frequency feature extraction module is further configured to: (1) calculate the total number of dynamic frequency bands and center frequencies through a clustering algorithm based on the Fourier transform amplitude of the preprocessed signal; (2) calculate a wavelet basis function for extracting local time-frequency features based on the center frequency and the time center point; (3) calculate an adaptive weight based on the Fourier transform amplitude and the center frequency; (4) calculate multi-scale time-frequency features through weighted aggregation in combination with the adaptive weight, the wavelet basis function, and the preprocessed signal.

5. The high-reliability communication system for unmanned vehicles of claim 4, wherein, The wavelet basis function for extracting local time-frequency features further comprises: The wavelet basis function for extracting local time-frequency features is calculated based on the center frequency and the time center point, and the time window width of the wavelet basis function is dynamically adjusted according to the center frequency, represented as: ； In the formula, is the wavelet basis function of the first frequency band, used to extract local time-frequency features; is the ratio of a circle; is the wavelet basis function of the first frequency band, used to extract local time-frequency features; is the time window width, ; is the central angle frequency; is the imaginary unit; t is the time; is the central frequency of the first frequency band, representing the energy concentration area of the frequency band; is the frequency band clustering index.

6. The high-reliability communication system for unmanned vehicles of claim 5, wherein, The data enhancement model training module is further configured to: (1) construct a communication signal data enhancement model composed of a symmetrical three-level structure of "encoder-bottleneck layer-decoder" using a deep encoding and decoding architecture; (2) generate an initialization weight matrix of the encoder through a dual mechanism of feature statistical mean and frequency band clustering correction; (3) construct a deep encoder through a gated recurrent unit and an improved adaptive activation function; (4) generate an actual compression rate of the encoded feature vector through a dynamic sparse mask and an energy adaptive threshold mechanism; (5) generate a reconstructed time-frequency feature matrix through a gated recurrent unit-deconvolution hybrid decoder and a residual connection mechanism; (6) Calculate the total loss function by combining the reconstruction loss with the signal-to-noise ratio quality loss ; (7) based on the total loss function using an adaptive moment estimation optimizer to perform backpropagation update data augmentation model parameters.

7. The high-reliability communication system for unmanned vehicles of claim 6, wherein, The construction of the communication signal data enhancement model composed of a symmetrical three-level structure of "encoder-bottleneck layer-decoder" further comprises: Implement end-to-end communication signal enhancement using a deep encoding and decoding architecture to form a data enhancement model, including an encoder, a bottleneck layer, and a decoder; The encoder is configured to receive a time-frequency feature matrix and gradually extract deep time sequence features through stacked gated recurrent unit layers; The bottleneck layer is configured to dynamically select key feature dimensions through adaptive sparse compression; The decoder adopts a hybrid design of a gated recurrent unit and a deconvolution layer; The primary gated recurrent unit reconstructs the time sequence dependency, and the secondary deconvolution layer performs feature up-sampling to restore the original dimension; The residual connection inputs the high-frequency component of the original time-frequency feature into the output end of the decoder to form a high-frequency compensation path; The model output layer reconstructs the enhanced time signal through an inverse time-frequency transform.

8. The high-reliability communication system for unmanned vehicles of claim 7, wherein, the computing total loss function further comprising: based on the reconstructed time-frequency feature, the time-frequency feature, the reconstructed time signal, and the original communication signal, a joint reconstruction loss is calculated ; computing an enhanced signal-to-noise ratio based on the original signal spectrum and the reconstructed signal spectrum and computing a quality loss from the enhanced signal-to-noise ratio ; based on a joint reconstruction loss and quality loss computing a total loss function ; ； In the formula, is a dynamic mass weight, calculated as ; is calculated for the moving average window; is the total loss function.

9. The high-reliability communication system for unmanned vehicles of claim 8, wherein, The signal guaranteeing module is further configured to: Output a final enhanced signal through a triple anti-interference design, and transmit the enhanced signal to a decision system; The triple anti-interference design includes: At a signal level, adaptive de-noising eliminates multipath effects and dynamic amplitude fluctuations; At a feature level, dynamic sparse compression retains key frequency band information; At a reconstruction level, residual connection restores high-frequency details.

Images