Marine microwave signal time-frequency joint equalization method, system and equipment based on Transform model and medium

Abstract

The invention discloses an offshore microwave signal time-frequency joint equalization method, system and device based on a Transform model and a medium, and belongs to the technical field of deep learning models and signal processing, and the method comprises the steps: collecting and preprocessing offshore microwave radio frequency signal data, and obtaining a complex baseband signal sequence; generating a time-frequency spectrum matrix; constructing an auxiliary feature vector sequence; dividing the auxiliary feature vector sequence into time frequency units to form a token sequence; the token sequence is input into an improved Linformer-Transformer model, and a hidden representation sequence is obtained; outputting a combined equalization coefficient vector; carrying out amplitude-phase correction to obtain a balanced complex baseband signal sequence; and obtaining an equalized time-frequency spectrum matrix and a reconstructed complex baseband signal sequence. According to the invention, time-frequency joint equalization of signals is realized, the transmission reliability and signal quality of offshore microwave communication are improved, and the problems of amplitude distortion and phase distortion caused by multipath propagation, frequency selective fading and phase random drift in an offshore microwave channel are effectively solved.

Description

Technical Field

[0001] This invention relates to the fields of deep learning models and signal processing technology, specifically to a method, system, device, and medium for joint time-frequency equalization of marine microwave signals based on the Transformer model. Background Technology

[0002] With the increasing demand for maritime communication, microwave signals have been widely used in marine environments. Maritime microwave communication systems offer advantages such as high bandwidth, high capacity, and wide coverage, providing reliable data transmission services for ships, offshore platforms, and shore-based facilities. However, due to the complexity and instability of the marine environment, microwave signals are inevitably affected by multipath effects, frequency-selective fading, random noise, and time-varying interference caused by climate change during propagation. These factors lead to amplitude fading and phase distortion of the received signal, severely degrading the transmission quality and reliability of the system. Therefore, effectively achieving equalization of maritime microwave signals has become a core problem that current technologies urgently need to solve.

[0003] Existing technologies typically employ two types of equalization methods: time-domain equalization and frequency-domain equalization. Time-domain equalization is relatively simple to implement, but in complex environments like maritime microwave channels, which change rapidly over time and exhibit strong frequency selectivity, it often struggles to achieve compensation across the entire frequency band, easily resulting in residual distortion. Simple frequency-domain equalization typically fails to accurately capture time-varying channel characteristics, especially when multipath and phase distortion are superimposed; it often cannot achieve joint optimization in both time and frequency aspects, and the equalization effect remains limited. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention provides a method, system, device and medium for joint time-frequency equalization of marine microwave signals based on the Transformer model.

[0005] Therefore, the technical problem solved by this invention is: how to solve the problem that time-domain equalization often fails to take into account the compensation effect of the entire frequency band and is prone to residual distortion, and frequency-domain equalization usually fails to accurately capture time-varying channel characteristics.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a joint time-frequency equalization method for marine microwave signals based on the Transformer model, comprising: acquiring marine microwave radio frequency signal data and preprocessing it to obtain a complex baseband signal sequence; performing a short-time Fourier transform on the complex baseband signal sequence to generate a time-frequency spectrum matrix; performing amplitude normalization and phase decentering on the time-frequency spectrum matrix to construct an auxiliary feature vector sequence; dividing the auxiliary feature vector sequence into time-frequency units, performing linear mapping on each time-frequency unit, and superimposing time position coding and frequency position coding to form a token sequence; inputting the token sequence into an improved Linformer-Transformer model to obtain an implicit representation sequence; performing linear mapping and activation operations on the implicit representation sequence to output a joint equalization coefficient vector; using the joint equalization coefficient vector to perform amplitude and phase correction on the time-frequency spectrum matrix, and performing de-filtering compensation on the complex baseband signal sequence in the time domain to obtain an equalized complex baseband signal sequence; performing an inverse short-time Fourier transform and sign decision on the equalized complex baseband signal sequence to obtain the equalized time-frequency spectrum matrix and the reconstructed complex baseband signal sequence.

[0007] As a preferred embodiment of the time-frequency joint equalization method for marine microwave signals based on the Transformer model described in this invention, the step of acquiring marine microwave radio frequency signal data and performing preprocessing to obtain a complex baseband signal sequence includes: acquiring marine microwave radio frequency signals through a receiving antenna; performing down-conversion on the radio frequency signals to obtain an intermediate frequency signal; performing analog-to-digital conversion on the intermediate frequency signals and sampling at discrete sampling times to obtain a discrete complex sequence; performing frame synchronization and carrier frequency offset compensation on the discrete complex sequence to obtain a compensated discrete complex sequence; and performing sampling clock recovery on the compensated discrete complex sequence to output a complex baseband signal sequence.

[0008] As a preferred embodiment of the time-frequency joint equalization method for marine microwave signals based on the Transformer model described in this invention, the step of performing a short-time Fourier transform on the complex baseband signal sequence to generate a time-spectrum matrix includes: setting the sampling frequency and total length of the complex baseband signal sequence; dividing the complex baseband signal sequence into fixed frames according to the sampling frequency and sequence length to obtain a frame sequence; applying a window function consistent with the frame length to each frame of the frame sequence; padding the windowed frames with zeros at the end to a preset number of transform points to obtain an updated frame; performing a short-time Fourier transform on the updated frames to obtain a complex-valued coefficient sequence; and arranging all frames of the complex-valued coefficient sequence under each frequency index to form a time-spectrum matrix.

[0009] As a preferred embodiment of the time-frequency joint equalization method for marine microwave signals based on the Transformer model described in this invention, the step of performing amplitude normalization and phase decentering processing on the time-frequency spectrum matrix to construct an auxiliary feature vector sequence includes: performing amplitude normalization processing on the complex-valued coefficients of each time-frequency unit in the time-frequency spectrum matrix by dividing the amplitude of each complex-valued coefficient by the maximum value of the amplitudes of all complex-valued coefficients to obtain the amplitude normalization result; performing phase decentering processing on the complex-valued coefficients of each time-frequency unit in the time-frequency spectrum matrix by subtracting the average value of the phases of all complex-valued coefficients from the phase of each complex-valued coefficient to obtain the phase decentering result; and combining and arranging the amplitude normalization result and the phase decentering result to construct an auxiliary feature vector sequence.

[0010] As a preferred embodiment of the time-frequency joint equalization method for marine microwave signals based on the Transformer model described in this invention, the step of dividing the auxiliary feature vector sequence into time-frequency units, performing linear mapping on each time-frequency unit, and superimposing time position encoding and frequency position encoding to form a token sequence includes: dividing the auxiliary feature vector sequence into time-frequency units according to time index and frequency index, where the time index refers to the time dimension position of the time-spectrum matrix, and the frequency index refers to the frequency dimension position of the time-spectrum matrix; performing linear mapping on each time-frequency unit, and combining the auxiliary feature vectors under the time index and frequency index with the linear... Multiply the mapping weight matrices and add the bias vector to obtain the linearly mapped time-frequency unit vectors. Then, superimpose time position codes and frequency position codes onto each linearly mapped time-frequency unit vector to obtain a superimposed position-coded time-frequency unit vector. The time position code refers to a vector with a one-to-one correspondence between time indices, and the frequency position code refers to a vector with a one-to-one correspondence between frequency indices. Arrange all superimposed position-coded time-frequency unit vectors sequentially to form a token sequence. The length of the token sequence is equal to the product of the number of time indices and the number of frequency indices. The token sequence is a one-dimensional vector with the same dimension as the linearly mapped time-frequency unit vectors.

[0011] This preferred scheme introduces time and frequency dual-dimensional position coding, enabling the model to more accurately capture the time-frequency local features of the signal, thereby improving the completeness of feature representation and equalization accuracy.

[0012] As a preferred embodiment of the joint time-frequency equalization method for marine microwave signals based on the Transformer model described in this invention, the step of inputting the token sequence into an improved Linformer-Transformer model to obtain a latent representation sequence includes: performing a linear mapping on the token sequence to obtain a query vector, a key vector, and a value vector; and inputting the query vector, key vector, and value vector into the improved Linformer-Transformer model. The improved Linformer-Transformer model includes a self-attention module, a multi-scale convolution module, and a feedforward network module. The self-attention module calculates an attention output based on the query vector, key vector, and value vector. The multi-scale convolution module introduces local modeling capabilities based on the attention output. The feedforward network module performs nonlinear transformation and stabilization processing on the convolution fusion output to output the final latent representation sequence. In the self-attention module, learnable low-rank projection moments are applied to the key vector and value vector. The method involves obtaining the projected key vector and the projected value vector using a matrix approach, and calculating the attention output by the dot product of the query vector and the projected key vector. In the multi-scale convolution module, one-dimensional convolution operations with different kernels are performed on the attention output, and the convolution results are concatenated and fused to obtain the convolution fusion output. The one-dimensional convolution operation refers to the sliding weighted calculation of the attention output with different kernel sizes along the sequence length dimension. In the feedforward network module, residual connections and layer normalization are performed on the attention output and the convolution fusion output, and nonlinear mapping is performed through a feedforward fully connected network to obtain the hidden representation sequence. The process of performing linear mapping and activation operations on the hidden representation sequence to output a joint equalization coefficient vector includes: performing linear mapping on the hidden representation sequence, which consists of a linear mapping weight matrix and a bias vector; multiplying the hidden representation sequence with the linear mapping weight matrix and adding the bias vector to obtain the linear mapping result; performing nonlinear activation operations on the linear mapping result to obtain the activation result; and mapping the activation result to each time-frequency unit to output the joint equalization coefficient vector.

[0013] This preferred approach combines self-attention and multi-scale convolutional structures, taking into account both global dependencies and local modeling capabilities, thereby improving the model's expressiveness and anti-interference performance.

[0014] As a preferred embodiment of the joint time-frequency equalization method for marine microwave signals based on the Transformer model described in this invention, the step of performing amplitude and phase correction on the time-spectrum matrix using the joint equalization coefficient vector and performing de-filtering compensation on the complex baseband signal sequence in the time domain to obtain the equalized complex baseband signal sequence includes: performing amplitude correction on the time-spectrum matrix using the joint equalization coefficient vector by multiplying the amplitude of the complex coefficients of the complex coefficient sequence by the amplitude correction coefficient in the joint equalization coefficient vector to obtain the amplitude-corrected complex coefficients; and performing phase correction on the amplitude-corrected complex coefficients using the joint equalization coefficient vector by multiplying the phase value of the amplitude-corrected complex coefficients by the phase value of the joint equalization coefficients. The phase correction coefficients in the coefficient vector sequence are added to obtain the complex-valued coefficients after amplitude-phase correction; the complex-valued coefficients after amplitude-phase correction are arranged to obtain the amplitude-phase corrected time-spectrum matrix; based on the amplitude-phase corrected time-spectrum matrix, de-filtering compensation is performed on the complex baseband signal sequence to obtain the equalized complex baseband signal sequence. The de-filtering compensation refers to mapping the time-frequency domain correction information into a time-domain complex gain sequence and multiplying it sample by sample to construct an equivalent inverse filter and deconvolve it with the complex baseband signal sequence; the step of performing inverse short-time Fourier transform and symbol decision on the equalized complex baseband signal sequence to obtain the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence includes performing inverse short-time Fourier transform and symbol decision on the equalized complex baseband signal sequence. A continuous complex baseband signal sequence is formed by performing an inverse short-time Fourier transform (ISF) and then adding the results using frame-shifting and overlapping. The ISF refers to the operation of restoring the complex coefficients in the time-frequency domain to a continuous complex baseband signal sequence in the time domain frame by frame. The frame-shifting and overlapping addition involves weighting the frame-by-frame time-domain segments obtained from the ISF by a synthesis window function, then misaligning them on the time axis according to the frame shift, and adding them sample by sample in the overlapping region. Point-to-point normalization is performed based on the window energy superposition term to obtain a continuous complex baseband signal sequence. A sign decision is then performed on the continuous complex baseband signal sequence, assigning the real and imaginary parts of the sampled points to the nearest sign constellation point to obtain the decided complex signal. The baseband signal sequence, where the symbol constellation point refers to the complex values ​​of each sampling point in a continuous complex baseband signal sequence mapped to a set of symbol constellation points defined under a preset modulation scheme, with each symbol constellation point corresponding to a unique bit combination; the time-spectrum reconstruction of the decided complex baseband signal sequence is performed to obtain an equalized time-spectrum matrix. The time-spectrum reconstruction refers to framing and windowing the decided complex baseband signal sequence according to the frame length, frame shift, number of transform points, and analysis window function consistent with the forward time-frequency analysis, and performing consistent analysis transformation on each frame, arranging the obtained complex coefficients according to the time index and frequency index to generate the equalized time-spectrum matrix; the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence are output.

[0015] This preferred solution effectively enhances signal reconstruction quality and reduces amplitude and phase distortion and bit error rate through amplitude and phase separation correction and time-domain defiltering compensation.

[0016] This invention provides a joint time-frequency equalization system for marine microwave signals based on the Transformer model.

[0017] To address the aforementioned technical problems, this invention provides the following technical solution: a joint time-frequency equalization system for marine microwave signals based on the Transformer model, comprising: a data acquisition and processing module, a time-spectrum matrix generation module, an auxiliary feature vector construction module, a token sequence generation module, a model input module, an equalization coefficient generation module, a correction module, and an output module; the data acquisition and processing module is used to acquire marine microwave radio frequency signal data and perform preprocessing to obtain a complex baseband signal sequence; the time-spectrum matrix generation module is used to perform a short-time Fourier transform on the complex baseband signal sequence to generate a time-spectrum matrix; the auxiliary feature vector construction module is used to perform amplitude normalization and phase decentering processing on the time-spectrum matrix to construct an auxiliary feature vector sequence; the token sequence generation module is used to generate the auxiliary feature vector sequence. The vector sequence is divided into time-frequency units. A linear mapping is performed on each time-frequency unit, and time position encoding and frequency position encoding are superimposed to form a token sequence. The model input module is used to input the token sequence into the improved Linformer-Transformer model to obtain the latent representation sequence. The equalization coefficient generation module is used to perform linear mapping and activation operations on the latent representation sequence and output a joint equalization coefficient vector. The correction module is used to perform amplitude and phase correction on the time-spectrum matrix using the joint equalization coefficient vector and perform de-filtering compensation on the complex baseband signal sequence in the time domain to obtain the equalized complex baseband signal sequence. The output module is used to perform inverse short-time Fourier transform and sign decision on the equalized complex baseband signal sequence to obtain the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence.

[0018] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the aforementioned method for joint time-frequency equalization of marine microwave signals based on the Transformer model.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned method for joint time-frequency equalization of marine microwave signals based on the Transformer model.

[0020] The beneficial effects of this invention are as follows: By introducing an improved Linformer-Transformer model, this invention organically combines a self-attention mechanism, a low-rank projection matrix, and a multi-scale convolution module. This effectively reduces computational complexity while maintaining global dependency modeling capabilities and enhances local feature extraction capabilities. Compared to existing methods that rely solely on time-domain or frequency-domain equalization, this invention enables joint time- and frequency-domain modeling and equalization, effectively overcoming amplitude distortion and phase aberration problems caused by multipath propagation, frequency-selective fading, and random phase drift in maritime microwave channels.

[0021] This invention ensures the stability of input features by performing amplitude normalization and phase decentering on the time-spectrum matrix; it maintains the correspondence of the signal in the time-frequency structure by introducing time position encoding and frequency position encoding into the auxiliary feature vector sequence, thereby improving the model's ability to perceive the dynamic characteristics of the signal; and it achieves accurate equalization of complex baseband signal sequences by generating a joint equalization coefficient vector on the implicit representation sequence and performing amplitude and phase correction and defiltering compensation in the time domain.

[0022] This invention recovers the time-spectrum matrix after equalization and the reconstructed complex baseband signal sequence through inverse short-time Fourier transform and symbol decision, thereby significantly improving the reliability and equalization accuracy of microwave signal transmission in complex maritime environments and ensuring the stability and signal quality of the communication link. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 The following is a flowchart of a time-frequency joint equalization method for marine microwave signals based on the Transformer model, provided as an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the module structure of an improved Linformer-Transformer model for a joint time-frequency equalization method for marine microwave signals based on the Transformer model, provided as an embodiment of the present invention. Detailed Implementation

[0026] To make the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a joint time-frequency equalization method for marine microwave signals based on the Transformer model, comprising: S1. Collect marine microwave radio frequency signal data and perform preprocessing to obtain complex baseband signal sequences.

[0028] S2. Perform a short-time Fourier transform on the complex baseband signal sequence to generate a time-spectrum matrix.

[0029] S3. Perform amplitude normalization and phase decentering on the time-frequency spectrum matrix to construct an auxiliary feature vector sequence.

[0030] S4. Divide the auxiliary feature vector sequence into time-frequency units, perform linear mapping on each time-frequency unit, and superimpose time position encoding and frequency position encoding to form a token sequence.

[0031] S5. Input the token sequence into the improved Linformer-Transformer model to obtain the hidden representation sequence.

[0032] S6. Perform linear mapping and activation operations on the hidden representation sequence and output the joint equilibrium coefficient vector.

[0033] S7. The amplitude and phase of the time spectrum matrix are corrected by using the joint equalization coefficient vector, and the complex baseband signal sequence is defiltered and compensated in the time domain to obtain the equalized complex baseband signal sequence.

[0034] S8. Perform inverse short-time Fourier transform and sign decision on the equalized complex baseband signal sequence to obtain the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence.

[0035] It should be noted that during the propagation of marine microwave radio frequency signals in complex environments, they are often subject to various interferences such as multipath effects, frequency drift, and wind and wave disturbances, resulting in significant distortion of the signal in both the time and frequency domains. Traditional equalization methods are difficult to simultaneously take into account both time and frequency characteristics and non-stationary distortion compensation, which in turn affects modulation recognition and bit error rate performance.

[0036] Therefore, to address the above problems, this invention constructs an end-to-end time-frequency joint equalization processing flow, utilizes an improved Transformer structure to achieve deep feature modeling, and performs amplitude and phase synchronization compensation of the signal through joint coefficient control, effectively improving the equalization accuracy and signal reconstruction capability, and enhancing the system's signal processing robustness in dynamic and complex marine environments.

[0037] Example 2, refer to Figure 2 As one embodiment of the present invention, based on the previous embodiment, a joint time-frequency equalization method for marine microwave signals based on the Transformer model is provided, comprising: Furthermore, in step S1, marine microwave radio frequency signal data is acquired and preprocessed to obtain a complex baseband signal sequence, including the following steps A1-A5: A1. Acquire marine microwave radio frequency signals through the receiving antenna.

[0038] A2. Perform a down-conversion operation on the radio frequency signal to obtain an intermediate frequency signal.

[0039] A3. Perform analog-to-digital conversion on the intermediate frequency signal and sample at discrete sampling times to obtain a discrete complex number sequence.

[0040] A4. Perform frame synchronization and carrier frequency offset compensation on the discrete complex sequence to obtain the compensated discrete complex sequence.

[0041] A5. Perform sampling clock recovery on the compensated discrete complex sequence and output a complex baseband signal sequence.

[0042] Specifically, in step A1, the radio frequency signal is a modulated signal at the carrier frequency.

[0043] Specifically, in step A2, the down-conversion operation involves multiplying and mixing the local oscillator signal generated by the local oscillator with the radio frequency signal in a mixer. This shifts the center frequency of the radio frequency signal to a low intermediate frequency (IF), and then filtering out frequency and image interference using a frequency filter to obtain the IF signal. In this embodiment, the discrete complex sequence can be used to sample the intermediate frequency signal at discrete sampling times after analog-to-digital conversion. By performing I / Q quadrature demodulation in the analog domain and acquiring the I and Q components respectively, combined with low jitter clock, anti-aliasing filtering, automatic gain control, quantization and formatted output processing, a discrete sampling point sequence in complex form containing amplitude and phase information can be obtained.

[0044] In one alternative implementation, the discrete complex sequence can also be demodulated in the digital domain using digital down-conversion technology, and the I / Q data can be extracted in the FPGA in a dual-channel manner to form a discrete complex sequence.

[0045] In another alternative implementation, the discrete complex sequence can also be directly output by the general-purpose radio frequency receiver module in the software-defined radio platform. The complex sampled data stream, after digital demodulation and synchronization processing, can be buffered and calibrated before being input into the subsequent processing module as a discrete complex sequence.

[0046] This invention constructs a high-fidelity discrete complex sequence by performing high-precision I / Q demodulation and complex sampling in the analog domain. This effectively preserves the phase continuity and amplitude characteristics of the signal in a highly dynamic multipath environment, providing an accurate and low-noise basic data source for subsequent time-spectrum construction and equalization processing, thereby improving the overall demodulation robustness and equalization model training efficiency of the system.

[0047] Specifically, the sampling process in step A3 is as follows: during the analog-to-digital conversion, the intermediate frequency signal is first subjected to anti-aliasing filtering and automatic gain control, and a high-speed ADC is used for sampling. I / Q quadrature demodulation is performed in the analog domain and both I and Q channels are sampled simultaneously. A low-jitter clock is used during sampling, and quantization and calibration are performed simultaneously. Finally, the formatted discrete complex number sequence is output.

[0048] Specifically, in step A4, the compensation process involves multiplying the frame-aligned discrete complex sequence by an exponential compensation term with frequency offset.

[0049] Specifically, in step A5, sampling clock recovery refers to resampling the sampling points at the symbol time using an interpolator.

[0050] In this embodiment, the time-spectrum matrix in step S2 can divide the complex baseband signal sequence into fixed frames according to the set sampling frequency and total length. After applying a window function consistent with the frame length to each frame, zeros are padded to the preset number of transform points. Then, a short-time Fourier transform is performed on the updated frame to obtain a complex coefficient sequence. Finally, the complex coefficients of each frame are arranged according to the frequency index to form a matrix. The time-spectrum matrix includes a real part matrix and an imaginary part matrix, which reflects the amplitude and phase information of the signal at different time and frequency positions.

[0051] In an alternative implementation, the time-frequency matrix can also be reconstructed from the time-frequency distribution map generated by wavelet transform of the complex baseband signal sequence to obtain a two-dimensional coefficient matrix, thus forming a time-frequency structure that replaces the Fourier transform.

[0052] In another alternative implementation, the time-spectrum matrix can also be processed by a sliding window plus fast Fourier transform method to process complex baseband signal sequences, and an overlapping windowing and energy spectrum calibration method is introduced to form a time-spectrum matrix by splicing the spectra generated in different window segments.

[0053] This invention uses a high-resolution time-frequency matrix constructed based on short-time Fourier transform to accurately describe the time-frequency evolution characteristics of signals in complex marine environments under the influence of multipath effects and frequency drift. This provides a structurally complete and physically consistent data foundation for subsequent auxiliary feature extraction and equalization processing, thereby improving the learning ability of the equalization model and the accuracy of signal restoration.

[0054] Furthermore, in step S2, a short-time Fourier transform is performed on the complex baseband signal sequence to generate a time-spectrum matrix, including the following steps B1-B5: B1. Set the sampling frequency and total length of the complex baseband signal sequence. Based on the sampling frequency and sequence length, divide the complex baseband signal sequence into fixed frames to obtain the frame sequence.

[0055] B2. Apply a window function with the same length as the frame to each frame of the frame sequence, and pad the windowed frame with zeros at the end to the preset number of transformation points to obtain the updated frame.

[0056] B3. Perform a short-time Fourier transform on the updated frame to obtain a sequence of complex-valued coefficients.

[0057] B4. Arrange all frames of the complex-valued coefficient sequence under each frequency index to form a time-spectrum matrix.

[0058] In this embodiment, the window function in step B2 can apply a window function with the same frame length to each frame of complex baseband signal after division, perform edge smoothing transition processing, and suppress spectral leakage caused by frame truncation. The window function is a weighted coefficient sequence used to control the main lobe width and side lobe suppression performance in the frequency domain.

[0059] In an alternative implementation, the window function can also be a Kaiser window or a Blackman window, which have stronger sidelobe suppression capabilities, thereby enhancing the spectral separation capability and making it suitable for signal time-frequency analysis in multi-frequency interference scenarios.

[0060] In another alternative implementation, the window function can also be a dynamic window function, which adaptively adjusts the window function parameters according to the signal energy distribution within the frame, thereby achieving responsive weighting for local changes in the signal and improving the time-frequency local resolution of the analysis results.

[0061] This invention significantly reduces spectral leakage and boundary distortion by introducing a window function before the short-time Fourier transform, resulting in a more efficient time-frequency spectrum matrix with better frequency domain resolution and amplitude fidelity. This is beneficial for subsequent feature extraction and the training stability of the equalization model, and improves the overall time-frequency modeling accuracy of the system.

[0062] Specifically, in step B2, the window function refers to truncating a small segment of signal and smoothing the edges of that segment.

[0063] Specifically, in step B3, the short-time Fourier transform refers to mapping the updated frame to the frequency domain after Fourier transform. Each frame is decomposed into a set of complex-valued coefficients, and the frequency distribution that changes with time is obtained by processing frame by frame, forming a sequence of complex-valued coefficients. The amplitude of the complex-valued coefficients reflects the energy of the frequency components, and the phase of the complex-valued coefficients reflects the relative displacement.

[0064] Specifically, in step B4, the frequency spectrum matrix includes a real part matrix and an imaginary part matrix.

[0065] Furthermore, in step S3, the time-spectrum matrix is ​​subjected to amplitude normalization and phase decentering to construct an auxiliary feature vector sequence, including the following steps C1-C3: C1. Perform amplitude normalization on the complex-valued coefficients of each time-frequency unit in the time-frequency matrix. Divide the amplitude of each complex-valued coefficient by the maximum value of the amplitudes of all complex-valued coefficients to obtain the amplitude normalization result.

[0066] C2. Perform phase decentering processing on the complex-valued coefficients of each time-frequency unit in the time-frequency matrix. Subtract the average phase of all complex-valued coefficients from the phase of each complex-valued coefficient to obtain the phase decentering result.

[0067] C3. Combine and arrange the amplitude normalization result with the phase decentering result to construct an auxiliary feature vector sequence.

[0068] Furthermore, in step S4, the auxiliary feature vector sequence is divided into time-frequency units, a linear mapping is performed on each time-frequency unit, and time position encoding and frequency position encoding are superimposed to form a token sequence, including the following steps D1-D4: D1. Divide the auxiliary feature vector sequence into time-frequency units according to the time index and frequency index. The time index refers to the time dimension position of the time spectrum matrix, and the frequency index refers to the frequency dimension position of the time spectrum matrix.

[0069] D2. Perform linear mapping on each time-frequency unit. Multiply the auxiliary feature vectors under the time index and frequency index with the linear mapping weight matrix and add the bias vector to obtain the time-frequency unit vector after linear mapping.

[0070] D3. Superimpose time position code and frequency position code on each time-frequency unit vector after linear mapping to obtain the time-frequency unit vector after superimposed position code. Time position code refers to a vector with one-to-one time index, and frequency position code refers to a vector with one-to-one frequency index.

[0071] D4. Arrange all the time-frequency unit vectors after superimposed position encoding in order to form a token sequence. The length of the token sequence is equal to the product of the number of time indices and the number of frequency indices. The token sequence is a one-dimensional vector with the same dimension as the time-frequency unit vector after linear mapping.

[0072] In this embodiment, the improved Linformer-Transformer model in step S5 can take the token sequence as input, first generate query vector, key vector, and value vector through linear mapping, and then enter the improved Linformer-Transformer structure including a self-attention module, a multi-scale convolution module, and a feedforward network module. Among them, the self-attention module reduces the dimensionality of the key vector and value vector by applying a learnable low-rank projection matrix to obtain the attention output. The multi-scale convolution module introduces one-dimensional convolution operations with different kernel sizes on the basis of the attention output and splices and fuses them. The feedforward network module performs nonlinear mapping on the basis of residual connections and layer normalization to output the final hidden representation sequence.

[0073] In one alternative implementation, the improved Linformer-Transformer model can retain only the low-rank self-attention mechanism in the original Linformer structure, enhance representation capabilities with positional encoding, omit multi-scale convolutional modules and feedforward networks, simplify network complexity through shallow structure, and achieve lightweight deployment.

[0074] In another alternative implementation, the improved Linformer-Transformer model can replace the multi-scale convolutional module with a gated convolutional or sparse convolutional module, introduce a dynamic channel selection mechanism to enhance local feature adaptation, and add batch normalization to the feedforward network to improve numerical stability.

[0075] This invention significantly enhances the model's ability to model complex time-frequency coupling relationships of marine microwave signals by introducing a low-rank attention compression mechanism and a multi-scale convolution feature fusion structure. At the same time, it reduces the resource overhead caused by high-dimensional attention computation in traditional Transformer models, thereby improving the system's real-time processing performance and model convergence stability.

[0076] Furthermore, in step S5, the token sequence is input into the improved Linformer-Transformer model to obtain the hidden representation sequence, including the following steps E1-E4: E1. Perform a linear mapping on the token sequence to obtain the query vector, key vector, and value vector. Input these vectors into the improved Linformer-Transformer model. The improved Linformer-Transformer model includes a self-attention module, a multi-scale convolution module, and a feedforward network module. The self-attention module calculates the attention output based on the query vector, key vector, and value vector. The multi-scale convolution module introduces local modeling capabilities based on the attention output. The feedforward network module performs nonlinear transformations and stabilization on the convolutional fusion output, outputting the final latent representation sequence. Figure 2 As shown.

[0077] E2. In the self-attention module, a learnable low-rank projection matrix is ​​applied to the key vector and value vector to obtain the projected key vector and the projected value vector. The attention output is then calculated by the dot product of the query vector and the projected key vector.

[0078] E3. In the multi-scale convolution module, one-dimensional convolution operations with different convolution kernels are performed on the attention output, and the convolution results are spliced ​​and fused to obtain the convolution fusion output. One-dimensional convolution operation refers to the sliding weighted calculation of the attention output with different convolution kernel sizes in the sequence length dimension.

[0079] E4. In the feedforward network module, residual connections and layer normalization are performed on the attention output and the convolution fusion output, and nonlinear mapping is performed through the feedforward fully connected network to obtain the hidden representation sequence.

[0080] Specifically, the attention output is calculated in step E2 as follows: Where O represents the attention output and Q represents the query vector. This is the projected key vector. This is the projected value vector. The dimension of the key vector. For transpose, This is a normalization operation.

[0081] A low-rank projection matrix is ​​a learnable linear mapping that reduces the dimensionality of the key vector and value vector along the length dimension of the sequence.

[0082] Specifically, in step E4, the implicit representation sequence is represented as follows: Where H is the hidden representation sequence. and Here is the weight matrix of the feedforward network. and For bias terms, It is a non-linear activation function. For convolutional fusion output, This is a layer normalization operation.

[0083] Residual connection refers to adding the attention output and the convolutional output together, and then adding them together with the feedforward network output. Layer normalization refers to standardizing the result after residual connection in the feature dimension and scaling and shifting it using learnable parameters.

[0084] Furthermore, in step S6, linear mapping and activation operations are performed on the hidden representation sequence to output a joint equalization coefficient vector, including the following steps F1-F3: F1. Perform a linear mapping on the implicit representation sequence. The linear mapping consists of a linear mapping weight matrix and a bias vector. Multiply the implicit representation sequence by the linear mapping weight matrix and add the bias vector to obtain the linear mapping result.

[0085] F2. Perform a nonlinear activation operation on the linear mapping result to obtain the activation result.

[0086] F3. Map the activation results to each time-frequency unit and output the joint equalization coefficient vector.

[0087] Furthermore, in step S7, the amplitude and phase of the time-spectrum matrix are corrected using the joint equalization coefficient vector, and de-filtering compensation is performed on the complex baseband signal sequence in the time domain to obtain the equalized complex baseband signal sequence, including the following steps G1-G4: G1. Perform amplitude correction on the time spectrum matrix using the joint equalization coefficient vector. Multiply the amplitude of the complex coefficients in the complex coefficient sequence with the amplitude correction coefficient in the joint equalization coefficient vector to obtain the amplitude-corrected complex coefficients.

[0088] G2. Perform phase correction on the amplitude-corrected complex coefficients using the joint equalization coefficient vector. Add the phase value of the amplitude-corrected complex coefficients to the phase correction coefficients in the joint equalization coefficient vector sequence to obtain the amplitude-phase corrected complex coefficients.

[0089] G3. Arrange the complex-valued coefficients after amplitude and phase correction to obtain the amplitude and phase corrected time-frequency matrix.

[0090] G4. Based on the time-frequency spectrum matrix after amplitude and phase correction, perform de-filtering compensation on the complex baseband signal sequence to obtain the equalized complex baseband signal sequence. De-filtering compensation refers to mapping the time-frequency domain correction information into a time-domain complex gain sequence and multiplying it sample by sample to construct an equivalent inverse filter and deconvolve it with the complex baseband signal sequence.

[0091] Furthermore, in step S8, an inverse short-time Fourier transform and symbol decision are performed on the equalized complex baseband signal sequence to obtain the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence, including the following steps H1-H4: H1. Perform an inverse short-time Fourier transform on the equalized complex baseband signal sequence, and form a continuous complex baseband signal sequence by frame-shifting and overlapping addition. The inverse short-time Fourier transform refers to the operation of restoring the complex coefficients in the time-frequency domain to a continuous complex baseband signal sequence in the time domain frame by frame. Frame-shifting and overlapping addition refers to the process of weighting the frame-by-frame time domain segments obtained by the inverse short-time Fourier transform through a synthesis window function, placing them in a staggered manner on the time axis according to the frame shift, and adding them sample by sample in the overlapping area. Point-to-point normalization is performed based on the window energy superposition term to obtain a continuous complex baseband signal sequence.

[0092] H2. Perform symbol decision on the continuous complex baseband signal sequence, and decide the real and imaginary parts of the sampling points in the continuous complex baseband signal sequence to the nearest symbol constellation point, so as to obtain the complex baseband signal sequence after decision. The symbol constellation point refers to the set of symbol constellation points defined under the preset modulation mode, which maps the complex values ​​of each sampling point in the continuous complex baseband signal sequence. Each symbol constellation point corresponds to a unique bit combination.

[0093] H3. Perform time-spectrum reconstruction on the complex baseband signal sequence after the decision to obtain the equalized time-spectrum matrix. Time-spectrum reconstruction refers to dividing the complex baseband signal sequence after the decision into frames and applying windows according to the same frame length, frame shift, number of transform points and analysis window function as the forward time-frequency analysis, and performing consistent analysis transformation on each frame. The obtained complex-valued coefficients are arranged according to the time index and frequency index to generate the equalized time-spectrum matrix.

[0094] H4, output equalized time-frequency matrix and reconstructed complex baseband signal sequence.

[0095] Example 3 is an embodiment of the present invention, which provides a joint time-frequency equalization method for marine microwave signals based on the Transformer model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0096] This study focuses on a microwave communication link between an offshore platform and a near-shore base station. The link operates in the 15 GHz microwave band with a transmission bandwidth of 20 MHz and a transmission distance of approximately 25 kilometers. Due to the complexity of the marine environment, the signal is significantly affected by multipath effects, frequency-selective fading, and time-varying noise during transmission. Particularly when the sea is rough, the channel fading depth can exceed 20 dB, severely impacting communication stability and bit error rate. The proposed time-frequency joint equalization method for marine microwave signals based on an improved Transformer model is validated in this scenario, aiming to address the problem that existing time-domain or frequency-domain equalization methods cannot simultaneously consider global dependencies and local characteristics.

[0097] In the experiment, the acquired marine microwave radio frequency signal data was first preprocessed to obtain a complex baseband signal sequence. Then, a short-time Fourier transform was performed on the complex baseband signal sequence to generate a time-spectrum matrix, followed by amplitude normalization and phase decentering to ensure the stability of the input features. Subsequently, the processed features were divided into time-frequency units and positional codes were superimposed to form a token sequence. This token sequence was input into an improved Linformer-Transformer model. The model applied low-rank projection matrices to the key and value vectors in its self-attention mechanism and introduced multi-scale convolution modules into the network layers, which reduced computational complexity and enhanced local feature extraction capabilities. The model's output latent representation sequence was further mapped to a joint equalization coefficient vector, used to perform amplitude and phase correction on the time-spectrum matrix and de-filtering compensation on the complex baseband signal sequence in the time domain, thus obtaining the equalized complex baseband signal sequence. Finally, the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence were recovered through inverse short-time Fourier transform and sign decision.

[0098] To verify the beneficial effects of this invention, the proposed method is compared with traditional time-domain equalization, frequency-domain equalization, and equalization methods based on convolutional neural networks (CNN-EQ). Under the same experimental conditions, the performance of different methods in terms of multipath fading, bit error rate, and signal recovery quality is compared, as shown in Table 1.

[0099] Table 1 Comparative Experimental Results of Microwave Signal Equalization Methods at Sea

[0100] As shown in Table 1, traditional time-domain equalization methods can only improve the signal-to-noise ratio (SNR) to a limited extent when dealing with multipath fading, with an average improvement of less than 5 dB, while the bit error rate remains high, failing to meet the stability requirements of maritime communication. Frequency-domain equalization, to some extent, mitigates the impact of frequency-selective fading, increasing the SNR to 6.2 dB and reducing the bit error rate to 9.6 × 10⁻⁶. -3 However, it remains insufficient when dealing with rapidly time-varying channels. The CNN-EQ method compensates for channel distortion through deep learning, resulting in a significant performance improvement and reducing the bit error rate to 6.2×10⁻⁶. -3 However, due to its reliance on convolutional structures and lack of global dependency modeling capabilities, the signal recovery effect remains limited, and the computational overhead increases significantly.

[0101] In contrast, the method proposed in this invention combines low-rank projection and multi-scale convolution in the self-attention mechanism, which not only ensures the ability to capture long-range dependencies but also enhances local feature modeling. This results in an average signal-to-noise ratio improvement of 12.7 dB, a peak signal-to-noise ratio as high as 34.5, a mean squared error reduction to 0.0038, and a bit error rate of only 2.1 × 10⁻⁶. -3 Meanwhile, since low-rank projection effectively reduces the computational complexity of attention, the computational cost of this method is only 17.6 GFLOPs, which is lower than CNN-EQ, and the average latency is controlled at 22.3ms, which has good real-time performance.

[0102] Therefore, the experimental results fully verify the effectiveness of the present invention in performing time-frequency joint equalization of microwave signals in complex marine environments. This method can significantly reduce signal distortion and bit error rate, improve signal recovery quality and transmission reliability, while balancing accuracy and computational efficiency, and has practical engineering application value.

[0103] Example 4 is an embodiment of the present invention. This embodiment provides a joint time-frequency equalization system for marine microwave signals based on the Transformer model, including a data acquisition and processing module, a time-spectrum matrix generation module, an auxiliary feature vector construction module, a token sequence generation module, a model input module, an equalization coefficient generation module, a correction module, and an output module.

[0104] The data acquisition and processing module is used to acquire marine microwave radio frequency signal data and perform preprocessing to obtain complex baseband signal sequences.

[0105] The time-spectrum matrix generation module is used to perform short-time Fourier transform on complex baseband signal sequences to generate time-spectrum matrices.

[0106] The auxiliary feature vector construction module is used to perform amplitude normalization and phase decentering on the time-spectrum matrix to construct an auxiliary feature vector sequence.

[0107] The token sequence generation module is used to divide the auxiliary feature vector sequence into time-frequency units, perform linear mapping on each time-frequency unit, and superimpose time position encoding and frequency position encoding to form a token sequence.

[0108] The model input module is used to input the token sequence into the improved Linformer-Transformer model to obtain the hidden representation sequence.

[0109] The equalization coefficient generation module is used to perform linear mapping and activation operations on the implicit representation sequence and output a joint equalization coefficient vector.

[0110] The correction module is used to perform amplitude and phase correction on the time spectrum matrix using the joint equalization coefficient vector, and to perform de-filtering compensation on the complex baseband signal sequence in the time domain to obtain the equalized complex baseband signal sequence.

[0111] The output module is used to perform inverse short-time Fourier transform and sign decision on the equalized complex baseband signal sequence to obtain the equalized time-spectrum matrix and the reconstructed complex baseband signal sequence.

[0112] The present invention also provides an electronic device applicable to a time-frequency joint equalization method for marine microwave signals based on a Transformer model, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement a time-frequency joint equalization method for marine microwave signals based on a Transformer model as proposed in the above embodiments.

[0113] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements a joint time-frequency equalization method for marine microwave signals based on the Transformer model as proposed in the above embodiments.

[0114] The storage medium proposed in this invention and the implementation method of joint time-frequency equalization of marine microwave signals based on the Transformer model proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this invention can be found in the above embodiments, and this invention has the same beneficial effects as the above embodiments.

[0115] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for time-frequency joint equalization of offshore microwave signals based on a Transformer model, characterized in that: The method comprises the following steps: Collecting and preprocessing microwave radio frequency signal data on the sea to obtain a complex baseband signal sequence; Performing short-time Fourier transform on the complex baseband signal sequence to generate a time-frequency spectrum matrix; Performing amplitude normalization and phase decentralization processing on the time-frequency spectrum matrix to construct an auxiliary feature vector sequence; Dividing the auxiliary feature vector sequence into time-frequency units, performing linear mapping on each time-frequency unit, and superimposing time position coding and frequency position coding to form a token sequence; Inputting the token sequence into an improved Linformer-Transformer model to obtain a hidden representation sequence; Performing linear mapping and activation operation on the hidden representation sequence to output a joint equalization coefficient vector; Performing amplitude and phase correction on the time-frequency spectrum matrix by using the joint equalization coefficient vector, and performing de-filtering compensation on the complex baseband signal sequence in the time domain to obtain an equalized complex baseband signal sequence; Performing inverse short-time Fourier transform and symbol decision on the equalized complex baseband signal sequence to obtain an equalized time-frequency spectrum matrix and a reconstructed complex baseband signal sequence.

2. The offshore microwave signal time-frequency joint equalization method based on the Transformer model according to claim 1, characterized in that: The method comprises the following steps: Collecting and preprocessing microwave radio frequency signal data on the sea to obtain a complex baseband signal sequence, The method comprises the following steps: Collecting microwave radio frequency signals on the sea through a receiving end antenna; Performing frequency down-conversion operation on the radio frequency signals to obtain intermediate frequency signals; Performing analog-to-digital conversion on the intermediate frequency signals to sample at discrete sampling time to obtain a discrete complex sequence; Performing frame synchronization and carrier frequency offset compensation on the discrete complex sequence to obtain a compensated discrete complex sequence; 3. The offshore microwave signal time-frequency joint equalization method based on the Transformer model according to claim 2, characterized in that: Performing sampling clock recovery on the compensated discrete complex sequence to output a complex baseband signal sequence. The method comprises the following steps: Setting the sampling frequency and sequence total length of the complex baseband signal sequence, dividing the complex baseband signal sequence into fixed frames according to the sampling frequency and sequence length to obtain a frame sequence; Applying a window function consistent with the frame length to each frame of the frame sequence, zero-padding the windowed frame at the tail to a preset transformation point number to obtain an updated frame; Performing short-time Fourier transform on the updated frame to obtain a complex coefficient sequence; 4. The offshore microwave signal time-frequency joint equalization method based on the Transformer model according to claim 3, characterized in that: Arranging all frames of the complex coefficient sequence under each frequency index to form a time-frequency spectrum matrix. The method comprises the following steps: Performing amplitude normalization processing on the complex values of each time-frequency unit in the time-frequency spectrum matrix, dividing the amplitude of each complex value by the maximum value of all complex value amplitudes to obtain an amplitude normalization result; Performing phase decentralization processing on the complex values of each time-frequency unit in the time-frequency spectrum matrix, subtracting the average value of all complex value phases from the phase of each complex value to obtain a phase decentralization result; 5. The offshore microwave signal time-frequency joint equalization method based on the Transformer model according to claim 4, characterized in that: Combining and arranging the amplitude normalization result and the phase decentralization result to construct an auxiliary feature vector sequence. The method comprises the following steps: The auxiliary feature vector sequence is divided into time-frequency units according to a time index and a frequency index, the time index refers to a time dimension position of a time-frequency spectrum matrix, and the frequency index refers to a frequency dimension position of the time-frequency spectrum matrix; A linear mapping is performed on each time-frequency unit, the auxiliary feature vector at the time index and the frequency index is multiplied by a linear mapping weight matrix and added with a bias vector to obtain a time-frequency unit vector after linear mapping; A time position code and a frequency position code are superimposed in each time-frequency unit vector after linear mapping to obtain a time-frequency unit vector after superimposing the position code, the time position code is a vector corresponding to the time index one by one, and the frequency position code is a vector corresponding to the frequency index one by one; All time-frequency unit vectors after superimposing the position code are sequentially arranged to form a token sequence, a sequence length of the token sequence is equal to a product of a time index quantity and a frequency index quantity, and the token sequence is a one-dimensional vector consistent with a time-frequency unit vector dimension after linear mapping.

6. The offshore microwave signal time-frequency joint equalization method based on the Transformer model according to claim 4, characterized in that: The token sequence is input into the improved Linformer-Transformer model to obtain a hidden representation sequence, including, A linear mapping is performed on the token sequence to obtain a query vector, a key vector and a value vector, and the query vector, the key vector and the value vector are input into the improved Linformer-Transformer model, the improved Linformer-Transformer model includes a self-attention module, a multi-scale convolution module and a feedforward network module, the self-attention module calculates attention output based on the query vector, the key vector and the value vector, the multi-scale convolution module introduces local modeling capability based on the attention output, and the feedforward network module performs nonlinear transformation and stabilization processing on convolution fusion output to output a final hidden representation sequence; In the self-attention module, a learnable low-rank projection matrix is applied to the key vector and the value vector to obtain a projected key vector and a projected value vector, and the attention output is calculated by dot product of the query vector and the projected key vector; In the multi-scale convolution module, one-dimensional convolution operations with different convolution kernels are performed on the attention output, and each convolution result is spliced and fused to obtain convolution fusion output, the one-dimensional convolution operation refers to sliding weighted calculation of the attention output in the sequence length dimension with different convolution kernel sizes; In the feedforward network module, residual connection and layer normalization are performed on the attention output and the convolution fusion output, and nonlinear mapping is performed through a feedforward full connection network to obtain the hidden representation sequence; The linear mapping and the activation operation are performed on the hidden representation sequence to output a joint equalization coefficient vector, including, The linear mapping is performed on the hidden representation sequence, the linear mapping is composed of a linear mapping weight matrix and a bias vector, the hidden representation sequence is multiplied by the linear mapping weight matrix, and the bias vector is added to obtain a linear mapping result; The nonlinear activation operation is performed on the linear mapping result to obtain an activation result; The activation result is corresponded to each time-frequency unit to output the joint equalization coefficient vector.

7. The offshore microwave signal time-frequency joint equalization method based on the Transformer model according to claim 4, characterized in that: The amplitude and phase correction of the time-frequency spectrum matrix is performed by using the joint equalization coefficient vector, and de-filtering compensation is performed on the complex baseband signal sequence in the time domain to obtain an equalized complex baseband signal sequence, including, The amplitude correction of the time-frequency spectrum matrix is performed by using the joint equalization coefficient vector, and the amplitude of the complex value coefficient of the complex value coefficient sequence is multiplied by the amplitude correction coefficient in the joint equalization coefficient vector to obtain the amplitude-corrected complex value coefficient. The phase correction of the amplitude-corrected complex value coefficient is performed by using the joint equalization coefficient vector, and the phase value of the amplitude-corrected complex value coefficient is added to the phase correction coefficient in the joint equalization coefficient vector sequence to obtain the amplitude and phase corrected complex value coefficient. The amplitude and phase corrected complex value coefficient is arranged to obtain the amplitude and phase corrected time-frequency spectrum matrix. Based on the amplitude and phase corrected time-frequency spectrum matrix, de-filtering compensation is performed on the complex baseband signal sequence to obtain an equalized complex baseband signal sequence, and the de-filtering compensation refers to mapping the time-frequency domain correction information into a time domain complex gain sequence and multiplying sample by sample to construct an equivalent inverse filter and a complex baseband signal sequence for deconvolution. The inverse short-time Fourier transform and symbol decision are performed on the equalized complex baseband signal sequence to obtain the equalized time-frequency spectrum matrix and the reconstructed complex baseband signal sequence, including, The inverse short-time Fourier transform is performed on the equalized complex baseband signal sequence, and the frame shift overlap addition method is used to form a continuous complex baseband signal sequence, the inverse short-time Fourier transform refers to the operation of restoring the time domain continuous complex baseband signal sequence from the time-frequency domain complex value coefficient, and the frame shift overlap addition refers to placing the frame-by-frame time domain segment obtained by the inverse short-time Fourier transform on the time axis by frame shift after weighting by a synthesis window function, and adding sample by sample in the overlapping area, and point-to-point normalization is performed according to the window energy superposition term to obtain a continuous complex baseband signal sequence. The symbol decision is performed on the continuous complex baseband signal sequence, and the real part and the imaginary part of the sampling point in the continuous complex baseband signal sequence are respectively judged to the nearest symbol constellation point to obtain a judged complex baseband signal sequence, and the symbol constellation point refers to that the complex value of each sampling point in the continuous complex baseband signal sequence is mapped to a symbol constellation point set defined under a preset modulation mode, and each symbol constellation point corresponds to a unique bit combination. The time-frequency spectrum reconstruction is performed on the judged complex baseband signal sequence to obtain the equalized time-frequency spectrum matrix, and the time-frequency spectrum reconstruction refers to that the judged complex baseband signal sequence is framed and windowed according to the frame length, frame shift, transform point number and analysis window function consistent with the forward time-frequency analysis, and consistent analysis transformation is performed on each frame, the obtained complex value coefficient is arranged according to the time index and the frequency index, and the equalized time-frequency spectrum matrix is generated. The equalized time-frequency spectrum matrix and the reconstructed complex baseband signal sequence are output.

8. A sea-based microwave signal time-frequency joint equalization system based on a Transformer model, applying a sea-based microwave signal time-frequency joint equalization method based on a Transformer model according to any one of claims 1-7, characterized in that, It includes: a data acquisition and processing module, a time-frequency spectrum matrix generation module, an auxiliary feature vector construction module, a token sequence generation module, a model input module, an equalization coefficient generation module, a correction module and an output module. The data acquisition and processing module is configured to acquire and pre-process the microwave radio frequency signal data on the sea to obtain a complex baseband signal sequence. The time-frequency spectrum matrix generation module is configured to perform a short-time Fourier transform on the complex baseband signal sequence to generate a time-frequency spectrum matrix. The auxiliary feature vector construction module is configured to perform amplitude normalization and phase decentralization processing on the time-frequency spectrum matrix to construct an auxiliary feature vector sequence. The token sequence generation module is configured to divide the auxiliary feature vector sequence into time-frequency units, perform linear mapping on each time-frequency unit, and superimpose time position encoding and frequency position encoding to form a token sequence. The model input module is configured to input the token sequence into the improved Linformer-Transformer model to obtain a hidden representation sequence. The equalization coefficient generation module is configured to perform linear mapping and activation operations on the hidden representation sequence to output a joint equalization coefficient vector. The correction module is configured to perform amplitude and phase correction on the time-frequency spectrum matrix using the joint equalization coefficient vector, and perform de-filtering compensation on the complex baseband signal sequence in the time domain to obtain an equalized complex baseband signal sequence. The output module is configured to perform inverse short-time Fourier transform and symbol decision on the equalized complex baseband signal sequence to obtain an equalized time-frequency spectrum matrix and a reconstructed complex baseband signal sequence. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8. The processor executes the computer program to implement the steps of the sea microwave signal time-frequency joint equalization method based on the Transformer model according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the sea microwave signal time-frequency joint equalization method based on the Transformer model according to any one of claims 1-7.

Images