Radar clutter suppression method based on end-to-end model

Abstract

The invention discloses a radar clutter suppression method based on an end-to-end model, and the method comprises the steps: collecting radar echo data, target labeling information and environment data, carrying out the data preprocessing, including data cleaning, normalization, feature extraction and data enhancement, and dividing the data into a training set, a verification set and a test set; the method comprises the following steps: constructing a model based on a Transform architecture, wherein the model comprises an input layer, an encoder, a decoder and an output layer; performing model training by using the training set, and performing verification by using the verification set; performing model evaluation by using the test set, and optimizing the model through hyper-parameter adjustment and model pruning; and acquiring radar echo signals in real time, preprocessing the radar echo signals, inputting the preprocessed radar echo signals into the trained model, and performing model reasoning and result decoding to realize radar clutter suppression. The invention aims to improve the recognition and suppression capability of the radar on various clutters in a complex water surface environment.

Description

Technical Field

[0001] The present invention relates to a method for suppressing radar clutter, and particularly to a method for suppressing radar clutter based on an end-to-end model. Background Art

[0002] In the field of surface situation awareness, surface surveillance radars and ship navigation radars, as key monitoring means, their accuracy and reliability are crucial. However, the echo signals received by the radar during operation are often interfered by various types of clutter, which poses a great challenge to the accurate detection and tracking of surface targets.

[0003] Traditional radar clutter suppression methods have significant defects. For example, STC (Sensitivity Time Control) clutter suppression mainly reduces the influence of strong clutter at close range by adjusting the sensitivity of the radar receiver in the time dimension. However, this method often falls short when dealing with long-range or complex clutter distributions, which may lead to over-weakening of target signals or incomplete clutter suppression. Although the CFAR (Constant False Alarm Rate) clutter suppression method can set thresholds according to the statistical characteristics of clutter to distinguish targets from clutter to a certain extent, it highly depends on accurate assumptions about the clutter distribution. In the actual complex surface environment, the intensity and distribution of clutter often show high uncertainty and dynamic changes, making it easy for the CFAR method to have inaccurate threshold setting problems, resulting in target misjudgment or missed detection.

[0004] Cloud and rain clutter suppression methods usually rely on prior knowledge of the physical characteristics of clouds and rain and adopt specific filtering algorithms to reduce the influence of cloud and rain clutter. However, the formation and evolution processes of clouds and rain are extremely complex, and their influence on radar echoes also has great randomness and variability. This makes the cloud and rain clutter suppression method have unstable effects when facing different types and intensities of clouds and rain, and it is difficult to ensure the effective extraction of target echoes. Most of the sea clutter suppression methods rely on the analysis of the sea clutter spectrum and corresponding filtering processing. However, the motion laws of sea waves are affected by various factors such as wind speed and wind direction, resulting in diverse characteristics of sea clutter. In addition, the interaction between sea clutter and other types of clutter further increases its complexity, making it difficult for traditional sea clutter suppression methods to cope with complex actual situations.

[0005] At the same time, these traditional clutter suppression methods often operate independently and lack comprehensive consideration of the mutual relationships between different types of clutter. In practical applications, operators need to have rich experience and professional knowledge to select and adjust different methods. The operation process is cumbersome and prone to errors. Moreover, since they are respectively used to process specific types of clutter, it is difficult to adapt to complex and changeable surface environments and the situation of multiple types of clutter existing simultaneously.

[0006] With the increasing frequency and diversification of surface activities, the requirements for radar performance are constantly rising. Existing clutter suppression technologies are already difficult to meet the needs of high-precision, high-reliability, and real-time surface situation awareness. There is an urgent need for a brand-new clutter suppression method that can comprehensively consider various clutter effects and has adaptive and efficient processing capabilities. Summary of the Invention

[0007] Object of the Invention: The object of the present invention is to provide a radar clutter suppression method based on an end-to-end model, which improves the performance and efficiency of radar clutter suppression and provides more accurate and reliable information support for surface situation awareness.

[0008] Technical Solution: A radar clutter suppression method based on an end-to-end model according to the present invention includes:

[0009] (1) Collect radar echo data, target annotation information, and environmental data, and perform data preprocessing, including data cleaning, normalization, feature extraction, and data augmentation. Divide the data into a training set, a validation set, and a test set;

[0010] (2) Build a model based on the Transformer architecture, including an input layer, an encoder, a decoder, and an output layer;

[0011] (3) Use the training set to train the model and use the validation set for validation;

[0012] (4) Use the test set to evaluate the model and optimize the model through hyperparameter tuning and model pruning;

[0013] (5) Real-time obtain the radar echo signal, perform preprocessing, and then input it into the trained model. Through model inference and result decoding, radar clutter suppression is achieved.

[0014] Preferably, for the feature extraction, radar signal features are extracted, including Doppler frequency shift, radar echo intensity, and target trajectory; the data augmentation is to simulate different radar operating modes, meteorological conditions, or target motion states, and mix the simulated radar echo signals with the collected radar echo signals to generate diverse training data.

[0015] Preferably, the input layer takes the preprocessed radar echo signal and its extracted features as input data, passes through an embedding layer, uses an embedding algorithm to map the input data into a vector space, and adds encodings to the vectors at each position according to the time or space order information of the data;

[0016] The encoder includes multiple encoder layers. Each encoder layer consists of two parts: a multi-head self-attention mechanism and a feed-forward neural network. Among them, the multi-head self-attention mechanism divides the input data into multiple subspaces, calculates the attention weights in parallel in each subspace, allowing the model to learn the features of the input signal from different perspectives and levels; the feed-forward neural network performs non-linear transformation on the feature vectors at each position after being processed by the self-attention mechanism to further extract and refine the features; between each encoder layer, a residual connection technique is adopted, adding the input of each layer to the output after intra-layer calculation and then performing normalization processing;

[0017] The decoder includes a masked self-attention mechanism. The decoder uses the encoder-decoder attention mechanism, taking the output information of the encoder as a reference to generate the final output result;

[0018] The output layer maps the output of the decoder to the output space of the target task.

[0019] Preferably, for model training, the cross-entropy loss function is selected to measure the difference between the class probability distribution of the model output and the true label; the Adam optimizer is selected for model training.

[0020] Preferably, for model training, the preprocessed training set is loaded into the model, and the input data is sequentially passed through the input layer, encoding layer, decoding layer, and output layer of the model for feature extraction and transformation to generate the output result.

[0021] Preferably, model training includes calculating the difference between the model output and the true label using the loss function to obtain the loss value. For the clutter suppression task, the cross-entropy calculation is performed between the class probability vector of the model output and the true class label; the loss value calculates the gradient of the loss value with respect to the model parameters through the backpropagation algorithm, and the Adam optimizer updates the model parameters according to the calculated gradient to minimize the loss value.

[0022] Preferably, for model evaluation, evaluation metrics are selected for calculation, including signal-to-clutter ratio, target detection rate, and target false alarm rate.

[0023] Preferably, for hyperparameter tuning, the parameters include but are not limited to learning rate, batch size, number of layers, and number of heads. The Bayesian optimization search method is used for hyperparameter optimization.

[0024] Preferably, after the model training is completed, through ensemble learning, multiple trained models are combined to generate the final prediction result by voting, averaging, or stacking.

[0025] Preferably, the model inference and result decoding include inputting the processed input data into the trained model. The model processes the input signal through an encoder and a decoder to generate a corresponding output result; decoding the result output by the model and converting it into an actual result; for the clutter suppression task, processing the input radar echo signal according to the clutter suppression weight or mask output by the model to obtain the radar echo data after clutter suppression.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: (1) The present invention takes multiple clutter suppressions as an overall learning process. By learning multi-type clutter data, capturing commonalities and characteristics, a unified and effective suppression strategy is formed, avoiding complex and independent switching between modules in traditional methods, and improving the processing efficiency and effect. (2) Utilizing a multi-layer self-attention mechanism and a feed-forward neural network structure, complex features of radar echo data are mined, including surface and deep semantic and spatio-temporal correlation features, accurately understanding the essence of clutter to achieve accurate suppression. (3) During training, the parameters are automatically adjusted according to the change of the input data distribution, adapting to different clutters, reducing manual intervention, and maintaining the optimal performance. (4) Processing the global information of radar echoes, considering the clutter relationships under different positions, different times, and different meteorological conditions, understanding the clutter rules, and accurately making radar clutter suppression decisions. (5) The advanced architecture design enables it to have excellent parallel computing capabilities, quickly processing a large amount of data, and meeting the real-time requirements of radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The technical solution of the present invention will be further described below with reference to the drawings.

[0029] The present invention provides a radar clutter suppression method based on an end-to-end model, including:

[0030] (1) Collecting radar echo data, target annotation information, and environmental data, performing data preprocessing, including data cleaning, normalization, feature extraction, and data augmentation, and dividing the data into a training set, a validation set, and a test set;

[0031] Data collection: Collecting radar echo data, target annotation information data, and environmental data from an actual radar system; among them, the target annotation information details the key attributes of the target such as position, size, and category, and the environmental data parameters cover meteorological conditions (such as wind speed, wind direction, precipitation intensity, etc.) and sea conditions (such as wave height, wave direction, tide, etc.).

[0032] Data preprocessing: Apply data cleaning algorithms to identify and remove outliers, missing values, and interference signals irrelevant to the radar clutter suppression task. For example, using a statistical analysis-based method, set a reasonable threshold range to screen out outliers that significantly deviate from the normal data distribution.

[0033] This invention uses the 5-fold standard deviation method to eliminate outliers, and the formula is as follows:

[0034] Outlier determination: |x - μ| > 5σ

[0035] Where x is the amplitude value of the radar echo signal, μ is the mean value of the batch data, σ is the standard deviation, and each batch defaults to 10 data.

[0036] Feature extraction: According to the specific requirements of the radar clutter suppression task, extract representative radar signal features. Among them, the Doppler frequency shift reflects the motion speed information of the target, the echo intensity is related to the reflection characteristics and distance of the target, and the target trajectory reflects the motion path and trend of the target. These features are used as the input of the model to help the model better learn and distinguish clutter from target signals.

[0037] Doppler frequency shift: The Doppler frequency shift is an important feature for distinguishing moving targets from stationary clutter. By analyzing the Doppler frequency shift, the model can more accurately identify moving targets and suppress stationary or low-speed moving clutter.

[0038] Among them, the Doppler frequency shift f d The calculation formula is as follows:

[0039]

[0040] Where v is the radial velocity of the target relative to the radar, calculated through the phase change rate of the radar echo; λ is the wavelength of the radar transmitted signal, determined by the radar operating frequency f c , λ = c / f c , where c = 3×10 8 m / s; typical value: for maritime surveillance radar, commonly used f c = 9.4 GHz (X-band), and the corresponding λ ≈ 3.19 cm.

[0041] Echo intensity: Directly extract the amplitude information from the radar echo signal as a measure of the echo intensity. The echo intensity is closely related to the reflection characteristics and distance of the target. Stronger echoes usually correspond to larger targets or closer distances. By analyzing the echo intensity, the model can distinguish targets of different sizes and distances, as well as the corresponding clutter.

[0042] The calculation formula for the radar echo amplitude value A is as follows:

[0043]

[0044] Among them, ρ is the target distance, and η is a constant related to radar system parameters and target characteristics, which specifically depends on the following factors: radar transmit power: the greater the radar transmit power, the greater the received signal amplitude, and the value of η will increase accordingly; antenna gain: the higher the antenna gain, the greater the received signal amplitude, and the value of η will also increase; target radar cross-section area: the larger the target's radar cross-section area, the stronger the reflected signal, and the value of η will also increase; propagation loss: the signal will be affected by atmospheric attenuation, rain fade, etc. during propagation, which will also affect the value of η.

[0045] In the radar equation, the relationship between the received signal power P r and the target distance ρ can be expressed as:

[0046]

[0047] where: P t is the radar transmit power; G t is the transmit antenna gain; G r is the receive antenna gain; λ is the radar wavelength; σ is the target's radar cross-section area (RCS); L is the system loss factor.

[0048] The radar echo amplitude value A and the received signal power P r are related as:

[0049]

[0050] Therefore, the relationship between the radar echo amplitude value A and the target distance ρ is expressed as:

[0051]

[0052] where: the constant η can be expressed as:

[0053]

[0054] where, P t is the radar transmit power, G t , G r are the antenna gains, λ is the radar wavelength, σ is the target radar cross-section area, and L is the system loss factor;

[0055] For example: P t = 10kW, G t = G r = 1000 (30dB), λ = 0.03m (X-band), σ = 10m 2 , L = 10 (10dB), then the constant η can be estimated as:

[0056]

[0057] Therefore, the relationship between the radar echo amplitude value A and the target distance can be expressed as:

[0058]

[0059] Normalization: Through linear transformation or other normalization algorithms, echo signals with different amplitude ranges are made to have a unified numerical scale, reducing the impact of data magnitude differences on gradient calculation and parameter update during model training, and improving the stability and convergence speed of model training. In the present invention, the amplitude of the radar echo signal is normalized to the interval [0, 255], and the formula is as follows:

[0060]

[0061] In the above formula, A norm is the amplitude value of the normalized radar echo signal, A is the original amplitude value of the radar echo signal, A min is the minimum value of the original amplitude of the radar echo signal, A max is the maximum value of the original amplitude of the radar echo signal.

[0062] Target trajectory: By analyzing consecutive multiple frames of radar echo signals, the movement trajectory of the target is tracked. It is achieved through algorithms such as Kalman filtering and particle filtering. The target trajectory provides information on the movement history and future trend of the target. By analyzing the target trajectory, the model predicts the state information of the target and effectively suppresses clutter that does not conform to the target trajectory.

[0063] Data augmentation: By simulating different radar operating modes, meteorological conditions, or target motion states, more diverse training data is generated using data transformation and synthesis techniques. This helps enhance the generalization ability of the model to various complex situations and avoid overfitting. The following are the specific steps:

[0064] Simulating different radar operating modes: Adjust parameters such as the radar's transmit power, pulse repetition frequency (PRF), and antenna beam width to simulate different radar operating modes. Changes in transmit power affect the propagation distance and intensity of the radar signal, and its relationship with the radar received power P r can be expressed by the radar equation:

[0065]

[0066] Where: P t is the radar transmit power; G t is the transmit antenna gain; G r is the receive antenna gain; λ is the radar wavelength; σ is the radar cross section (RCS) of the target; ρ is the target distance; L is the system loss factor.

[0067] The pulse repetition frequency fprf Determines the number of pulses the radar emits per unit time and affects the ranging range ρ of the radar max and Doppler ambiguity:

[0068]

[0069] where c = 3×10 8 m / s, f prf Generally has a value range from 500 Hz to 3000 Hz.

[0070] The antenna beamwidth affects the spatial resolution of the radar. The narrower the beam, the higher the resolution. The accuracy Δθ in angle measurement is related to the beamwidth and can be approximately expressed as where d is the antenna aperture size.

[0071] Generate radar echo data under different operating modes through software simulation or hardware adjustment. In software simulation, based on the above formulas, change parameters such as transmit power, pulse repetition frequency, and antenna beamwidth, and combine with the simulated parameters of the target and environment, and use radar signal simulation algorithms to generate corresponding radar echo data.

[0072] Simulate different meteorological conditions: According to meteorological data (such as wind speed, wind direction, precipitation intensity), simulate the influence of meteorological conditions such as cloud, rain, fog, and snow on radar echoes. Taking rainfall as an example, the attenuation of raindrops on radar waves can be calculated through a specific attenuation model, such as the Liebe model. For X-band radar, the relationship between the rainfall attenuation coefficient k (unit: dB / km) and the rainfall intensity I (unit: mm / h) is approximately k = aI b , where a and b are coefficients related to the radar frequency and raindrop size distribution.

[0073] Use meteorological models or empirical formulas to adjust the amplitude, phase, and frequency characteristics of the radar echo signal to simulate echo signals under different meteorological conditions. Calculate the influence amount of meteorological conditions on the amplitude of the radar echo signal, and then make corresponding adjustments on the basis of the original radar echo signal to generate echo signals simulating different meteorological conditions.

[0074] Simulate different target motion states: Define the motion model of the target (such as uniform linear motion, uniformly accelerated linear motion, curved turning motion, etc.) and generate corresponding radar echo signals. For a target in uniform linear motion, its Doppler frequency shift where v r is the radial velocity of the target relative to the radar. In uniformly accelerated linear motion, the Doppler frequency shift at time t: where v0 is the initial velocity and a is the acceleration. For curved turning motion, assuming the target is moving in a circular motion, its radial velocity then the Doppler frequency shift where \(v\) is the target motion speed and \(\omega\) is the angular velocity, is the initial phase. The relationship between the echo intensity \(S\) and the target distance \(\rho\) is (in the case of an ideal point target), as the target moves, the change in distance will cause the echo intensity to change.

[0075] Through software simulation, the radar echoes of the target under different motion states are simulated, including the changes in Doppler frequency shift and echo intensity. According to the above motion model and related formulas, the motion parameters of the target are set in the software, combined with the working parameters of the radar, and the radar echo signals of the target under different motion states are simulated and generated.

[0076] Data synthesis: Mix the above-simulated radar echo signals with the actually collected radar echo signals to generate diverse training data. Let the actually collected radar echo signal be \(x\) real , and the simulated radar echo signal be \(x\) sim , and the mixed signal \(x\) mix can be generated by the following formula:

[0077] \(x\) mix =\(\alpha x\) real +(1 - \(\alpha\))\(x\) sim (\(\alpha\in[0.2,0.8]\))

[0078] where \(\alpha\) is the mixing ratio.

[0079] By randomly selecting different simulated signals \(x\) sim and real signals \(x\) real , and randomly adjusting the mixing ratio \(\alpha\), a training data set containing different clutter types and target motion states is generated for model training to enhance the generalization ability of the model.

[0080] The preprocessed data is divided into a training set, a validation set, and a test set according to the ratio of 75%, 15%, and 10%. The training set is used for model training, the validation set is used to evaluate the model performance and adjust the hyperparameters of the model during training, and the test set is used to finally evaluate the generalization ability and actual application performance of the model to ensure that the performance of the model on different data sets can be effectively evaluated.

[0081] (2) Construct a model based on the Transformer architecture, including an input layer, an encoder, a decoder, and an output layer;

[0082] Input layer: Use the preprocessed radar echo signals and their extracted features as input data. Through the embedding layer, the input data is mapped to a vector space with a fixed dimension (in the present invention, the dimension is selected as 256) by using an embedding algorithm, enabling the data to be effectively calculated and processed in the model.

[0083] Meanwhile, positional encoding is added. The positional encoding can adopt methods such as sine and cosine functions to assign unique encodings to the vectors at each position according to the time or spatial order information of the data, ensuring that the model can capture the spatio-temporal order characteristics of the radar signal, which is crucial for processing radar echo data with time series and spatial distribution characteristics.

[0084] The positional encoding in the present invention is as follows:

[0085]

[0086] In the above formula, PE is the positional encoding matrix; pos is the data position index; i is the vector dimension index; dim is the dimension of the embedding layer (the dimension in the present invention is 256).

[0087] Encoder: The encoder consists of multiple encoder layers, forming a deep neural network structure. Each layer of the encoder is mainly composed of a multi-head self-attention mechanism and a feed-forward neural network. The multi-head self-attention mechanism divides the input data into multiple subspaces, calculates the attention weights in parallel in each subspace, allows the model to learn the features of the input signal from different angles and levels, and can effectively capture the long-range dependence relationships between signals. The feed-forward neural network performs non-linear transformation on the feature vectors at each position after being processed by the self-attention mechanism, further extracts and refines the features, and enhances the expression ability of the model. Between each layer of the encoder, a residual connection technique is adopted, that is, the input of this layer is added to the output after the intra-layer calculation, and then normalized. The residual connection can effectively alleviate the problem of gradient disappearance, enabling the model to maintain good training stability and generalization ability when the depth increases, and avoiding training difficulties and performance degradation caused by excessive network depth.

[0088] Decoder: The structure of the decoder is similar to that of the encoder, but includes a masked self-attention mechanism. The masked self-attention mechanism ensures the rationality and accuracy of the output result by setting a mask matrix when calculating the attention weights. The decoder uses the output information of the encoder as a reference through the encoder-decoder attention mechanism to generate the final output result. The attention mechanism enables the decoder to focus on the key information in the input data, combines the global information of the encoder and its own local information processing ability to generate high-quality output.

[0089] Output layer: The output layer maps the output of the decoder to the output space of the target task. The output layer is adjusted according to the requirements of the specific task and the characteristics of the data to ensure that the model can accurately output results that meet the actual application requirements.

[0090] (3) Use the training set to train the model and use the validation set to verify it;

[0091] Training Preparation: Loss Function Selection. Select an appropriate loss function according to the specific task type. For clutter suppression tasks, the cross-entropy loss function is selected. The cross-entropy loss function can measure the difference between the class probability distribution output by the model and the true labels. By minimizing the cross-entropy loss, the model is prompted to learn accurate classification decision boundaries. In practical applications, the loss function can also be appropriately adjusted and improved, such as adding regularization terms to prevent overfitting, etc.

[0092] The formula for the improved loss function is as follows:

[0093]

[0094] In the above formula, M is the number of clutter categories. In the present invention, M = 5, namely, five types of clutter: sea waves, cloud and rain, ship wakes, object reflections, and frequency interference.

[0095] y c is the one-hot encoding of the true label, that is:

[0096]

[0097] If it is cloud and rain clutter, then its y c in vector form is [0, 1, 0, 0, 0]; p c is the predicted probability of the c-th class, that is, the probability value output by the Softmax function of this model. p c ∈(0, 1) and ∑p c = 1; β is the L2 regularization coefficient, β ∈ [0.0001, 0.01]; w is the model weight parameter, that is, all trainable parameters in this model (including attention weights, feed-forward network parameters, etc.); Θ is the set of model parameters, that is, Θ = {w1, w2, …, w N}, and N is the total number of parameters.

[0098] Optimizer Selection: It is preferred to use the Adam optimizer for model training. The Adam optimizer combines the advantages of adaptive learning rate adjustment and momentum methods, and dynamically adjusts the learning rate according to the magnitude of the gradient and historical gradient information.

[0099] Device Configuration: To improve the efficiency of model training, high-performance computing devices should be used for training. Before training, the model and data need to be loaded onto a GPU (Graphics Processing Unit) or TPU (Tensor Processing Unit) device, and appropriate configurations should be made according to the characteristics of the device, such as setting parameters such as computing precision and memory allocation. Use a GPU (such as NVIDIA RTX5090), set mixed-precision training (FP16 / FP32), and allocate 80% of the video memory to balance speed and stability. By making full use of the advantages of high-performance computing devices, the training time of the model can be greatly shortened and the R & D efficiency can be improved.

[0100] Training process: Data loading, the preprocessed training data is loaded into the model batch by batch. Each batch contains a certain number of radar echo data samples. The selection of the batch size needs to comprehensively consider computing resources and the stability of model training. A smaller batch size may lead to instability in gradient calculation during training, but it can reduce memory occupancy; a larger batch size can improve computing efficiency, but may require higher computing resource support. During the data loading process, it is also necessary to ensure that the order of the data is random to avoid the model relying on the data order during training and improve the generalization ability of the model.

[0101] Forward propagation: The input data is sequentially passed through the input layer, encoder, decoder, and output layer of the model. During this process, the model extracts and transforms the features of the input data according to its own structure and parameters, and gradually generates the output result.

[0102] Loss calculation: Use the selected loss function to calculate the difference between the model output and the true label to obtain the loss value. For the clutter suppression task, calculate the cross-entropy between the class probability vector output by the model and the true class label; the loss value reflects the fitting degree of the model to the training data under the current parameters. The smaller the loss value, the closer the prediction result of the model is to the real situation.

[0103] Backward propagation: Calculate the gradient of the loss value with respect to the model parameters through the backward propagation algorithm. Use the selected optimizer to update the model parameters according to the calculated gradient to minimize the loss value. When updating the parameters, the optimizer will adjust the learning rate and the step size of parameter update according to its own algorithm rules, so that the model can be optimized in the direction of reducing the loss value.

[0104] Verification and saving: After each training epoch, use the validation set to evaluate the model performance and record the validation loss value. The validation set is used to monitor whether the model is overfitting or underfitting. If the validation loss value of the current model is better than the best model saved previously, save the parameters of the current model for subsequent application or further training.

[0105] (4) Use the test set to evaluate the model and optimize the model through hyperparameter tuning and model pruning;

[0106] Model evaluation: Selection of evaluation metrics. For the clutter suppression task, select the signal-to-clutter ratio as the evaluation metric. The greater the improvement in the signal-to-clutter ratio, the better the clutter suppression effect.

[0107] Test set evaluation: Evaluate the model on the test set to verify the generalization ability and actual performance of the model. Input the test set data into the trained model to obtain the prediction results of the model, and calculate the corresponding values according to the selected evaluation metrics.

[0108] The performance indicators selected for this invention are as follows:

[0109] Improvement in signal-to-clutter ratio (SCR):

[0110]

[0111] In the above formula, P signal is the target signal power (W), is the input clutter power (W), is the output clutter power (W);

[0112] Target detection rate (TDR):

[0113]

[0114] In the above formula, N correct is the number of correctly detected targets, and N total is the total number of true targets;

[0115] False alarm rate (FAR) of the target:

[0116]

[0117] In the above formula, N false is the number of false alarms of incorrect detections, and N detected is the total number of detected targets (including false alarms)

[0118] By analyzing the evaluation results, understand the clutter suppression effect of the model in different scenarios for different types of clutter.

[0119] Model optimization: Hyperparameter tuning, adjust the hyperparameters of the model, such as learning rate, batch size, number of layers, number of heads, etc., to further improve the model performance.

[0120] This invention uses Bayesian optimization to search for the optimal learning rate, number of layers, number of heads, etc., and selects the parameter combination by maximizing the expected improvement (EI) criterion. The specific steps are as follows:

[0121] Randomly sample 30 groups of parameters within the search space; Train a Gaussian model based on the initial samples; Select the candidate parameters that maximize EI: where θ is the hyperparameter vector (learning rate, number of layers, number of heads); Evaluate the objective function; Update the Gaussian model; Select the parameter combination with the minimum validation loss.

[0122] Model pruning and quantization: Prune and quantize the trained model, remove redundant parameters and weights, reduce the storage space and computational amount of the model, improve the inference speed of the model, and make it more suitable for deployment in practical applications.

[0123] Structured pruning includes:

[0124] Calculate the L1 norm of each channel in the attention head:

[0125]

[0126] In the above formula, s j is the importance score of the j-th channel; w ij is the i-th weight of the j-th channel; N is the total number of weights. Remove the channels with scores below the 5th percentile (keep the top 95%):

[0127] Kept channels = {j|s j ≥ quantile(s, 0.05)}

[0128] In the above formula, s is the score vector of all channels.

[0129] Continue to train the pruned model for 50 epochs, and reduce the learning rate to 1 / 10 of the original value.

[0130] Dynamic quantization includes:

[0131] Statistically calculate the dynamic ranges of weights and activation values on the validation set:

[0132]

[0133] In the above formula, s w is the quantization scaling factor for weights; s a is the scaling factor for activation values.

[0134] Linear quantization includes:

[0135] w quant = round(w · s w ), a quant = round(a · s a )

[0136] In the above formula, w quant is the quantized 8-bit integer weight; s a is the quantized activation value.

[0137] Dequantization inference:

[0138]

[0139] In the above formula, y is the output value after dequantization, which restores the integer calculation result to a floating-point value for use by subsequent network layers (such as activation functions, normalization layers) to ensure that the output result is compatible with the calculations of subsequent neural network layers.

[0140] Ensemble learning: An ensemble learning method is adopted to combine multiple trained models, and the final prediction result is generated through voting, averaging, stacking, etc., further improving the accuracy and stability of the model.

[0141] Among them, the models include: Informer model, Transformer model, Autoformer model;

[0142] Weight allocation strategy, where the accuracy rates of each model on the independent validation set are:

[0143]

[0144] Weight calculation:

[0145]

[0146] Weighted voting:

[0147]

[0148] In the above formula, Π(·) is the indicator function (taking 1 when predicting as class c, otherwise taking 0).

[0149] In the specific implementation process, the validation set accuracy rates are: Informer (85.9%), Transformer (93.1%), Autoformer (88.7%)

[0150] Weight calculation:

[0151]

[0152]

[0153] Voting result: If Transformer and Autoformer predict as class 3, and Informer predicts as class 2, then:

[0154] Score(3) = 0.348 + 0.331 = 0.679

[0155] Score(2) = 0.321

[0156] That is

[0157] y final = 3

[0158] (5) Real-time acquisition of radar echo signals, preprocessing them, and then inputting them into the trained model. Through model inference and result decoding, radar clutter suppression is achieved.

[0159] Inference process: Input processing, preprocess the echo signals collected by the actual radar, including operations such as normalization and feature extraction, to make their formats consistent with the training data. In practical applications, the echo signals collected by the radar may have different amplitude and feature distributions, and need to be processed according to the data preprocessing methods during training to ensure that the model can correctly process them. The normalization operation converts the amplitude of the echo signal to the range [0, 255], and feature extraction extracts the same features as the training data, such as Doppler frequency shift, echo intensity, etc., to provide a suitable data form for model input.

[0160] Model inference: Input the processed input data into the trained model. The model processes the input signal through the encoder and decoder to generate the corresponding output result.

[0161] Result decoding: Decode the result output by the model to convert it into the actual result; for the clutter suppression task, according to the clutter suppression weights or masks output by the model, process the input radar echo signals to obtain the radar echo data after clutter suppression.

[0162] By learning the clutter features in the radar signals through the model, effective suppression of clutter is achieved, the signal-to-noise ratio of the radar signals is improved, and the radar's target detection ability is enhanced. In a complex water surface environment, there are various clutter interferences such as sea clutter and rain / cloud clutter. The model can accurately identify and suppress these clutters, enabling the radar to more clearly detect the target signals, and improving the reliability and accuracy of the radar. This is of great significance for improving the water surface situation awareness ability and ensuring the safety of water activities.

Claims

1. A radar clutter suppression method based on an end-to-end model, characterized in that, Including: (1) Collect radar echo data, target annotation information, and environmental data, and perform data preprocessing, including data cleaning, normalization, feature extraction, and data augmentation. Divide the data into a training set, a validation set, and a test set. (2) Construct a model based on the Transformer architecture, including an input layer, an encoder, a decoder, and an output layer. (3) Use the training set to train the model and use the validation set for validation. (4) Use the test set to evaluate the model, and optimize the model through hyperparameter tuning and model pruning. (5) Real-time obtain the radar echo signal, perform preprocessing, and then input it into the trained model. Through model inference and result decoding, radar clutter suppression is achieved.

2. The radar clutter suppression method based on an end-to-end model according to claim 1, wherein For the feature extraction, extract radar signal features, including Doppler frequency shift, radar echo intensity, and target trajectory. The data augmentation generates diverse training data by simulating different radar operating modes, meteorological conditions, or target motion states, and mixing the simulated radar echo signals with the collected radar echo signals.

3. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that The input layer takes the preprocessed radar echo signal and its extracted features as input data. Through the embedding layer, the input data is mapped to a vector space using an embedding algorithm, and encoding is added to the vectors at each position according to the temporal or spatial order information of the data. The encoder includes multiple encoder layers. Each encoder layer consists of a multi-head self-attention mechanism and a feed-forward neural network. Among them, the multi-head self-attention mechanism divides the input data into multiple subspaces, calculates the attention weights in parallel in each subspace, allowing the model to learn the features of the input signal from different perspectives and levels. The feed-forward neural network performs a non-linear transformation on the feature vectors at each position after being processed by the self-attention mechanism to further extract and refine the features. Between each encoder layer, a residual connection technique is adopted, adding the input of each layer to the output after intra-layer calculation, and then performing normalization processing. The decoder includes a masked self-attention mechanism. The decoder uses the output information of the encoder as a reference through the encoder-decoder attention mechanism to generate the final output result. The output layer maps the output of the decoder to the output space of the target task.

4. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that, For the model training, the cross-entropy loss function is selected to measure the difference between the class probability distribution of the model output and the true label. The Adam optimizer is selected for model training.

5. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that, In the model training, the preprocessed training set is loaded into the model, and the input data passes through the input layer, encoding layer, decoding layer, and output layer of the model in sequence for feature extraction and transformation to generate an output result.

6. The radar clutter suppression method based on an end-to-end model according to claim 1, wherein, The model training includes calculating the difference between the model output and the true label using the loss function to obtain a loss value. For the clutter suppression task, the cross-entropy calculation is performed on the class probability vector of the model output and the true class label. The loss value calculates the gradient of the loss value with respect to the model parameters through the backpropagation algorithm, and the Adam optimizer updates the model parameters according to the calculated gradient to minimize the loss value.

7. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that For the model evaluation, evaluation metrics are selected for calculation, including signal-to-clutter ratio, target detection rate, and target false alarm rate.

8. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that, For the above hyperparameter tuning, the parameters include but are not limited to learning rate, batch size, number of layers, and number of heads. The Bayesian optimization search method is used for hyperparameter optimization.

9. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that, After the model training is completed, through ensemble learning, multiple trained models are combined to generate the final prediction result by voting, averaging, or stacking.

10. A method for suppressing radar clutter based on an end-to-end model according to claim 1, characterized in that, The model inference and result decoding include inputting the processed input data into the trained model. The model processes the input signal through the encoder and decoder to generate the corresponding output result; decoding the result output by the model and converting it into the actual result; for the clutter suppression task, according to the clutter suppression weight or mask output by the model, processing the input radar echo signal to obtain the radar echo data after clutter suppression.

Images