A Constant False Alarm Rate Detection System and Method for Sea Surface Small Targets Based on Depth Features and Feature Game Theory

Abstract

This invention discloses a constant false alarm rate (CFAR) detection system and method for small sea surface targets based on feature game theory, relating to the field of radar signal processing technology. The system and method provided by this invention can be used for small sea surface target detection. It includes: extracting potential high-dimensional nonlinear features from the radar received signal using a feature extractor module; transforming the target detection problem into a binary classification problem using a target detector module to obtain the probability prediction result of the target detection task; and using the output feature vector of the feature extractor module and the probability density function of the target detector module as the input of a sea state discriminator module; and training the sea state discriminator module based on feature game learning to obtain the label distribution for different sea state environments. Therefore, using the above method, the detection of small sea surface targets under time-varying sea states is achieved, improving the detection rate and reducing the false alarm rate, while also exhibiting strong robustness.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a constant false alarm rate detection system and method for depth features of small targets on the sea surface based on feature game theory. Background Technology

[0002] Deep learning is a branch of machine learning that uses the structure and function of neural networks in the human brain for data analysis and learning. A deep learning model consists of multi-layered neural networks, each containing numerous neurons, each connected to neurons in the layers above and below. Data is passed between layers through these connections, with each layer responsible for extracting increasingly higher-level features from the input data. Through the backpropagation algorithm, deep learning models can automatically adjust weights and biases to minimize the error between predicted and actual outputs, thus achieving efficient processing and prediction of complex data.

[0003] In the radar field, deep learning has made significant progress, primarily in target detection and tracking, signal processing, and waveform recognition. By leveraging the powerful learning capabilities of deep neural networks, radar systems can achieve more efficient and accurate target identification and localization, improving the accuracy and stability of target tracking. Furthermore, deep learning can optimize radar signal processing methods, such as automatically suppressing clutter and extracting effective target information, thereby enhancing radar performance. In waveform recognition, deep learning helps radar systems automatically identify the characteristics of complex targets, improving the ability to distinguish between different targets and further enhancing the radar's perception and decision-making capabilities.

[0004] Small target detection on the sea surface is a key technology in radar detection, widely used in marine surveillance, marine resource exploration, and marine environmental monitoring. However, the detection of small targets on the sea surface faces challenges such as strong sea clutter and small target radar cross-sections, resulting in poor detection performance and high false alarm probabilities for traditional physics-based algorithms (such as constant false alarm rate (CFAR) detectors) under complex sea conditions. Data-driven target detectors can improve detection performance by learning the high-dimensional statistical characteristics of target echoes and sea clutter, but they have poor robustness and poor adaptability to unknown sea conditions.

[0005] While deep learning-based target detectors can significantly improve the detection performance of small targets on the sea surface, outperforming traditional CFAR detectors, these detectors are highly sensitive to data distribution. If the statistical characteristics of the test and training data are not entirely consistent, the performance of the target detection model will be affected. For small target detection on the sea surface, sea state is time-varying, and the statistical characteristics of sea clutter fluctuate with changes in sea state. This means that the performance of deep learning-based target detectors depends on the feature extractor and the sea state; the extracted features are more likely to be related to the sea state, thus compromising the performance of the target detection model under unknown sea states. Therefore, there is an urgent need for a deep learning-based feature extractor and target detector that can identify more effective statistical features, both representing and distinguishing targets from sea clutter, while remaining independent of sea state, thereby improving the detection performance of small targets on the sea surface. Summary of the Invention

[0006] The purpose of this invention is to provide a constant false alarm rate (CFAR) detection system and method for small sea surface targets based on feature game theory, which can effectively mine target echo features independent of sea state and achieve target detection that meets the requirements of CFAR detection.

[0007] To achieve the above objectives, this invention provides a constant false alarm rate (CFAR) detection system for small sea surface targets based on feature game theory depth features, comprising:

[0008] The feature extractor module is used to process radar data to extract depth feature vectors;

[0009] The target detector module is used to distinguish target signals from sea clutter signals;

[0010] The sea state identifier module is used to predict sea states.

[0011] A constant false alarm rate (CFAR) detection method for small sea surface targets based on depth features, using feature game theory, includes the following steps:

[0012] S1. Use the time-domain data of the radar received signal as input to the feature extractor module to extract the potential high-dimensional nonlinear features of the radar received signal;

[0013] S2. The extracted high-dimensional nonlinear features are used as input to the target detector module to transform the target detection problem into a binary classification problem and obtain the probability prediction results of the target detection task.

[0014] S3. Concatenate the output feature vector of the feature extractor module and the probability density function output of the target detector module, and use them as the input of the sea state discriminator module;

[0015] S4. Based on feature game learning, the sea state discriminator module is trained to obtain the label distribution of different sea state environments.

[0016] Preferably, step S1 includes normalizing the input data and using a multilayer perceptron to enhance the features of the normalized data.

[0017] Preferably, step S2 includes evaluating the performance of the target detector module using a first loss function, as follows:

[0018]

[0019] In the formula, L t Let F and Q represent the first loss function, respectively, and let y represent the feature extractor module and the object detector module. i z represents the true label of the i-th target signal. i (ρ) represents the i-th time-dimensional fragmented data, F(z) i (ρ)) represents the depth feature of the i-th temporal fragmented data, Q(F(z)) i (ρ))) is the predicted value of the target detector module, and N represents the total number of target signals.

[0020] Preferably, in step S3, the concatenated feature expression is as follows:

[0021]

[0022] In the formula, d i (ρ) represents the input features of the sea state discriminator module. This is represented as a serial operation.

[0023] Preferably, the performance of the sea state discriminator module is evaluated using the second loss function, as follows:

[0024]

[0025] In the formula, L s Let ρ represent the second loss function. ij Let L represent the true label of the i-th target signal data under the j-th sea state, and L represent the total number of sea states.

[0026] Preferably, in step S4, the objective function for feature game learning is:

[0027]

[0028] In the formula, L represents the objective function. This indicates maximizing the performance of the sea state analyzer module. Let λ represent minimizing the first loss function, and let λ represent the trade-off parameter.

[0029] Therefore, the present invention employs the above-mentioned constant false alarm rate detection system and method for small sea surface targets based on feature game theory, which has the following technical effects:

[0030] (1) High detection performance and low false alarm probability: It can effectively eliminate the influence of different sea conditions, significantly improve the detection performance of small targets on the sea surface, effectively reduce the false alarm probability, and improve the reliability of target detection;

[0031] (2) Strong robustness: By training the feature extractor module and the sea state discriminator module through game learning, the feature extractor can effectively extract target echo features that are unrelated to sea state and deceive the sea state discriminator, adapting to various unknown sea state conditions and improving the robustness of target detection.

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the DF-CFAR detector in an embodiment of a constant false alarm rate (CFAR) detection system and method for small sea surface targets based on feature game theory.

[0034] Figure 2 This is a performance analysis of the DF-CFAR detector under different equivalent signal-to-noise ratios in the embodiments of the constant false alarm rate detection system and method for small sea surface targets based on feature game theory.

[0035] Figure 3 This is a t-SNE visualization result of the features extracted by the DF-CFAR detector in the embodiment of the constant false alarm rate detection system and method for small targets on the sea surface based on feature game theory.

[0036] Figure 4 This is a t-SNE visualization result of the features extracted by the MLP-CFAR detector in the embodiment of the constant false alarm rate detection system and method for depth features of small targets on the sea surface based on feature game theory. Detailed Implementation

[0037] The present invention will be explained in more detail through the following embodiments. The purpose of disclosing the present invention is to protect all changes and modifications within the scope of the present invention. The present invention is not limited to the following embodiments.

[0038] The present invention provides a constant false alarm rate (CFAR) detection system for small sea surface targets based on feature game theory depth features, which includes building a DF-CFAR detector model, as detailed below:

[0039] Suppose that the radar transmits T0 consecutive coherent pulses within one coherent processing interval (CPI). After sampling, matched filtering, coherent accumulation and other processes, it obtains echo data of N consecutive range cells. The echo data can be encapsulated in an N-dimensional complex vector z = [z(1); z(2); ...; z(n)].

[0040] Considering sea state ρ, the target detection problem under a specific sea state ρ is formulated as a binary hypothesis test, as follows:

[0041]

[0042] In the formula, z(n|ρ), c(n|ρ), and s(n) represent the complex-valued radar received signal, sea clutter signal, and target signal (CUT) of the unit under test, respectively. p (n|ρ) and c p (n|ρ) represent the radar received signal and sea clutter signal in the reference cell, respectively, and ρ is the number of reference cells.

[0043] like Figure 1 As shown, the DF-CFAR detector consists of three parts: a feature extractor F, a target detector Q, and a sea state discriminator D. The feature extractor F processes the input radar data to extract depth feature vectors, the target detector Q distinguishes target signals from sea clutter signals, and the sea state discriminator D predicts sea state based on the extracted features.

[0044] A feature extractor F is trained to learn and extract latent high-dimensional nonlinear features of small target echoes and sea clutter on the sea surface. The extracted feature vectors are used as input to a target detector Q. The target detector Q predicts the presence of small targets on the sea surface based on the input. The feature vectors extracted by the feature extractor F and the target detection probability of the target detector Q are matrix-concatenated and input into a sea state discriminator D to fully exploit the feature differences in different environments. Error backpropagation is used to optimize the network parameters. The accuracy of the environment label is minimized to deceive the sea state discriminator D, while the accuracy of the target detector Q is maximized. Information specific to different sea states is discarded through the minimax criterion. By statistically analyzing the probability density function of the output sample probability of the target detector Q when there is no target, a probability density threshold is determined to achieve DF-CFAR.

[0045] The method for a constant false alarm rate (CFAR) detection system for small sea surface targets based on depth features, using feature game theory, includes the following steps:

[0046] S1. Using the feature extractor module, the depth features of the radar data are extracted, as follows:

[0047] First, the radar data is processed by accumulating phase parameters to obtain segmented time-domain data, which is then used as input data to the feature extractor F to obtain high-dimensional nonlinear features. Max-min normalization is used to normalize the input data within the range [0,1], accelerating the model's learning process. Next, the feature extractor F, a five-layer deep neural network, is used to extract deep features from the normalized features. Finally, the ReLU (Rectified Linear Unit) nonlinear activation function is used to enhance the model's ability to model nonlinear relationships and its feature extraction capabilities.

[0048] The ReLU function expression is:

[0049] Relu(x) = max(0,x)

[0050] The high-dimensional latent nonlinear features of the radar received signal output by feature extractor F are calculated using the following formula:

[0051] F(z i (ρ))=σ(W N σ(W N-1 …σ(W1z i (ρ)+b1)+b N-1 )+b N )

[0052] Where σ is the activation function, W i and b i Let be the weight matrix and bias of the i-th layer, respectively.

[0053] S2. Using the target detector module, the target signal and sea clutter signal are detected, as follows:

[0054] First, the high-dimensional features extracted by the feature extractor F are input into the object detector Q, and min-max normalization is used to map them to the range [0,1]. Then, the object detection problem is transformed into a binary classification problem using the sigmoid activation function, and the probability prediction results for the object detection task are output.

[0055] Using the first loss function L t The specific expression for evaluating the performance of the object detector is as follows:

[0056]

[0057] In the formula, y i Let z be the true label of the i-th CUT data. i (ρ) represents the i-th time-dimensional fragmented data, F(z) i (ρ)) represents the deep features of the i-th temporal fragmented data obtained from the feature extraction network, Q(F(z)) i (ρ))) is the predicted value of the target detection network for the extracted features.

[0058] In this embodiment, the weights W of each perceptron layer in the target detector are updated using the backpropagation algorithm. i and bias b i This makes the first loss function L t minimize.

[0059] S3. Use the sea state detector module to determine the sea state, as follows:

[0060] The output feature vector F(z) of the feature extractor F is... i (ρ)) and the probability density function output Q(F(z) of the target detector Q i (ρ)))Serial input to sea state discriminator D.

[0061] The input expression for the sea state discriminator D is:

[0062]

[0063] in, This is represented as a serial operation.

[0064] The binary cross-entropy loss function, i.e., the second loss function, is obtained after passing through the sea state discriminator D, and is expressed as:

[0065]

[0066] Where, ρ ij Let L represent the true label of the i-th CUT data under the j-th sea state, and L be the total number of sea states.

[0067] By calculating the derivative of the second loss function with respect to each layer, and using an iterative gradient descent algorithm, the parameters of each layer are updated to make L... s Minimize and obtain the distribution of different environment labels.

[0068] S4. Based on feature game theory learning, the sea state discriminator is trained as follows:

[0069] The learning process of the sea state discriminator D aims to mine more detailed and effective features that represent scattered sea points under different sea states, thereby improving sea state discrimination performance. However, the feature extractor F is designed to mine target echo features that are unrelated to sea states, which creates a contradiction between the feature extractor F and the sea state discriminator D. To resolve this contradiction, a feature game learning model is proposed to train both the feature extractor F and the sea state discriminator D simultaneously, i.e., by minimizing the loss function L. s To train the sea state discriminator D, in order to maximize the performance of sea state prediction.

[0070] The objective function of the feature game is to make the loss function L t and L s Minimizing while maximizing the performance of sea state prediction can be expressed as:

[0071]

[0072] In the formula, λ is the trade-off parameter.

[0073] In summary, the model training process is as follows:

[0074] (1) Initialize the weight matrix W of the feature extractor F and the sea state discriminator D. i and bias b i ;

[0075] (2) Obtain the deep feature vector F(z) i (ρ)) and target detection prediction Q(F(z) i (ρ)));

[0076] (3) Calculate the depth feature vector d i (ρ);

[0077] (4) By minimizing the loss function L t Update the weight matrix and bias of the feature extractor F;

[0078] (5) By minimizing the loss function L s Update the weight matrix and bias of the sea state discriminator D;

[0079] (6) Using the percentile function P k = (k / 100)*N to calculate the probability threshold γ ρ ;

[0080] (7) Calculate the target detection result Q(F(z) i (ρ))).

[0081] Experimental data

[0082] In the experimental data testing, a publicly available X-band radar dataset was used to verify the DF-CFAR detector of this embodiment. The sea scattering data ranged from the 100th to the 450th range unit, and the target echo data was approximately at the 500th range unit. The sliding window length was set to K = 51. Specific details of the radar signal data are shown in Table 1.

[0083] Table 1. Actual X-band radar data for different bands

[0084]

[0085] To verify the performance of this method under different sea clutter ratios (SCR) and sea states, the target echo was multiplied by a constant coefficient related to the SCR and superimposed on the sea clutter. The detection ratio P was then used. d and false alarm rate P fa The following definition is used to evaluate the performance of the DF-CFAR detector:

[0086]

[0087] In the formula, N d N represents the number of targets detected. tN represents the number of true target samples. fa N represents the number of false alarms. c This represents the number of clutter samples.

[0088] To evaluate the robustness of the proposed DF-CFAR detector under unknown sea conditions, data labeled 3#1 to 3#5 were used to train the feature extractor F and the object detector Q, and data labeled 3#6 and 4#6 were used for testing. The training and testing datasets were generated using 100 and 1000 Monte Carlo simulations, respectively, and the SCR of the datasets ranged from 5 dB to 15 dB.

[0089] The DF-CFAR detector in this embodiment is compared with the classic CFAR detector, SVM, random forest, KNN, decision tree, Naive Bayes, and AdaBoost. Since the methods compared are binary classifiers with outputs of 0 or 1, they are not suitable for specific P values. fa Target detection.

[0090] P fa =10 -5 The statistical detection probability P of the DF-CFAR detector d ,like Figure 2 As shown. From Figure 2 As can be seen, the performance of the classic CFAR detector (marked by the cyan line) is poor in any sea state, making it unsuitable for target detection in complex sea conditions. In contrast, the DF-CFAR detector (marked by the magenta line) outperforms the MLP-CFAR detector (marked by the blue line), exhibiting a higher detection probability.

[0091] Specifically, for sea states ρ∈[3,4], represented by brown and yellow double-headed arrows, the detection probability of the DF-CFAR detector is improved by 2 dB (represented by the equivalent silicon controlled rectifier). Notably, the DF-CFAR detector exhibits the best detection performance at sea state ρ=3. In ocean states ρ∈[3,4], represented by red double-headed arrows, the detection probability of the DF-CFAR detector is 30% higher at sea state ρ=3 than at ρ=4. For the unseen sea state ρ=5, represented by black circles, the DF-CFAR detector outperforms the MLP-CFAR detector and significantly surpasses the classic CFAR detector, demonstrating its robustness in handling unseen sea states.

[0092] Compared with the classic CFAR detector in sea state ρ=3 ( Figure 2Compared to the MLP-CFAR (shown by the green double-headed arrows), the DF-CFAR detector improves the detection probability of equivalent silicon controlled rectifiers by 4 dB. The results show that the DF-CFAR detector is more robust than the MLP-CFAR in invisible sea conditions, and far surpasses the classic CFAR.

[0093] Compared to the classic CFAR detector in sea state ρ=3 (green double arrow line), the DF-CFAR detector improves the detection probability of equivalent silicon controlled rectifiers by 4dB.

[0094] This embodiment also utilizes the t-SNE algorithm to visualize the features extracted by the DF-CFAR detector and the MLP-CFAR detector, as detailed below:

[0095] The deep feature vector F(z) is obtained through the t-SNE algorithm. i (ρ)) is reduced to 3D, and the depth feature vectors F(z) of the DF-CFAR detector and the MLP-CFAR detector are reduced to 3D. i (ρ)) is visualized as sea state ρ∈[3,4,5], such as Figure 3 and Figure 4 As shown. From Figure 3 and Figure 4 As can be seen from this, the depth feature vector F(z) of the DF-CFAR detector i The DF-CFAR detector is more concentrated than the MLP-CFAR detector, which indicates that the DF-CFAR detector can effectively capture target echo features independent of sea state and characterize them under different sea states.

[0096] Furthermore, the DF-CFAR detector exhibits the best detection performance under different SCR and sea states, verifying the effectiveness of this method for target detection in both visible and invisible sea states, as shown in Tables 2 and 3.

[0097] Table 2. Detection ratio P of each algorithm under different sea clutter ratios (SCR) d

[0098] SCR(dB) KNN SVM Decisiontree Randomforest NaiveBayes adaBoost 0dB 0.9319 0.7053 0.6125 0.5883 0.4283 0.6328 2dB 0.9953 0.9186 0.7889 0.8461 0.7245 0.7447 4dB 0.9996 0.9916 0.8975 0.9642 0.9020 0.8181 6dB 1.0000 0.9997 0.9532 0.9979 0.9853 0.8703 8dB 1.0000 1.0000 0.9805 1.0000 0.9983 0.9061 10dB 1.0000 1.0000 0.9909 1.0000 1.0000 0.9325

[0099] Table 3 False Alarm Rates P of Each Algorithm under Different Sea States fa =10 -5

[0100] P KNN SVM Decisiontree Randomforest NaiveBayes adaBoost 3 0.0607 0.0039 0.0735 0.0012 0.0002 0.1190 4 0.0891 0.0074 0.0749 0.0031 0.0001 0.1582 5 0.0599 0.0039 0.0810 0.0015 0.0002 0.1160

[0101] Therefore, the present invention employs the above-mentioned constant false alarm rate detection system and method for small sea surface targets based on feature game theory, which can achieve a high detection rate and low false alarm rate for small sea surface targets, while maintaining excellent constant false alarm characteristics and strong robustness to unknown sea conditions.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A constant false alarm rate (CFAR) detection method for small sea surface targets based on depth features using feature game theory, characterized in that, Includes the following steps: S1. Use the time-domain data of the radar received signal as input to the feature extractor module to extract the potential high-dimensional nonlinear features of the radar received signal; S2. The extracted high-dimensional nonlinear features are used as input to the target detector module to transform the target detection problem into a binary classification problem and obtain the probability prediction results of the target detection task. The performance of the target detector module is evaluated using the first loss function, as follows: In the formula, Denotes the first loss function. , These represent the feature extractor module and the object detector module, respectively. Indicates the first The true label of a target signal Indicates the first Time-dimension fragmented data, Indicates the first Deep features of time-dimension fragmented data It is the predicted value from the target detector module. Indicates the total number of target signals; S3. Concatenate the output feature vector of the feature extractor module and the probability density function output of the target detector module, and use them as the input of the sea state discriminator module; The performance of the sea state discriminator module is evaluated using the second loss function, as follows: In the formula, This represents the second loss function. Indicates the first The target signal data in the first... The actual markings under the given sea conditions. Indicates the total number of sea states. This represents the input characteristics of the sea state discriminator module; S4. Based on feature game learning, the sea state discriminator module is trained to obtain the label distribution of different sea state environments. The objective function for feature-based game learning is: In the formula, Describe the objective function. This indicates maximizing the performance of the sea state analyzer module. This represents minimizing the first loss function. Indicates the trade-off parameters; Represents the first loss function; This represents the second loss function.

2. The constant false alarm rate detection method for depth features of small targets on the sea surface based on feature game theory as described in claim 1, characterized in that, Step S1 includes normalizing the input data and using a multilayer perceptron to enhance the features of the normalized data.

3. The constant false alarm rate detection method for depth features of small targets on the sea surface based on feature game theory as described in claim 1, characterized in that, In step S3, the concatenated feature expression is as follows: In the formula, This is represented as a serial operation.

4. A constant false alarm rate (CFAR) detection system for small sea surface targets based on feature game theory, used to implement the method described in any one of claims 1-3, characterized in that, include: The feature extractor module is used to process radar data to extract depth feature vectors; The target detector module is used to distinguish target signals from sea clutter signals; The sea state identifier module is used to predict sea states.

Images