A multi-modal fusion radar tide level intelligent adaptive filtering method

Abstract

The application provides a multi-modal fusion radar tide intelligent adaptive filtering method, and belongs to the technical field of radar tide.The application collects dual-polarized radar signals, performs three-layer wavelet decomposition on vertical and horizontal polarization digital intermediate frequency signals, and marks clutter coefficients based on variance and skewness threshold values, performs adaptive soft threshold processing, and utilizes a convolutional neural network to evaluate residual clutter confidence and dynamically adjust suppression strength, calculates a polarization phase difference sequence and inputs a long short-term memory network to identify false target distance units, combines ray tracing calculation of a digital elevation model to calculate multipath path parameters for phase compensation and amplitude correction, and finally detects real tide height distance Doppler spectrum peak values to obtain accurate tide height values, thereby solving the technical problem that radar tide measurement in a complex marine environment cannot effectively distinguish clutter interference and multipath false targets, resulting in a decrease in measurement accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar tide, and particularly relates to a multi-modal fusion radar tide intelligent adaptive filtering method. BACKGROUND

[0002] Radar tide measurement technology measures the tide height by transmitting electromagnetic waves to the sea surface and receiving the echo signals, and has important application value in marine monitoring and port management. The traditional method uses fixed threshold filtering and single polarization signal processing to extract tide information. However, in the actual marine environment, the intensity and spectral characteristics of sea clutter and weather clutter change dynamically with the sea state, and the fixed threshold filtering method is difficult to adaptively adjust the suppression parameters, resulting in that the effective signal is excessively suppressed in the strong clutter environment, and the residual clutter affects the measurement accuracy in the weak clutter environment. At the same time, the multipath propagation effect caused by the near-shore terrain and building reflections will form false tide targets in the range-Doppler spectrum, and the traditional single polarization signal processing cannot effectively identify and eliminate these false targets, so that the peak value detection algorithm extracts the wrong distance information. That is to say, in the prior art, there is a technical problem that the radar tide measurement cannot effectively distinguish the clutter interference and multipath false targets in the complex marine environment, resulting in the decline of the measurement accuracy. SUMMARY

[0003] Therefore, the application provides a multi-modal fusion radar tide intelligent adaptive filtering method, which can solve the technical problem that the radar tide measurement cannot effectively distinguish the clutter interference and multipath false targets in the complex marine environment, resulting in the decline of the measurement accuracy.

[0004] The application is implemented in the following manner: the application provides a multi-modal fusion radar tide intelligent adaptive filtering method, a radar transmitting antenna transmits dual-polarized electromagnetic wave signals to the sea surface, a receiving antenna synchronously receives vertical polarization echo signals and horizontal polarization echo signals and performs down-conversion processing and analog-to-digital conversion to obtain vertical polarization digital intermediate frequency signals and horizontal polarization digital intermediate frequency signals, the vertical polarization digital intermediate frequency signals and the horizontal polarization digital intermediate frequency signals are respectively subjected to three-layer wavelet decomposition to obtain wavelet coefficients, the wavelet coefficients are marked and subjected to adaptive soft threshold processing, and then wavelet reconstruction is performed to obtain vertical polarization preliminary clutter suppression signals and horizontal polarization preliminary clutter suppression signals, features of the vertical polarization preliminary clutter suppression signals are extracted, a clutter recognition model is input, a clutter confidence score is output, and a clutter suppression intensity coefficient is adjusted, the vertical polarization preliminary clutter suppression signals and the horizontal polarization preliminary clutter suppression signals are respectively subjected to range direction pulse compression and Doppler frequency shift compensation to obtain vertical polarization range Doppler spectra and horizontal polarization range Doppler spectra, a polarization phase difference sequence is calculated and input into a multi-path false target recognition model to output a target authenticity probability value, digital elevation model data are acquired and a ray tracing algorithm is used to establish a multi-path parameter set, the false target range cells are phase compensated and amplitude corrected according to the multi-path parameter set to obtain a real tide range Doppler spectrum, peak value detection is performed on the real tide range Doppler spectrum, and peak value distance information is calculated, a current time tide height value is calculated according to the radar antenna altitude and the peak value distance information to form a tide time sequence.

[0005] The three-layer wavelet decomposition obtains first-layer wavelet coefficients, second-layer wavelet coefficients and third-layer wavelet coefficients, the first-layer variance and the first-layer skewness of the first-layer wavelet coefficients are calculated, the second-layer variance and the second-layer skewness of the second-layer wavelet coefficients are calculated, and the third-layer variance and the third-layer skewness of the third-layer wavelet coefficients are calculated.

[0006] The first-layer variance is compared with a first variance threshold value, when the first-layer variance is greater than the first variance threshold value and the first-layer skewness is less than a first skewness threshold value, the first-layer wavelet coefficients are marked as first-layer clutter coefficients, the second-layer variance is compared with a second variance threshold value, when the second-layer variance is greater than the second variance threshold value and the second-layer skewness is less than a second skewness threshold value, the second-layer wavelet coefficients are marked as second-layer clutter coefficients.

[0007] The three-layer wavelet decomposition uses a Daubechies wavelet base function, and the first-layer wavelet decomposition decomposes the vertical polarization digital intermediate frequency signals into first-layer low-frequency approximation coefficients and first-layer high-frequency detail coefficients.

[0008] The second-layer wavelet decomposition further decomposes the first-layer low-frequency approximation coefficients into second-layer low-frequency approximation coefficients and second-layer high-frequency detail coefficients, and the third-layer wavelet decomposition further decomposes the second-layer low-frequency approximation coefficients into third-layer low-frequency approximation coefficients and third-layer high-frequency detail coefficients.

[0009] wherein the variance threshold is determined according to 1.8 to 2.2 times of the variance of the background signal collected in a clutter-free environment, and the skewness threshold is set to 0.5.

[0010] wherein the threshold function calculation of the adaptive soft threshold processing is to divide the amplitude of each wavelet coefficient in the clutter coefficient by the root mean square value of the amplitudes of all wavelet coefficients in the clutter coefficient to obtain a normalized amplitude, and multiply the normalized amplitude by a reference threshold factor to obtain an adaptive threshold.

[0011] wherein the wavelet reconstruction adopts inverse transformation of a Daubechies wavelet basis function, and inverse transformation is performed on the third layer processed wavelet coefficient to obtain reconstructed second layer coefficients, and inverse transformation is performed on the reconstructed second layer coefficients and the second layer processed wavelet coefficients to obtain reconstructed first layer coefficients.

[0012] wherein the feature of the vertically polarized preliminary clutter suppression signal is extracted, including extracting time domain waveform features and frequency domain spectrum features, the time domain waveform features including a peak amplitude mean ratio and a rising slope, and the frequency domain spectrum features including a main peak frequency position and a spectral energy concentration degree.

[0013] wherein the clutter recognition model is a classifier based on a convolutional neural network architecture, and the input layer receives a feature vector composed of time domain waveform features and frequency domain spectrum features.

[0014] wherein the first convolutional layer of the clutter recognition model contains 32 first convolutional kernels for convolution operation on the feature vector to obtain first convolutional features, and the first convolutional features are subjected to batch normalization and rectified linear unit activation to obtain first activation features, the second convolutional layer contains 64 second convolutional kernels for convolution operation on the first activation features to obtain second convolutional features, and the second convolutional features are subjected to batch normalization and rectified linear unit activation to obtain second activation features, and the maximum pooling layer performs maximum pooling operation on the second activation features to obtain pooling features.

[0015] wherein the clutter recognition model adopts a low-precision computing framework based on neural network quantization, and 32-bit floating-point weights in the first convolutional layer, the second convolutional layer and the fully connected layer are quantized to 8-bit fixed-point number representation, and a dynamic quantization strategy is adopted for the activation features to determine the quantization range according to the 5th and 95th percentiles of the activation value distribution.

[0016] wherein the range-to-pulse compression is implemented by using a matched filter, the impulse response of the matched filter is the time reversal and conjugate of the transmitted signal waveform, and the Doppler shift compensation is achieved by multiplying the vertically polarized preliminary clutter suppression signal by a phase correction factor to eliminate the frequency offset caused by the radial velocity of the target.

[0017] The polarization phase difference sequence is calculated, the vertical polarization phase is extracted for each distance unit in the vertical polarization range-doppler spectrum, the horizontal polarization phase is extracted for each distance unit in the horizontal polarization range-doppler spectrum, the vertical polarization phase is subtracted from the horizontal polarization phase to obtain a polarization phase difference, and the polarization phase difference is arranged in sequence according to the distance to form the polarization phase difference sequence.

[0018] The structure of the multi-path false target identification model is that the input layer receives the polarization phase difference sequence, the first long short-term memory network layer extracts time sequence features from the polarization phase difference sequence to obtain first time sequence features, the discard layer performs a discard operation on the first time sequence features, and the second long short-term memory network layer extracts time sequence features from the discarded time sequence features to obtain second time sequence features.

[0019] The ray tracing algorithm is calculated, a virtual ray is emitted from the radar antenna position determined by the radar antenna longitude, the radar antenna latitude and the radar antenna altitude, the virtual ray is reflected on the surface of the digital elevation model data, the propagation path of each virtual ray from emission to reception is recorded as a multi-path propagation path, and the path delay difference and the path phase offset of the multi-path propagation path are calculated to establish a multi-path parameter set.

[0020] The application solves the defects of the traditional fixed threshold method in different sea conditions, such as insufficient clutter suppression ability, by collecting dual-polarized signals and multi-layer wavelet decomposition adaptive clutter suppression, combining a clutter identification model based on deep learning to dynamically adjust the filtering strength. By extracting the polarization phase difference sequence of the dual-polarized range-doppler spectrum and inputting the long short-term memory network for false target identification, combined with the ray tracing multi-path path compensation of the digital elevation model, the real tidal level echo and the multi-path false target are effectively distinguished, and the limitation of single-polarization processing that cannot identify the multi-path effect is overcome. The application uses wavelet domain statistical characteristics to adaptively mark clutter coefficients, extracts time-frequency domain features through a convolutional neural network to evaluate the residual clutter level, realizes closed-loop feedback adjustment of the clutter suppression strength, and uses the multi-path propagation feature difference of the dual-polarized phase difference and the physical constraints of the terrain model to accurately locate and eliminate false distance units, so that the peak value detection can correctly extract the real tidal level distance information. In summary, the application solves the technical problem of the radar tidal level measurement in the background art that the clutter interference and the multi-path false target cannot be effectively distinguished in the complex marine environment, resulting in a decrease in measurement accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is a vertical polarization digital intermediate frequency signal time domain waveform diagram.

[0022] Figure 2 It is a vertical polarization range-doppler spectrum two-dimensional distribution diagram.

[0023] Figure 3 It is a polarization phase difference sequence along the distance dimension variation curve diagram.

[0024] Figure 4 A three-dimensional spatial distribution map of multipath propagation paths.

[0025] Figure 5 A real tidal level range Doppler spectrum peak feature map.

[0026] Figure 6 A tidal level time series periodic change curve map. DETAILED DESCRIPTION

[0027] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, 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.

[0028] The present application provides a multi-modal fusion radar tidal level intelligent adaptive filtering method, comprising the following steps:

[0029] S01, the radar transmitting antenna transmits a dual-polarized electromagnetic wave signal to the sea surface, the receiving antenna synchronously receives a vertical polarization echo signal and a horizontal polarization echo signal, and the vertical polarization echo signal and the horizontal polarization echo signal are subjected to down-conversion processing and analog-digital conversion to obtain a vertical polarization digital intermediate frequency signal and a horizontal polarization digital intermediate frequency signal;

[0030] S02, the vertical polarization digital intermediate frequency signal and the horizontal polarization digital intermediate frequency signal are respectively subjected to three-layer wavelet decomposition to obtain first-layer wavelet coefficients, second-layer wavelet coefficients and third-layer wavelet coefficients, the first-layer variance and the first-layer skewness of the first-layer wavelet coefficients are calculated, the second-layer variance and the second-layer skewness of the second-layer wavelet coefficients are calculated, and the third-layer variance and the third-layer skewness of the third-layer wavelet coefficients are calculated;

[0031] S03, the first-layer variance is compared with a first variance threshold value, when the first-layer variance is greater than the first variance threshold value and the first-layer skewness is less than a first skewness threshold value, the first-layer wavelet coefficients are marked as first-layer clutter coefficients, the second-layer variance is compared with a second variance threshold value, when the second-layer variance is greater than the second variance threshold value and the second-layer skewness is less than a second skewness threshold value, the second-layer wavelet coefficients are marked as second-layer clutter coefficients, and the third-layer variance is compared with a third variance threshold value, when the third-layer variance is greater than the third variance threshold value and the third-layer skewness is less than a third skewness threshold value, the third-layer wavelet coefficients are marked as third-layer clutter coefficients;

[0032] S04, performing adaptive soft threshold processing on the first layer clutter coefficient, the second layer clutter coefficient and the third layer clutter coefficient to obtain a first layer processed wavelet coefficient, a second layer processed wavelet coefficient and a third layer processed wavelet coefficient, and performing wavelet reconstruction on the first layer processed wavelet coefficient, the second layer processed wavelet coefficient and the third layer processed wavelet coefficient to obtain a vertical polarization preliminary clutter suppression signal and a horizontal polarization preliminary clutter suppression signal;

[0033] S05, extracting time domain waveform features and frequency domain spectrum features of the vertical polarization preliminary clutter suppression signal, inputting the time domain waveform features and the frequency domain spectrum features into a pre-trained clutter recognition model, outputting a clutter confidence score, increasing a clutter suppression intensity coefficient by an adjustment amplitude when the clutter confidence score is greater than a confidence threshold, keeping the clutter suppression intensity coefficient unchanged when the clutter confidence score is less than the confidence threshold, and re-executing the adaptive soft threshold processing according to the adjusted clutter suppression intensity coefficient;

[0034] S06, performing range pulse compression and Doppler frequency shift compensation on the vertical polarization preliminary clutter suppression signal and the horizontal polarization preliminary clutter suppression signal respectively to obtain a vertical polarization range Doppler spectrum and a horizontal polarization range Doppler spectrum, and calculating a polarization phase difference sequence between the vertical polarization range Doppler spectrum and the horizontal polarization range Doppler spectrum;

[0035] S07, inputting the polarization phase difference sequence into a pre-trained multi-path false target recognition model to output a target authenticity probability value of each range unit, and marking a range unit with a target authenticity probability value less than an authenticity threshold as a false target range unit;

[0036] S08, obtaining digital elevation model data of a measurement area, calculating multi-path propagation paths by using a ray tracing algorithm according to a radar antenna longitude, a radar antenna latitude, a radar antenna altitude and the digital elevation model data, calculating a path delay difference and a path phase offset for each multi-path propagation path, and establishing a multi-path parameter set;

[0037] S09, compensating the phase and correcting the amplitude of the vertical polarization range Doppler spectrum and the horizontal polarization range Doppler spectrum corresponding to the false target range unit according to the path delay difference and the path phase offset in the multi-path parameter set, eliminating false tidal level echoes in the false target range unit, and obtaining a real tidal level range Doppler spectrum;

[0038] S10, performing peak value detection on the real tidal level range Doppler spectrum, extracting a peak value position of the real tidal level range Doppler spectrum, calculating peak distance information according to the peak value position, calculating a current time tidal level height value according to the radar antenna altitude and the peak distance information, and recording the current time tidal level height value at multiple time points to form a tidal level time sequence.

[0039] The vertical polarization component and the horizontal polarization component of the dual-polarized electromagnetic wave signal have the same frequency and a phase difference of 90 degrees. The down-conversion process converts the radio frequency signal into an intermediate frequency signal, and the analog-to-digital conversion converts the analog intermediate frequency signal into a discrete digital signal.

[0040] The three-layer wavelet decomposition employs the Daubechies wavelet basis function. The first-layer wavelet decomposition decomposes the vertically polarized digital intermediate frequency signal into first-layer low-frequency approximation coefficients and first-layer high-frequency detail coefficients. The first-layer wavelet coefficients include first-layer low-frequency approximation coefficients and first-layer high-frequency detail coefficients. The second-layer wavelet decomposition further decomposes the first-layer low-frequency approximation coefficients into second-layer low-frequency approximation coefficients and second-layer high-frequency detail coefficients. The second-layer wavelet coefficients include second-layer low-frequency approximation coefficients and second-layer high-frequency detail coefficients. The third-layer wavelet decomposition further decomposes the second-layer low-frequency approximation coefficients into third-layer low-frequency approximation coefficients and third-layer high-frequency detail coefficients. The third-layer wavelet coefficients include third-layer low-frequency approximation coefficients and third-layer high-frequency detail coefficients.

[0041] Wherein, the first variance threshold is determined based on 1.8 times the variance of the background signal collected in a clutter-free environment, the second variance threshold is determined based on 2.0 times the variance of the background signal collected in a clutter-free environment, the third variance threshold is determined based on 2.2 times the variance of the background signal collected in a clutter-free environment, the first skewness threshold is set to 0.5, the second skewness threshold is set to 0.5, and the third skewness threshold is set to 0.5.

[0042] The threshold function calculation steps for the adaptive soft thresholding process are as follows: the amplitude of each wavelet coefficient in the first layer clutter coefficient is divided by the root mean square value of the amplitudes of all wavelet coefficients in the first layer clutter coefficient to obtain the normalized amplitude of the first layer; the normalized amplitude of the first layer is multiplied by the benchmark threshold factor to obtain the adaptive threshold of the first layer; when the amplitude of the wavelet coefficient in the first layer clutter coefficient is less than the adaptive threshold of the first layer, the wavelet coefficient is set to zero; when the amplitude of the wavelet coefficient in the first layer clutter coefficient is greater than the adaptive threshold of the first layer, the adaptive threshold of the first layer is subtracted from the wavelet coefficient to obtain the processed wavelet coefficient of the first layer. The benchmark threshold factor is set to 1.0.

[0043] Specifically, the wavelet reconstruction employs the inverse transform of the Daubechies wavelet basis functions. The wavelet coefficients after the third layer of processing are subjected to an inverse transform to obtain the reconstructed second-layer coefficients. The reconstructed second-layer coefficients and the wavelet coefficients after the second layer of processing are combined and subjected to an inverse transform to obtain the reconstructed first-layer coefficients. The reconstructed first-layer coefficients and the wavelet coefficients after the first layer of processing are combined and subjected to an inverse transform to obtain the vertically polarized preliminary suppressed clutter signal.

[0044] The time-domain waveform feature includes a peak amplitude mean value ratio and a rising edge slope, the peak amplitude mean value ratio is a peak amplitude of the vertical polarization preliminary suppression of spurious signals divided by a mean amplitude of the vertical polarization preliminary suppression of spurious signals, and the rising edge slope is a change amount of amplitude of a rising edge section of the vertical polarization preliminary suppression of spurious signals divided by a change amount of time.

[0045] The spurious signal recognition model is a classifier based on a convolutional neural network architecture, the confidence threshold is set to 0.7, and the adjustment amplitude is 20% of a current value of the spurious signal suppression intensity coefficient.

[0046] The specific structure of the spurious signal recognition model is that an input layer receives a feature vector composed of time-domain waveform features and frequency-domain spectrum features, the feature vector includes four elements of a peak amplitude mean value ratio, a rising edge slope, a main peak frequency position and a spectrum energy concentration degree, a first convolutional layer includes 32 first convolutional kernels with a size of 3, the first convolutional layer performs convolution operation on the feature vector to obtain first convolutional features, a first batch normalization layer performs batch normalization processing on the first convolutional features to obtain first normalized features, a first activation function layer performs rectified linear unit activation on the first normalized features to obtain first activated features, a second convolutional layer includes 64 second convolutional kernels with a size of 3, the second convolutional layer performs convolution operation on the first activated features to obtain second convolutional features, a second batch normalization layer performs batch normalization processing on the second convolutional features to obtain second normalized features, a second activation function layer performs rectified linear unit activation on the second normalized features to obtain second activated features, a maximum pooling layer performs maximum pooling operation on the second activated features to obtain pooled features, a fully connected layer includes 128 neurons, the fully connected layer performs fully connected operation on the pooled features to obtain fully connected features, and an output layer performs Sigmoid activation function operation on the fully connected features to obtain a spurious signal confidence score.

[0047] The clutter recognition model adopts a low-precision calculation framework based on neural network quantization, quantizes 32-bit floating-point weights in the first convolutional layer, the second convolutional layer and the fully connected layer to 8-bit fixed-point number representation, quantizes the first activation feature and the second activation feature and the pooling feature and the fully connected feature to 8-bit fixed-point number representation, and the quantization process is as follows: the maximum absolute value of the first convolutional layer weight matrix of the first convolutional layer is calculated as the first convolutional layer scaling factor, each weight element in the first convolutional layer weight matrix is divided by the first convolutional layer scaling factor and multiplied by 127, and then rounded to an integer to obtain the first convolutional layer quantization weight, the integer range of the first convolutional layer quantization weight is limited to -128 to 127, the maximum absolute value of the second convolutional layer weight matrix of the second convolutional layer is calculated as the second convolutional layer scaling factor, each weight element in the second convolutional layer weight matrix is divided by the second convolutional layer scaling factor and multiplied by 127, and then rounded to an integer to obtain the second convolutional layer quantization weight, the integer range of the second convolutional layer quantization weight is limited to -128 to 127, the maximum absolute value of the fully connected layer weight matrix of the fully connected layer is calculated as the fully connected layer scaling factor, each weight element in the fully connected layer weight matrix is divided by the fully connected layer scaling factor and multiplied by 127, and then rounded to an integer to obtain the fully connected layer quantization weight, and the integer range of the fully connected layer quantization weight is limited to -128 to 127.

[0048] The quantization of the first activation feature adopts a dynamic quantization strategy, the 5th percentile and the 95th percentile of the activation value distribution of the first activation feature are counted, the first activation feature quantization range is determined according to the 5th percentile and the 95th percentile, the maximum absolute value of the first activation feature quantization range is taken as the first activation feature scaling factor, each activation value in the first activation feature is divided by the first activation feature scaling factor and multiplied by 127, and then rounded to an integer to obtain the first activation feature quantization value, the integer range of the first activation feature quantization value is limited to -128 to 127, and the quantization methods of the second activation feature, the pooling feature and the fully connected feature are the same as those of the first activation feature.

[0049] The clutter recognition model adopts a quantization-aware training technique in the training stage, performs quantization operations on the first convolutional layer weight matrix, the second convolutional layer weight matrix and the fully connected layer weight matrix to obtain the first convolutional layer quantization weight, the second convolutional layer quantization weight and the fully connected layer quantization weight during forward propagation, and uses the first convolutional layer quantization weight, the second convolutional layer quantization weight and the fully connected layer quantization weight for forward calculation, adopts a pass-through estimator method during back propagation, regards the quantization function as an identity mapping during back propagation, and directly transmits the gradient of the quantization operation to the previous layer, and the training process simultaneously updates the floating-point weight and the quantization parameter, and only uses the quantized integer weight for calculation in the inference stage.

[0050] The low-precision computing framework based on neural network quantization reduces the storage space requirement of the clutter recognition model to one fourth of the original floating-point model, significantly reducing the memory occupation of the model in the embedded radar processor. The 8-bit integer operation after quantization has higher hardware execution efficiency than the 32-bit floating-point operation. The throughput of the arithmetic logic unit of the processor when performing integer multiplication and addition operations is several times that of floating-point operations, directly improving the calculation speed of the first and second convolution layers and the fully connected layer. The dynamic quantization strategy adaptively adjusts the quantization parameters according to the actual sea conditions, so that the clutter recognition model can maintain stable recognition accuracy under different clutter intensity and spectral distribution conditions, avoiding the performance degradation problem of the fixed quantization scheme in extreme sea conditions. The quantization-aware training technology simulates the influence of quantization error during the training stage, so that the first convolution layer weight matrix, the second convolution layer weight matrix and the fully connected layer weight matrix actively adapt to low-precision representation during the optimization process. The accuracy loss of the quantized model after training is controlled within an acceptable range. The pass-through estimator method solves the difficulty of back propagation caused by the non-differentiable quantization operation, enabling end-to-end training of neural networks containing quantization layers, and the entire training process does not require complex post-quantization calibration steps. The low-precision computing framework reduces the bandwidth requirement of data storage and transmission, and the power consumption of integer operation is lower than that of floating-point operation. Under the same power budget, it can support higher sampling rate and longer signal processing time, making the millisecond-level tidal level measurement response a reality. The entire filtering system maintains high accuracy while achieving a significant improvement in real-time processing capability.

[0051] The training data set of the clutter recognition model is established by the following steps: collecting radar echo data containing sea clutter and weather clutter under different sea conditions, synchronously recording wind speed data and rainfall intensity data measured by a weather station, manually labeling the collected radar echo data, marking the echo data segment containing obvious clutter characteristics as a clutter sample, marking the echo data segment containing the real tidal level echo as a signal sample, extracting the time-domain waveform features and frequency-domain spectral features from the clutter sample to obtain a clutter feature vector, extracting the time-domain waveform features and frequency-domain spectral features from the signal sample to obtain a signal feature vector, labeling the clutter feature vector as a class label 1, labeling the signal feature vector as a class label 0, and combining the clutter feature vector and the signal feature vector and the corresponding class label to form a training sample pair, and dividing the training sample pair according to the proportion of 70% training set and 30% validation set.

[0052] The specific steps of training the clutter recognition model include: initializing the weights of the first convolutional kernel of the first convolutional layer, the weights of the second convolutional kernel of the second convolutional layer, and the weights of the fully connected layer as random values ​​following a normal distribution with zero mean and a standard deviation of 0.01; inputting training sample pairs from the training set into the clutter recognition model in batches of 32; calculating the binary cross-entropy loss between the clutter confidence score output by the clutter recognition model and the category label; using the Adam optimization algorithm to update the weights of the first and second convolutional kernels and the fully connected layer based on the binary cross-entropy loss; setting the learning rate to 0.001 and the number of training epochs to 100; evaluating the performance of the clutter recognition model using a validation set after each training epoch; stopping training when the binary cross-entropy loss on the validation set no longer decreases for 10 consecutive training epochs; and saving the first and second convolutional kernel weights and the fully connected layer weights with the best performance on the validation set as the final model weights.

[0053] The range-direction pulse compression is achieved using a matched filter, the impulse response of which is the time reversal and conjugate of the transmitted signal waveform. The Doppler frequency shift compensation eliminates the frequency shift caused by the target radial velocity by multiplying the vertically polarized initial suppressed clutter signal by a phase correction factor. The vertically polarized range-Doppler spectrum is a two-dimensional matrix, where the row direction of the two-dimensional matrix represents the range dimension, the column direction represents the Doppler frequency dimension, and the element values ​​of the two-dimensional matrix represent the echo energy at the corresponding range and Doppler frequency.

[0054] The calculation steps of the polarization phase difference sequence are as follows: extract the vertical polarization phase for each distance unit in the vertical polarization range-Doppler spectrum, extract the horizontal polarization phase for each distance unit in the horizontal polarization range-Doppler spectrum, subtract the horizontal polarization phase from the vertical polarization phase to obtain the polarization phase difference, and arrange the polarization phase differences of all distance units in the distance order to form a polarization phase difference sequence.

[0055] The specific structure of the multi-path false target identification model is that: the input layer receives a polarization phase difference sequence, the first long short-term memory network layer includes a first number of hidden units, the first number of hidden units is determined according to a radar operating frequency and a radar antenna altitude, the first long short-term memory network layer performs time sequence feature extraction on the polarization phase difference sequence to obtain first time sequence features, a dropout layer performs a dropout operation on the first time sequence features to obtain dropped time sequence features, a dropout rate of the dropout layer is set to 0.3, the second long short-term memory network layer includes 32 hidden units, the second long short-term memory network layer performs time sequence feature extraction on the dropped time sequence features to obtain second time sequence features, the multi-path fully connected layer includes 16 neurons, the multi-path fully connected layer performs fully connected operation on the second time sequence features to obtain multi-path fully connected features, and the multi-path output layer performs Sigmoid activation function operation on the multi-path fully connected features to obtain a target authenticity probability value.

[0056] The calculation steps of the first number of hidden units are: dividing the radar operating frequency by a standard frequency of 10GHz to obtain a frequency normalization coefficient, dividing the radar antenna altitude by a standard height of 50m to obtain a height normalization coefficient, multiplying the frequency normalization coefficient and the height normalization coefficient, and then multiplying the product by a reference number of hidden units of 64 to obtain a first number of hidden units initial value, and rounding the first number of hidden units initial value to an integer to obtain the first number of hidden units.

[0057] The multi-path false target identification model adopts a low-precision calculation framework based on neural network quantization, quantizes 32-bit floating-point weights in the first long short-term memory network layer, the second long short-term memory network layer and the multi-path fully connected layer into 8-bit fixed-point number representations, the quantization method is the same as that of the clutter identification model, the gating mechanism in the first long short-term memory network layer includes an input gate, a forget gate and an output gate, the activation function calculation of the input gate, the forget gate and the output gate adopts quantized integer operation, the update of the hidden state and the cell state of the first long short-term memory network layer adopts quantized integer addition and integer multiplication operation, and the multi-path false target identification model adopts a quantization-aware training technique in the training stage, performs quantization operation in forward propagation, and uses a pass-through estimator to approximate gradient in backward propagation.

[0058] The step of establishing the training data set of the multi-path false target identification model specifically comprises: collecting radar dual-polarization echo data in a measurement environment known to have a multi-path effect, synchronously measuring a true tide value as a reference benchmark using a high-precision tide gauge, processing the radar dual-polarization echo data to obtain a vertical polarization range-Doppler spectrum and a horizontal polarization range-Doppler spectrum, calculating a polarization phase difference sequence between the vertical polarization range-Doppler spectrum and the horizontal polarization range-Doppler spectrum, determining which distance units in the radar dual-polarization echo data correspond to a true target according to the true tide value measured by the high-precision tide gauge, marking the distance units corresponding to the true target as positive sample distance units, marking the distance units deviating from the true tide value as negative sample distance units, extracting a polarization phase difference sequence segment corresponding to the positive sample distance units as a positive sample feature, extracting a polarization phase difference sequence segment corresponding to the negative sample distance units as a negative sample feature, marking the positive sample feature as a target label 1, marking the negative sample feature as a target label 0, and grouping the positive sample feature and the negative sample feature and the corresponding target label into a multi-path training sample pair, and dividing the multi-path training sample pair according to a proportion of 80% training set and 20% validation set.

[0059] The step of training the multi-path false target identification model specifically comprises: initializing weight matrices of a first long short-term memory network layer and a second long short-term memory network layer as random values subject to a Xavier initialization distribution, inputting the multi-path training sample pairs in the training set into the multi-path false target identification model in a batch size of 64, calculating a binary cross-entropy loss between target authenticity probability values output by the multi-path false target identification model and target labels, updating the weight matrices of the first long short-term memory network layer and the second long short-term memory network layer according to the binary cross-entropy loss using an RMSprop optimization algorithm, setting a learning rate to 0.0005, setting a number of training periods to 150, evaluating performance of the multi-path false target identification model on the validation set every 5 training periods, stopping training when accuracy on the validation set no longer improves for 15 consecutive training periods, and saving the weight matrices with optimal performance on the validation set as final model weights.

[0060] The digital elevation model data contains three-dimensional spatial coordinate information of terrains and buildings in a measurement region, the resolution of the digital elevation model data is 3 m, and the calculation step of the ray tracing algorithm is: a virtual ray is emitted from a radar antenna position determined by radar antenna longitude, radar antenna latitude and radar antenna altitude, the virtual ray is reflected on a surface of the digital elevation model data, and a propagation path of each virtual ray from emission to reception is recorded as a multi-path propagation path.

[0061] The path delay difference calculation step is: calculating the path length of the multipath propagation path, calculating the direct path length from the radar antenna position to the sea surface, subtracting the direct path length from the path length of the multipath propagation path to obtain the path length difference, and dividing the path length difference by the electromagnetic wave propagation speed to obtain the path delay difference.

[0062] The multipath path parameter set includes the path delay difference and the path phase offset corresponding to all multipath propagation paths.

[0063] The phase compensation step is: multiplying the vertical polarization range Doppler spectrum corresponding to the false target range cell by a phase correction factor, the phase of the phase correction factor being the negative value of the path phase offset, and the amplitude correction step is: calculating the propagation loss according to the path length of the multipath propagation path, the propagation loss being proportional to the square of the path length, and adjusting the amplitudes of the vertical polarization range Doppler spectrum and the horizontal polarization range Doppler spectrum corresponding to the false target range cell according to the propagation loss.

[0064] The real tidal level range Doppler spectrum is the range Doppler spectrum after the false tidal level echo is eliminated, and the real tidal level range Doppler spectrum presents a single peak value feature in the range dimension.

[0065] The peak detection identifies sample points with amplitudes greater than the amplitude threshold in the real tidal level range Doppler spectrum by setting the amplitude threshold, the peak position is the index of the peak sample point in the range dimension, and the peak distance information is obtained by multiplying the peak position by the range resolution, which is equal to the electromagnetic wave propagation speed divided by the radar signal bandwidth and then divided by 2.

[0066] The calculation step of the current time tidal level height value is: dividing the radar antenna altitude by the standard height 100m to obtain the height normalization value, dividing the peak distance information by the standard distance 200m to obtain the distance normalization value, subtracting the distance normalization value from the height normalization value and then multiplying the standard height 100m to obtain the current time tidal level height value.

[0067] The tidal level time sequence is a sequence of current time tidal level height values arranged in time sequence, and the time interval of the tidal level time sequence is 5s; the authenticity threshold is set to 0.6.

[0068] Optionally, the present application also provides a computer-implemented way to form a multi-modal fusion radar tidal level intelligent adaptive filtering system, the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions can execute the above-mentioned method when running in the computer.

[0069] The specific implementation of the above steps is described in detail below. It should be noted that the present application also solves the technical problem that the real-time performance of a deep learning model is insufficient when deployed on an embedded radar processor due to limited computing resources and storage space. The present application uses a low-precision computing framework based on neural network quantization to quantize the 32-bit floating-point weights of the convolutional layer and the long short-term memory network layer to 8-bit fixed-point representation. By calculating the maximum absolute value of the weight matrix as a scaling factor and normalizing the weights to the integer range of negative 128 to positive 127, the model storage requirement is significantly reduced to one-fourth of the original floating-point model. At the same time, a dynamic quantization strategy is used for the activation features, and the quantization range is adaptively determined according to the 5th and 95th percentiles of the activation value distribution, so that the model can adapt to the dynamic changes of signal amplitude in different sea conditions while maintaining stable recognition accuracy. In the training stage, a quantization-aware training technique is introduced. During forward propagation, the quantization operation is performed to make the weight matrix actively adapt to low-precision representation. During back propagation, the straight-through estimator method is used to directly pass the gradient of the quantization function to the previous layer, solving the training difficulty caused by the non-differentiable quantization operation, and making the accuracy loss of the quantized model controlled within an acceptable range. The hardware execution efficiency of 8-bit integer operations is significantly higher than that of 32-bit floating-point operations, and the throughput of the arithmetic logic unit of the processor when performing integer multiplication and addition operations is increased by several times. In addition, the reduced data transmission bandwidth and lower power consumption make the millisecond-level tidal level measurement response a reality. The entire filtering system maintains high accuracy while significantly improving real-time processing capability, meeting the strict requirements of embedded radar systems for computing efficiency and power consumption.

[0070] Specifically, the principle of the present application is that the technical solution of the present application can solve the root cause of the above technical problems due to the synergistic effect of multi-modal information fusion and intelligent adaptive processing mechanism. Wavelet decomposition decomposes the signal into different frequency subbands, and the variance and skewness of the wavelet coefficients at each layer are significantly different. By setting a multi-layer adaptive threshold, the clutter component can be accurately located and suppressed while retaining the effective signal, overcoming the problem of improper processing of different frequency components by a fixed global threshold. Convolutional neural networks can learn the complex time-frequency domain patterns of clutter through multi-layer feature extraction, and the non-linear expression ability of the model can be enhanced by combining batch normalization and activation functions. The output confidence score reflects the true level of residual clutter, and the soft threshold processing strength is dynamically adjusted to form a closed-loop feedback, enabling the clutter suppression process to have intelligent decision-making capabilities. The phase change rule of dual-polarized signals in multipath propagation is essentially different from that of direct waves, and the gating mechanism of long short-term memory networks can capture the time-dependent relationship of the polarization phase difference sequence, accurately identify the false target range cell corresponding to the abnormal phase pattern, and the terrain constraints provided by the digital elevation model enable the delay and phase shift of the multipath path to be accurately calculated through ray tracing. Phase compensation and amplitude correction can eliminate multipath interference from a physical level, and the complementarity of multi-modal information and the adaptability of deep learning models ensure the robustness of the algorithm in complex environments, so that the final extracted real tide distance Doppler spectrum has a single peak value feature, and the peak detection result accurately reflects the actual tide height.

[0071] A specific embodiment 1 of the present application is provided below, and the specific implementation of steps S01 and S06 in embodiment 1 is the same as described above, and will not be described in detail here. The specific implementation of other steps is described in detail as follows.

[0072] The specific implementation of step S02 is that the Daubechies wavelet basis function is used to perform three-layer wavelet decomposition on the vertical polarization digital intermediate frequency signal and the horizontal polarization digital intermediate frequency signal respectively. The first layer wavelet decomposition decomposes the vertical polarization digital intermediate frequency signal into a first layer low frequency approximation coefficient and a first layer high frequency detail coefficient , the second layer wavelet decomposition continues to decompose the first layer low frequency approximation coefficient into a second layer low frequency approximation coefficient and a second layer high frequency detail coefficient , and the third layer wavelet decomposition continues to decompose the second layer low frequency approximation coefficient into a third layer low frequency approximation coefficient and a third layer high frequency detail coefficient . The calculation formula of the first layer variance is:

[0073] ;

[0074] wherein, is the th coefficient in the first layer wavelet coefficients; is the total number of the first layer wavelet coefficients; is the mean value of the first layer wavelet coefficients, calculated by The calculation formula of the first layer skewness is:

[0075]

[0076] wherein, is a dimensionless parameter, reflecting the symmetry characteristics of the first layer wavelet coefficient distribution. The calculation method of the second layer variance and the second layer skewness is the same as the first layer, replacing with the second layer wavelet coefficient , with , wherein is the total number of the second layer wavelet coefficients, is the index of the second layer wavelet coefficients. The calculation method of the third layer variance and the third layer skewness is the same as the first layer, replacing with the third layer wavelet coefficient , with , wherein is the total number of the third layer wavelet coefficients, is the index of the third layer wavelet coefficients.

[0077] The specific implementation of step S03 is that the first variance threshold is determined according to the background signal variance collected in a clutter-free environment, and the calculation formula is:

[0078]

[0079] wherein, is a standard variance reference value, and the empirical value is 1.0, which is used for dimensionless processing. The calculation formula of the second variance threshold is:

[0080]

[0081] The calculation formula of the third variance threshold is:

[0082]

[0083] When and​​​​​ When and , the second layer wavelet coefficients are marked as second layer clutter coefficients. When and , the third layer wavelet coefficients are marked as third layer clutter coefficients.

[0084] The specific implementation of step S04 is that when the adaptive soft threshold processing is performed on the first layer clutter coefficients, the first layer normalized amplitude is first calculated, and the formula is:

[0085] ;

[0086] In the formula, is the amplitude of the kth wavelet coefficient in the first layer clutter coefficients; the denominator is the root mean square value of the amplitudes of all wavelet coefficients in the first layer clutter coefficients; is a summation index, and the value range is 1 to . The calculation formula of the first layer adaptive threshold is:

[0087] ;

[0088] In the formula, is a reference threshold factor, and the value is usually 1.0; is a normalization factor, and the default value is 1.0, which is used to ensure is a dimensionless parameter. When , the wavelet coefficient is set to zero, that is, ; when , the first layer adaptive threshold is subtracted from the wavelet coefficient to obtain the first layer processed wavelet coefficient , and the formula is:

[0089] ;

[0090] In the formula, is a sign function, which takes a value of 1 when , takes a value of -1 when , and takes a value of 0 when , which is used to retain the positive and negative signs of the original coefficients. The adaptive soft threshold processing of the second layer and the third layer is the same as that of the first layer, and the second layer processed wavelet coefficient and the third layer processed wavelet coefficient are obtained. The wavelet reconstruction adopts the inverse transform of the Daubechies wavelet basis function, and the specific implementation is the same as the foregoing, which will not be described in detail here.

[0091] ​The specific implementation of step S05 is: peak amplitude mean ratio The calculation formula is:

[0092] ;

[0093] In the formula, is the peak amplitude of the vertically polarized preliminary suppression spurs signal; is the amplitude of the vertically polarized preliminary suppression spurs signal at the i-th sampling point; is the total number of sampling points. The calculation formula of the rising slope is:

[0094] ;

[0095] In the formula, is the amplitude change of the rising edge section; is the amplitude normalization factor, and the empirical value is 1 volt; is the time change of the rising edge section; is the time normalization factor, and the empirical value is 1 second. The main peak frequency position is the frequency value corresponding to the amplitude maximum peak in the spectrum of the vertically polarized preliminary suppression spurs signal. The calculation formula of the spectral energy concentration is:

[0096] ;

[0097] In the formula, are the energies of the first three peaks in the spectrum, respectively; is the energy of the i-th frequency point in the spectrum, calculated as the square of the amplitude of the frequency point; is the total number of frequency points in the spectrum. The time domain waveform feature and the frequency domain spectrum feature form a feature vector , which is represented as:

[0098] ; In the formula,

[0099] is a four-dimensional column vector containing four elements of peak amplitude mean ratio, rising slope, main peak frequency position and spectral energy concentration; the superscript represents the transpose operation. When the spurs confidence score is, the spurs suppression intensity coefficient is increased according to the following formula:

[0100] ;

[0101] In the formula, is the adjusted spurs suppression intensity coefficient;​​ is the current clutter suppression intensity coefficient; is a normalization coefficient, which is 1.0 by default, used to ensure the dimensionless of the adjustment amplitude. When , the clutter suppression intensity coefficient remains unchanged.

[0102] The specific implementation of step S07 is: the calculation formula of the polarization phase difference sequence is as follows:

[0103] ;

[0104] In the formula, is the vertical polarization phase of the i-th distance unit; is the horizontal polarization phase of the i-th distance unit; is the distance unit index. The calculation formula of the number of first hidden units in the multipath false target identification model is as follows:

[0105] ;

[0106] In the formula, is the radar operating frequency, in hertz; is the standard frequency, which is 300000000 hertz; is the radar antenna altitude, in meters; is the standard height, which is 50 meters; and 64 is the reference number of hidden units. represents the rounding function. The polarization phase difference sequence is input into the multipath false target identification model, and the target authenticity probability value of each distance unit is output . When , the distance unit is marked as a false target distance unit.

[0107] The specific implementation of step S08 is: the ray tracing algorithm transmits a virtual ray from the radar antenna position , and calculates the reflection path of the virtual ray on the digital elevation model data surface, wherein is the radar antenna longitude, is the radar antenna latitude, is the radar antenna altitude. The calculation formula of the path delay difference is as follows:

[0108] ;

[0109] In the formula, is the path length of the multipath propagation path, in meters; ​​​​is the direct path length from the radar antenna position to the sea surface, in meters; is the path normalization factor, with an empirical value of 100 meters; is the electromagnetic wave propagation speed, with a value of meters per second; is the speed normalization factor, with a value of meters per second. The path phase offset is calculated as follows:

[0110] ;

[0111] wherein is the radar operating angular frequency, in radians per second; is the angular frequency normalization factor, with an empirical value of radians per second.

[0112] The specific implementation of step S09 is that the phase compensation is implemented by multiplying the vertical polarization range-Doppler spectrum corresponding to the false target distance unit by a phase correction factor , wherein is the imaginary unit, satisfying . The amplitude correction is performed according to the propagation loss , and the propagation loss calculation formula is as follows:

[0113] ;

[0114] wherein is the path length of the multipath propagation path, in meters; is the standard path length, with an empirical value of 100 meters. The amplitude of the corrected vertical polarization range-Doppler spectrum is calculated as follows:

[0115] ;

[0116] wherein is the amplitude of the vertical polarization range-Doppler spectrum corresponding to the false target distance unit before correction.

[0117] The specific implementation of step S10 is that the peak distance information is obtained by multiplying the peak position by the distance resolution , and the distance resolution calculation formula is as follows:

[0118] ;

[0119] wherein is the electromagnetic wave propagation speed, with a value of meters per second; is the speed normalization factor, with a value of meters per second; is the radar signal bandwidth, in hertz; is the bandwidth normalization factor, the empirical value is hertz. The current time tidal height value is calculated by the formula:

[0120] ;

[0121] In the formula, is the radar antenna altitude, in meters; is the standard height, taking the value of 100 meters; is the peak distance information, in meters; is the standard distance, taking the value of 200 meters; the calculation result is in meters. The tidal level time series is recorded continuously at a time interval of 5 seconds, wherein represents the tidal height value at the th time, is the total number of recorded time points.

[0122] In the above equation, the first layer normalized amplitude is calculated by dividing the amplitude of each wavelet coefficient by the root mean square value to achieve dimensionless processing, so that the wavelet coefficients of different scales have comparability, and the normalized amplitude range is usually between 0 and 3, which is convenient for subsequent threshold judgment. The adaptive soft threshold formula adopts a soft threshold shrinkage strategy, which retains the phase information of the coefficient through the sign function , and simultaneously realizes amplitude attenuation by subtracting the adaptive threshold, which avoids the oscillation phenomenon caused by discontinuity compared with the hard threshold method, and the soft threshold function is expressed as:

[0123] ;

[0124] The function has continuity near the threshold, which guarantees the smoothness of the reconstructed signal, effectively suppresses the clutter component and retains the main features of the tidal level echo. The peak amplitude mean ratio is dimensionless by ratio form, reflecting the kurtosis characteristics of the signal, and a high value indicates that the signal has obvious transient characteristics, and the tidal level echo usually presents a sharp peak with a value greater than 3, while the value of the clutter is usually less than 2. The spectral energy concentration degree also adopts a ratio form, with a numerical range of 0 to 1, and a larger value indicates that the signal energy is more concentrated in the main frequency component, and the value of the real tidal level echo is usually greater than 0.6, and the clutter usually presents energy dispersion characteristics and The value is less than 0.4. Eigenvector By unifying time-domain and frequency-domain features into points in four-dimensional space, a complete signal representation is provided for the clutter identification model. The four feature components complement each other, characterizing the time-frequency properties of the signal from different perspectives. (Clutter suppression intensity coefficient) The adaptive adjustment mechanism dynamically increases the suppression strength based on the clutter confidence score. When residual clutter is detected, the filtering strength is automatically increased, with the adjustment range set at 20% of the current coefficient. This ensures both flexibility in adjustment and avoids signal distortion caused by over-filtering. (Number of first hidden units) The computation employs a dual normalization strategy for frequency and altitude, adaptively adjusting the network capacity to match the multipath complexity under different radar configurations. High-frequency radars and tall antennas generate more complex multipath structures, requiring more hidden elements to capture temporal dependencies. The baseline hidden element count of 64 has been verified through extensive experiments to achieve a good balance between computational efficiency and recognition accuracy. Path delay difference The calculation achieves dimensionlessness through double normalization. The numerator is calculated by dividing the path length difference by a path normalization factor, and the denominator is calculated by dividing the electromagnetic wave propagation speed by a speed normalization factor. Dividing these two dimensionless quantities yields a dimensionless delay ratio, which directly reflects the time lag of the multipath echo relative to the direct echo. Path phase offset. By employing angular frequency normalization, the physical phase is converted into standardized phase units, eliminating the influence of the radar's operating frequency on phase compensation accuracy and making the phase compensation algorithm universal across radars of different frequency bands. Propagation loss Using the path length squared ratio conforms to the physical law that free-space propagation loss is proportional to the square of the distance, and normalization to the standard path length ensures the dimensionless nature of the loss coefficient. Distance resolution. The calculation achieves dimensionlessness through dual normalization of velocity and bandwidth. The numerator is the normalized velocity, and the denominator is twice the normalized bandwidth. Dividing the two yields the standardized distance unit scale. Current tide height value. The calculation formula employs a piecewise normalization strategy. The antenna height is divided by the standard height of 100 meters to obtain the height normalization value, and the peak distance information is divided by the standard distance of 200 meters to obtain the distance normalization value. The two dimensionless values ​​are subtracted and then multiplied by the standard height of 100 meters to restore the actual height. This method effectively compensates for systematic errors under different measurement environments, improves the absolute accuracy and relative stability of tide level measurement, and ensures that the radar tide level measurement results are consistent with the tide gauge reference. The entire filtering system, through multi-level dimensionless processing and adaptive parameter adjustment, effectively suppresses clutter and multipath effects in complex marine environments, ensuring the real-time performance and reliability of tide level measurement, with measurement accuracy reaching the centimeter level.

[0125] For better understanding and implementation of the present application, the following provides an embodiment 2 of a specific application scenario of the present application: in order to verify the effect of the present application, the skilled person builds a numerical simulation analysis environment, simulates the radar tide level measurement scene under complex sea condition, and analyzes the performance of the multi-modal fusion radar tide level intelligent adaptive filtering method in actual application. The simulation environment sets the radar working frequency to 9.4GHz, the antenna altitude to 35m, adopts dual polarization system to transmit electromagnetic wave signal, sets the phase difference of vertical polarization component and horizontal polarization component to 90 degrees, the signal bandwidth to 150MHz, the pulse repetition frequency to 2000Hz, and the receiving antenna to synchronously collect vertical polarization and horizontal polarization echo signals. The simulation scene simulates a complex measurement environment with strong sea clutter and weather clutter and multipath effect, sets the environmental wind speed to 12m / s, the rainfall intensity to 8mm / h, the sea wave effective wave height to 1.8m, and the surrounding of the measurement area to have a building with a height of 15m, which will form obvious multipath reflection path on the sea surface.

[0126] After the radar transmitting antenna transmits dual polarization electromagnetic wave signal to the sea surface, the vertical polarization echo signal and the horizontal polarization echo signal synchronously received by the receiving antenna contain various components such as real tide level echo, sea clutter, weather clutter and multipath false echo. The skilled person performs down-conversion processing on the received radio frequency signal, converts the radio frequency signal with a center frequency of 9.4GHz into an intermediate frequency signal with an intermediate frequency of 70MHz, and then samples the analog intermediate frequency signal through a 16-bit analog-to-digital converter at a sampling rate of 300MHz to obtain a vertical polarization digital intermediate frequency signal and a horizontal polarization digital intermediate frequency signal. As shown in FIG. 2, the time domain waveform of the vertical polarization digital intermediate frequency signal presents obvious multi-peak characteristics, the peak amplitude is 2.8V, the background noise amplitude is about 0.3V, the signal-to-noise ratio is about 9.3, and the waveform contains a clutter component with a frequency of about 1200Hz. Figure 1

[0127] The skilled person adopts Daubechies wavelet basis function to perform three-layer wavelet decomposition on the vertical polarization digital intermediate frequency signal and the horizontal polarization digital intermediate frequency signal respectively. The first layer wavelet decomposition divides the signal into low frequency approximation coefficient and high frequency detail coefficient, the first layer low frequency approximation coefficient contains 8192 sample points, and the first layer high frequency detail coefficient contains 8192 sample points. The first layer low frequency approximation coefficient is continuously subjected to second layer wavelet decomposition to obtain second layer low frequency approximation coefficient 4096 sample points and second layer high frequency detail coefficient 4096 sample points. The second layer low frequency approximation coefficient is continuously subjected to third layer wavelet decomposition to obtain third layer low frequency approximation coefficient 2048 sample points and third layer high frequency detail coefficient 2048 sample points. The variance of the first layer wavelet coefficient is 0.82, the skewness is 0.35, the variance of the second layer wavelet coefficient is 1.15, the skewness is 0.42, and the variance of the third layer wavelet coefficient is 1.68, the skewness is 0.48.​

[0128] The first variance threshold is 0.76, the second variance threshold is 0.84, and the third variance threshold is 0.92, calculated by the technician according to the background signal variance of 0.42 collected in the non-clutter environment. Comparing the first layer variance 0.82 with the first variance threshold 0.76, since 0.82 is greater than 0.76 and the first layer skewness 0.35 is less than the first skewness threshold 0.5, the first layer wavelet coefficient is marked as the first layer clutter coefficient. Comparing the second layer variance 1.15 with the second variance threshold 0.84, since 1.15 is greater than 0.84 and the second layer skewness 0.42 is less than the second skewness threshold 0.5, the second layer wavelet coefficient is marked as the second layer clutter coefficient. Comparing the third layer variance 1.68 with the third variance threshold 0.92, since 1.68 is greater than 0.92 and the third layer skewness 0.48 is less than the third skewness threshold 0.5, the third layer wavelet coefficient is marked as the third layer clutter coefficient.

[0129] The marked clutter coefficients are adaptively soft-thresholded, and the root mean square value of the amplitudes of all wavelet coefficients in the first layer clutter coefficients is 0.58. The amplitudes of each wavelet coefficient in the first layer clutter coefficients are divided by 0.58 to obtain the first layer normalized amplitudes, and the first layer normalized amplitudes are multiplied by the reference threshold factor 1.0 to obtain the first layer adaptive threshold. When the amplitude of a wavelet coefficient is less than the corresponding adaptive threshold, the coefficient is set to zero; when the amplitude of a wavelet coefficient is greater than the corresponding adaptive threshold, the adaptive threshold is subtracted from the coefficient. After the first layer processing, the first layer processed wavelet coefficients are obtained. The second layer and third layer clutter coefficients are processed in the same way to obtain the second layer processed wavelet coefficients and the third layer processed wavelet coefficients. The inverse transform of the Daubechies wavelet basis function is used for wavelet reconstruction, and the third layer processed wavelet coefficients are once inverse transformed to obtain the reconstructed second layer coefficients. The reconstructed second layer coefficients and the second layer processed wavelet coefficients are combined and once inverse transformed to obtain the reconstructed first layer coefficients. The reconstructed first layer coefficients and the first layer processed wavelet coefficients are combined and once inverse transformed to obtain the vertical polarization preliminary clutter suppression signal.

[0130] The technician extracts the time-domain waveform features and frequency-domain spectral features of the vertical polarization preliminary clutter suppression signal, as shown in Table 1:

[0131] Table 1 Characteristic parameter table of vertical polarization preliminary clutter suppression signal

[0132]

[0133] The four features are input into a pre-trained clutter recognition model as a feature vector. The model is a classifier based on a convolutional neural network architecture. The input layer of the model receives a 4-element feature vector. The first convolutional layer contains 32 convolutional kernels of size 3, which convolve the feature vector to obtain first convolutional features. The first batch of normalization layers and rectified linear unit activations are applied to obtain first activation features. The second convolutional layer contains 64 convolutional kernels of size 3, which convolve the first activation features to obtain second convolutional features. The second batch of normalization layers and rectified linear unit activations are applied to obtain second activation features. The maximum pooling layer performs maximum pooling on the second activation features to obtain pooled features. The fully connected layer contains 128 neurons, which perform fully connected operations on the pooled features to obtain fully connected features. The output layer applies a Sigmoid activation function to the fully connected features to obtain a clutter confidence score of 0.76. Since 0.76 is greater than the confidence threshold of 0.7, the technical personnel increase the clutter suppression intensity coefficient from the initial value of 1.0 by 20% to 1.2. The adaptive soft thresholding process is performed again according to the adjusted clutter suppression intensity coefficient, resulting in a more thorough clutter suppression effect.

[0134] The technical personnel perform range pulse compression and Doppler frequency shift compensation on the vertically polarized preliminary clutter suppression signal and the horizontally polarized preliminary clutter suppression signal, respectively. Range pulse compression is achieved using a matched filter, and the impulse response of the matched filter is the time reversal and conjugate of the transmitted signal waveform. The compressed range resolution is 1 m. Doppler frequency shift compensation eliminates the frequency offset caused by the target radial velocity by multiplying the phase correction factor to obtain the vertically polarized range-Doppler spectrum and the horizontally polarized range-Doppler spectrum. As shown in Figure 2

[0135] The technical personnel calculate the polarization phase difference sequence between the vertically polarized range-Doppler spectrum and the horizontally polarized range-Doppler spectrum. The vertically polarized phase is extracted for each range cell in the vertically polarized range-Doppler spectrum, and the horizontally polarized phase is extracted for each range cell in the horizontally polarized range-Doppler spectrum. The polarization phase difference is obtained by subtracting the horizontally polarized phase from the vertically polarized phase. The polarization phase difference of all 256 range cells is arranged in order of distance to form the polarization phase difference sequence. As shown in Figure 3 Figure 3

[0136] ​The polarized phase difference sequence is input into a pre-trained multi-path false target recognition model, which adopts a long short-term memory network architecture. According to the radar operating frequency of 9.4 GHz and the radar antenna altitude of 35 m, the frequency normalization coefficient is calculated as 0.94, the height normalization coefficient is calculated as 0.70, and the initial value of the number of first hidden units is calculated as 42.1. After rounding, the number of first hidden units is 42. The first long short-term memory network layer contains 42 hidden units, and the polarized phase difference sequence is subjected to time sequence feature extraction to obtain first time sequence features. The discarded time sequence features are obtained after the discarded layer with a discard rate of 0.3. The second long short-term memory network layer contains 32 hidden units, and the discarded time sequence features are subjected to time sequence feature extraction to obtain second time sequence features. The multi-path fully connected layer contains 16 neurons, and the second time sequence features are subjected to fully connected operation to obtain multi-path fully connected features. The multi-path output layer performs Sigmoid activation function operation on the multi-path fully connected features to obtain target authenticity probability values of each distance unit. As shown in Table 2:

[0137] Table 2 Target authenticity probability value table of part of distance units

[0138]

[0139] The distance units with target authenticity probability values less than the authenticity threshold value of 0.6 are marked as false target distance units, and the distance units 32, 41 and 48 are marked as false target distance units.

[0140] The technical personnel obtain digital elevation model data of the measurement area, which contains three-dimensional spatial coordinate information of the terrain and buildings, and the resolution is 3 m. According to the radar antenna position and the digital elevation model data, the ray tracing algorithm is used to calculate the multi-path propagation path. A virtual ray is emitted from the radar antenna position, and the virtual ray is reflected on the surface of the digital elevation model data. The propagation path of each virtual ray from emission to reception is recorded. Three main multi-path propagation paths are calculated, the path length of path 1 is 38 m, the direct path length is 35 m, the path length difference is 3 m, the path delay difference is 10 ns, and the path phase offset is 0.59 rad. The path length of path 2 is 42 m, the path length difference is 7 m, the path delay difference is 23 ns, and the path phase offset is 1.36 rad. The path length of path 3 is 51 m, the path length difference is 16 m, the path delay difference is 53 ns, and the path phase offset is 3.14 rad. As shown in Figure 4 The multi-path propagation path forms a complex reflection path structure in space.

[0141] The skilled person establishes a set of multi-path path parameters, including path delay difference and path phase offset corresponding to all multi-path propagation paths. According to the set of multi-path path parameters, the vertical polarization range Doppler spectrum and the horizontal polarization range Doppler spectrum corresponding to the false target distance unit are phase compensated and amplitude corrected. The phase compensation multiplies the vertical polarization range Doppler spectrum by a phase correction factor, and the phase of the phase correction factor is the negative value of the path phase offset. The amplitude correction calculates the propagation loss according to the path length of the multi-path propagation path, and the propagation loss is proportional to the square of the path length. The amplitude of the false target distance unit is adjusted according to the propagation loss. After eliminating the false tidal level echo, the real tidal level range Doppler spectrum is obtained, as shown in FIG. 8, the real tidal level range Doppler spectrum presents a single peak value characteristic in the range dimension, and the peak position corresponds to the distance unit 28. Figure 5

[0142] The skilled person performs peak detection on the real tidal level range Doppler spectrum, sets the amplitude threshold to 1.5V, identifies the sample point with an amplitude greater than the amplitude threshold, and the peak position is the distance unit 28. The peak distance information is obtained by multiplying the peak position by the distance resolution 1m, and the peak distance is 28m. The current tidal level height value is calculated, the height normalized value 0.35 is obtained by dividing the radar antenna altitude 35m by the standard height 100m, the distance normalized value 0.14 is obtained by dividing the peak distance information 28m by the standard distance 200m, and the current tidal level height value 21m is obtained by subtracting 0.14 from 0.35 and multiplying by the standard height 100m. The skilled person records the current tidal level height value at multiple time points to form a tidal level time sequence, and the time interval is 5s, as shown in FIG. 9. The tidal level time sequence presents a periodic change characteristic, the fluctuation amplitude is about 0.8m, and the average tidal level height is about 21.2m. Figure 6

[0143] ​​The present application significantly improves the accuracy and reliability of radar tide measurement through the intelligent adaptive filtering method of multi-modal fusion. The traditional method relies on single polarization signal and fixed threshold filtering, which cannot effectively distinguish real tide echo and clutter interference, especially in the complex environment where strong sea clutter and multipath effect coexist, the measurement error often reaches several meters. The present application can retain the characteristics of real signals while effectively suppressing clutter components by three-layer wavelet decomposition combined with adaptive soft threshold processing, dynamically adjusting the filtering strength according to the variance and skewness of each layer of wavelet coefficients. The clutter recognition model is based on the convolutional neural network architecture, which learns the multi-dimensional features in time and frequency domains, automatically judges the degree of clutter residue and feeds back the adjustment of filtering parameters, realizing intelligent clutter suppression effect. The multipath false target recognition model uses the polarization phase difference sequence of dual-polarization signals, combined with long short-term memory network to capture time sequence features, accurately identifies and marks the false target distance unit. The ray tracing algorithm accurately calculates the multipath propagation path parameters based on digital elevation model data, providing a physical basis for phase compensation and amplitude correction, fundamentally eliminating the false echoes caused by multipath effect. The neural network quantization technology compresses the model weights and activation values from 32-bit floating point numbers to 8-bit fixed point numbers, which significantly reduces the computational complexity and storage overhead while maintaining the recognition accuracy, so that the entire filtering system can run in real time on the embedded radar processor, meeting the actual application requirements of millisecond-level response. Multi-modal information fusion fundamentally breaks through the limitations of single signal processing, through the collaborative analysis of vertical and horizontal polarization signals, providing more rich target feature dimensions, providing reliable technical support for high-precision tide measurement in complex marine environment.

[0144] It should be noted that the variables involved in the present application are explained in detail as shown in Tables 3 and 4.

[0145] Table 3 Variable Explanation Table (First Part)

[0146]

[0147] Table 4 Variable Explanation Table (Second Part)

[0148]

[0149] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A multi-modal fusion-based intelligent adaptive filtering method for radar tide levels, characterized in that, Includes the following steps: The radar transmitting antenna transmits dual-polarized electromagnetic wave signals toward the sea surface, and the receiving antenna simultaneously receives vertically polarized echo signals and horizontally polarized echo signals, and performs down-conversion processing and analog-to-digital conversion to obtain vertically polarized digital intermediate frequency signals and horizontally polarized digital intermediate frequency signals. Three-level wavelet decomposition is performed on the vertically polarized digital intermediate frequency signal and the horizontally polarized digital intermediate frequency signal to obtain the first-level wavelet coefficients, the second-level wavelet coefficients and the third-level wavelet coefficients. The first-level variance and the first-level skewness of the first-level wavelet coefficients are calculated. The second-level variance and the second-level skewness of the second-level wavelet coefficients are calculated. The third-level variance and the third-level skewness of the third-level wavelet coefficients are calculated. The first-layer variance is compared with the first variance threshold. When the first-layer variance is greater than the first variance threshold and the first-layer skewness is less than the first skewness threshold, the first-layer wavelet coefficients are marked as first-layer clutter coefficients. The second-layer variance is compared with the second variance threshold. When the second-layer variance is greater than the second variance threshold and the second-layer skewness is less than the second skewness threshold, the second-layer wavelet coefficients are marked as second-layer clutter coefficients. The third-layer variance is compared with the third-layer variance threshold. When the third-layer variance is greater than the third-layer variance threshold and the third-layer skewness is less than the third skewness threshold, the third-layer wavelet coefficients are marked as third-layer clutter coefficients. After adaptive soft thresholding of the first-layer clutter coefficients, the second-layer clutter coefficients, and the third-layer clutter coefficients, wavelet reconstruction is used to obtain the vertically polarized preliminary suppressed clutter signal and the horizontally polarized preliminary suppressed clutter signal. The features of the vertically polarized preliminary suppressed clutter signal are extracted and input into the clutter identification model to output the clutter confidence score and adjust the clutter suppression intensity coefficient. Range pulse compression and Doppler frequency shift compensation are performed on the vertically polarized and horizontally polarized preliminary suppressed clutter signals to obtain the vertically polarized range-Doppler spectrum and the horizontally polarized range-Doppler spectrum, respectively. The polarization phase difference sequence is calculated and input into the multipath false target identification model to output the target authenticity probability value. Digital elevation model data is acquired and a multipath path parameter set is established using a ray tracing algorithm. Phase compensation and amplitude correction are performed on the false target range cells based on the multipath path parameter set to obtain the true tide level range-Doppler spectrum. Peak detection is performed on the true tide level range-Doppler spectrum and peak distance information is calculated. The current tide level height value is calculated based on the radar antenna altitude and peak distance information to form a tide level time series.

2. The method according to claim 1, characterized in that, The three-layer wavelet decomposition uses the Daubechies wavelet basis function. The first-layer wavelet decomposition decomposes the vertically polarized digital intermediate frequency signal into the first-layer low-frequency approximation coefficients and the first-layer high-frequency detail coefficients.

3. The method according to claim 2, characterized in that, The second-level wavelet decomposition further decomposes the first-level low-frequency approximation coefficients into second-level low-frequency approximation coefficients and second-level high-frequency detail coefficients. The third-level wavelet decomposition further decomposes the second-level low-frequency approximation coefficients into third-level low-frequency approximation coefficients and third-level high-frequency detail coefficients.

4. The method according to claim 3, characterized in that, The variance threshold is determined based on 1.8 to 2.2 times the variance of the background signal collected in a clutter-free environment, and the skewness threshold is set to 0.

5.

5. The method according to claim 4, characterized in that, The threshold function calculation for the adaptive soft thresholding process involves dividing the amplitude of each wavelet coefficient in the clutter coefficients by the root mean square value of the amplitudes of all wavelet coefficients in the clutter coefficients to obtain a normalized amplitude, and then multiplying the normalized amplitude by a benchmark threshold factor to obtain the adaptive threshold.

6. The method according to claim 5, characterized in that, The wavelet reconstruction employs the inverse transform of the Daubechies wavelet basis functions. The wavelet coefficients after the third layer of processing are inversely transformed to obtain the reconstructed second-layer coefficients. The reconstructed second-layer coefficients and the wavelet coefficients after the second layer of processing are combined and then inversely transformed to obtain the reconstructed first-layer coefficients.

7. The method according to claim 6, characterized in that, The extraction of features for the vertically polarized, initially suppressed clutter signal includes the extraction of time-domain waveform features and frequency-domain spectral features. The time-domain waveform features include the peak-to-average ratio and the rising slope, while the frequency-domain spectral features include the position of the main peak frequency and the concentration of spectral energy.

8. The method according to claim 7, characterized in that, The clutter identification model is a classifier based on a convolutional neural network architecture, and the input layer receives a feature vector composed of time-domain waveform features and frequency-domain spectral features.

Images