Ship radar identification method based on deep learning

Abstract

The invention provides a ship radar identification method based on deep learning, and relates to the technical field of ship radar signal processing. The method comprises the following steps: denoising and filtering echo signals reflected by a target, and generating spatial features and filtered target feature signals based on the denoised and filtered signals; carrying out enhancement and normalization processing on the filtered target characteristic signal, and further carrying out space-time decomposition and nonlinear fusion to obtain a fused space-time signal matrix; windowing processing is carried out on the fused space-time signal matrix, and space-time feature optimization is carried out; generating data after spatio-temporal feature fusion through a spatio-temporal feature adaptive optimization algorithm; and inputting the data after spatio-temporal feature fusion into the target recognition model, outputting a prediction result and a confidence value of the target category, and optimizing the target recognition model. The problems that a traditional radar target recognition method is difficult to effectively extract accurate spatio-temporal features and lacks adaptive ability, and the generalization ability of a target recognition model is insufficient are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of ship radar signal processing, and in particular to a ship radar recognition method based on deep learning. Background Art

[0002] Radar target recognition technology has important applications in areas such as ship navigation and collision avoidance warning, but its target recognition accuracy and efficiency in complex sea conditions still face challenges. Traditional radar target recognition methods rely on manually designed feature extraction and classification models, which are difficult to adapt to the dynamic nature of target features changing over time and in the environment. The variability of signal noise, target motion, and complex sea conditions will significantly reduce the feature extraction capabilities of traditional radar target recognition methods, resulting in reduced classification accuracy. In addition, due to the complex spatiotemporal characteristics of target signals, existing models often find it difficult to effectively capture significant features in the signal, especially in scenarios with noise interference and drastic environmental changes.

[0003] In recent years, deep learning technology has provided a new solution for radar target recognition. Through multi-layer nonlinear mapping, it can automatically extract key features from signals. However, deep learning models still have problems such as insufficient generalization ability, overfitting, and poor adaptability to changes in feature distribution in practical applications. At the same time, the spatiotemporal stability of target features is insufficient in a changing environment, and traditional deep learning models lack adaptive capabilities and are difficult to cope with the complex interaction between target signals and background noise. In addition, the under-optimization of feature fusion methods and insufficient samples also limit the performance improvement of deep learning models. To this end, there is an urgent need for an intelligent method that can adapt to complex environmental changes and efficiently extract target spatiotemporal features to improve the accuracy and efficiency of radar target recognition. Summary of the invention

[0004] The present invention provides a ship radar recognition method based on deep learning to solve the problems that, in radar target recognition, due to signal noise and environmental changes, traditional target recognition methods are difficult to effectively extract accurate spatiotemporal features; in a changing environment, the spatiotemporal features of the target are unstable, resulting in low classification accuracy; the existing target recognition model lacks adaptive capabilities and is difficult to cope with dynamic changes in target features and environmental conditions; and the target recognition model has insufficient generalization capabilities and is prone to overfitting, resulting in unsatisfactory recognition effects in practical applications.

[0005] A ship radar identification method based on deep learning comprises the following steps: S1. De-noising and filtering the echo signal reflected by the target, generating a spatial feature and a filtered target feature signal based on the de-noising and filtered signal; enhancing and normalizing the filtered target feature signal to obtain a normalized target feature signal; S2. Performing spatiotemporal decomposition on the normalized target feature signal to obtain a spatiotemporal matrix, and obtaining a fused spatiotemporal signal matrix through nonlinear fusion; performing windowing processing on the fused spatiotemporal signal matrix to obtain a windowed signal, and optimizing the spatiotemporal features of the windowed signal, and finally constructing a training data set; S3. Generate data after fusion of spatiotemporal features through spatiotemporal feature adaptive optimization algorithm to perform target classification; S4. Input the data after fusion of spatiotemporal features into the target recognition model, output the prediction results and confidence values ​​of the target category, and optimize the target recognition model.

[0006] Preferably, the S2 specifically includes: Through the improved space-time fluctuation decomposition and fusion algorithm, the normalized target feature signal is decomposed into a space-time matrix, and the formula is as follows: , in, is a spatiotemporal matrix used to decompose the normalized target feature signal at time and azimuth Frequency characteristics on ; It is at the moment Normalized target feature signal; is the wavelet basis matrix; is the translation parameter of the wavelet; is the scale parameter of the wavelet; is the frequency variable of the normalized target feature signal; is the direction weight; is a time variable; Is an imaginary unit.

[0007] Preferably, the S2 specifically includes: In the process of implementing windowing processing on the fused space-time signal matrix, an adaptive dynamic windowing algorithm is used to define an adaptive windowing function. The specific formula is: , in, is the adaptive windowing function; is the square gradient of the fused spatiotemporal signal matrix; is a discrete time index; is the azimuth; Adjust parameters for the window; is the fused spatiotemporal signal matrix.

[0008] Preferably, the S2 specifically includes: In the process of optimizing the spatiotemporal features of the windowed signal, the key spatiotemporal features of the windowed signal are extracted and standardized to obtain the standardized key spatiotemporal features, and the standardized key spatiotemporal features are expanded into a high-dimensional feature vector; based on the high-dimensional feature vector and the target label, a training data set is constructed.

[0009] Preferably, the S3 specifically includes: In the process of implementing the spatiotemporal feature adaptive optimization algorithm, the spatiotemporal features of the target are fused in a weighted manner. The specific formula is as follows: , in, Represents the data after the fusion of spatiotemporal features; A function representing the spatiotemporal feature fusion operation; is a high-dimensional feature vector; The target time characteristics; The target spatial features; and are the adaptive weights of temporal and spatial features, respectively; Represents an element-by-element dot product operation; Represents the product operation on spatial features; is the total number of time features; is the total number of spatial features.

[0010] Preferably, the S3 specifically includes: An adaptive weighting mechanism is introduced to optimize the adaptive weights of temporal and spatial features.

[0011] Preferably, the S3 specifically includes: Based on the data after the fusion of spatiotemporal features, a multi-layer perceptron is used for classification to obtain the probability that the target belongs to different categories; at the same time, a regularization term is introduced to constrain the parameters of the multi-layer perceptron; based on the regularization term and the cross entropy loss function, a total loss function is generated, and the multi-layer perceptron is trained by minimizing the total loss function.

[0012] Preferably, the S4 specifically includes: The target recognition model is optimized through a feedback mechanism, which consists of two parts: one is error correction based on the current recognition results, and the other is dynamic adjustment of the target recognition model based on environmental changes.

[0013] The beneficial effects of the technical solution of the present invention are: 1. Through multi-stage signal processing, including denoising, wavelet transform, empirical mode decomposition, frequency domain analysis and time domain feature extraction, the temporal and spatial characteristics of the target are effectively extracted. Through normalization processing, the dimensional influence between different features is eliminated to ensure the uniformity and reliability of the features, and to improve the efficiency and accuracy of subsequent target recognition and analysis.

[0014] 2. Enhance the dynamics and robustness of target recognition through improved spatiotemporal fluctuation decomposition and fusion algorithm and adaptive dynamic windowing algorithm. The adaptive dynamic windowing algorithm adjusts the window size according to the change of target feature signal to improve the clarity and recognizability of target feature signal. By expanding the training samples through data enhancement, the generalization ability of deep learning model is enhanced, and finally the anti-interference ability, accuracy and target recognition efficiency of deep learning model are improved.

[0015] 3. Combine the spatiotemporal feature adaptive optimization algorithm with the deep learning model to achieve efficient recognition and classification of radar targets. By weighted fusion of spatiotemporal features, information interaction and feature expression are enhanced, and classification accuracy is improved. By inputting the data after spatiotemporal feature fusion into the deep learning model, target prediction and confidence generation are achieved, and the target recognition model is optimized through incremental learning and feedback mechanism. The recognition window size is dynamically adjusted through the dynamic windowing mechanism to ensure the accuracy of feature extraction and improve the system's adaptability and recognition accuracy in different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a ship radar identification method based on deep learning described in the present invention. DETAILED DESCRIPTION

[0017] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0019] The following is a detailed description of a specific scheme of a ship radar recognition method based on deep learning provided by the present invention in conjunction with the accompanying drawings.

[0020] See attached Figure 1 , which shows a ship radar identification method based on deep learning provided by an embodiment of the present invention, the method comprising the following steps: S1. De-noising and filtering the echo signal reflected by the target, generating spatial features and filtered target feature signals based on the de-noising and filtered signals; enhancing and normalizing the filtered target feature signals to obtain normalized target feature signals.

[0021] The radar antenna receives the echo signal reflected by the target , which is a continuous time signal. Sampling is performed, and the sampling period is According to the sampling results, the discretized time signal is obtained .

[0022] , in, Represents analog-to-digital conversion; It is the echo signal reflected by the target; is the sampling period, which determines the time resolution after discretization; is a discrete time index, indicating the location of the sampling point; is the discretized radar signal; is a sampling pulse signal used to Extract the signal value.

[0023] Using wavelet transform Decompose it into different scales to obtain its high-frequency and low-frequency components, remove the main noise part in the high-frequency component, and retain the low-frequency component. Use inverse wavelet transform to reconstruct the processed components into denoised signals .

[0024] , in, is the wavelet transform operator; It's time Discretized radar signal; is the mother wavelet function; is the scale factor, indicating the frequency range of the decomposition; is the translation factor, representing the movement in time.

[0025] For denoised signal Through the Empirical Mode Decomposition (EMD), the signal is decomposed step by step to obtain several intrinsic mode functions (IMFs) and residual signals. The high-frequency noise is removed, the low-frequency effective components are reorganized, and the denoised and filtered signal is obtained. .

[0026] , in, It is An intrinsic mode function is used to reflect the different frequency components of the signal; is the residual signal, indicating the undecomposed part; is the total number of intrinsic mode functions obtained by decomposition.

[0027] After denoising, the filtered signal Apply fast Fourier transform to convert it from time domain to frequency domain and extract spectral features .

[0028] , in, is the frequency; is the input time domain signal; Is an imaginary unit.

[0029] Apply windowing to extract the denoised filtered signal The time domain characteristics of the target, such as peak value, average value, etc., are calculated to extract the time domain features. .

[0030] , in, is the signal after denoising and filtering; It is a windowing function used to extract local time characteristics.

[0031] The denoised filtered signal Mapping to the range-azimuth map to extract the spatial location information of the target and generate spatial features .

[0032] , in, is the target azimuth; is the target distance; Yes Mapping to a 2D coordinate system The distribution map generated above is used to visualize the spatial location information of the target.

[0033] Design a set of bandpass filters to retain the frequency range of the target characteristic signal. The target characteristic signal refers to the characteristic part related to the target in the echo signal reflected by the target received by the radar, specifically, the characteristic frequency component related to the target in the echo signal reflected by the target received by the radar, including the influence of the target motion and characteristics (such as speed, distance, azimuth, scattering characteristics, etc.) on the frequency spectrum of the echo signal reflected by the target received by the radar. The filter design adopts the window function method or the frequency domain sampling method to remove the noise of the non-target characteristic signal frequency and highlight the frequency component of the target characteristic signal to ensure that the characteristics of the target characteristic signal are not affected by excessive smoothing.

[0034] , in, is the target characteristic signal after filtering; is the coefficient of the filter, which determines the frequency characteristics of the filter; It's time De-noising filtered signal; is the delay factor, which represents the response length of the filter.

[0035] Then use histogram equalization or adaptive contrast enhancement method to improve the difference between the target and the background in the target feature signal. By stretching the dynamic range of the filtered target feature signal, the target recognition ability in low-contrast environment is enhanced.

[0036] , in, is the enhanced target characteristic signal; , are the minimum and maximum values ​​of the filtered target feature signal; , is the target dynamic range of the filtered target feature signal.

[0037] The enhanced target feature signal is linearly normalized to adjust its value to the standard range and eliminate the dimensional influence between different features.

[0038] , in, Normalized target feature signal; , are the minimum and maximum values ​​of the enhanced target feature signal.

[0039] S2. Perform spatiotemporal decomposition on the normalized target feature signal to obtain a spatiotemporal matrix, and obtain a fused spatiotemporal signal matrix through nonlinear fusion; perform windowing on the fused spatiotemporal signal matrix to obtain a windowed signal, and optimize the spatiotemporal features of the windowed signal, and finally construct a training data set.

[0040] The spatiotemporal fluctuation decomposition and fusion algorithm and the adaptive dynamic windowing algorithm are introduced to extract the dynamic spatiotemporal features of the normalized signal. Combined with data enhancement and normalization processing, a multidimensional feature data set is constructed. The normalized target feature signal is obtained by using an improved spatiotemporal fluctuation decomposition and fusion algorithm. Perform spatiotemporal decomposition to separate the time-frequency features in the target feature signal and perform multi-scale fusion.

[0041] The normalized target feature signal Decompose into space-time matrix , capturing the details of the normalized target feature signal changing over time and space.

[0042] , in, is a spatiotemporal matrix used to decompose the normalized target feature signal at time and azimuth Frequency characteristics on ; It is at the moment Normalized target feature signal; It is the wavelet basis matrix, which represents the kernel function of multi-scale wavelet transform and comes from wavelet transform theory; is the translation parameter of the wavelet; is the scale parameter of the wavelet; It is the frequency variable of the normalized target characteristic signal, which is related to the time-varying characteristics of the signal; is the direction weight, which is used to describe the characteristics of the normalized target feature signal in different directions; is a time variable, indicating the sampling time point; It is an imaginary unit, appears in the complex exponential, and is used to represent the frequency characteristics of the normalized target characteristic signal; It is the frequency component of the normalized target characteristic signal, which comes from Fourier transform.

[0043] Through nonlinear fusion, the significant features in the spatiotemporal matrix are enhanced while the interference of background noise is suppressed.

[0044] , in, is the fused spatiotemporal signal matrix; is the number of wavelet scales, indicating the number of different scales or frequency decomposition layers of the wavelet function used in the process of space-time decomposition; is the number of wavelet translations, which indicates the number of times or step size of translating the wavelet function in the wavelet transform; is the fusion weight, which is dynamically calculated by the frequency response of the space-time matrix; ) is the logarithm of the square of the spatiotemporal matrix modulus to improve weak signal response; Used to smooth the amplitude changes of signals; is the square of the second norm of the space-time matrix.

[0045] Adaptive dynamic windowing algorithm is used to adjust the window size according to the change of target feature signal after pre-processing, and further optimize the target feature extraction. Adaptive windowing function is defined so that the window size is inversely proportional to the signal change intensity, and a small window is used for fast changes and a large window is used for slow changes.

[0046] , in, is an adaptive windowing function, used to obtain the adaptive windowed signal; It is the square gradient of the fused spatiotemporal signal matrix, which is used to describe the rate of change of the target; is a discrete time index; is the azimuth; Adjust parameters for the window, related to the changing characteristics of the target; Used to enhance the saliency of the fused spatiotemporal signal matrix; is a normalization factor used to ensure the consistency of the adaptive windowing function.

[0047] Dynamic windowing is used to highlight the dynamic characteristics of the adaptive windowed signal and eliminate the interference of low-frequency background components. The specific implementation formula of dynamic windowing is: , in, is the windowed signal.

[0048] Next, the spatiotemporal features of the windowed signal are optimized. Specifically, the key spatiotemporal features of the windowed signal are extracted and standardized for use as input to the deep learning model. Specifically, the key spatiotemporal features are normalized through nonlinear mapping to improve the sparsity and robustness of the windowed signal, so that the deep learning model has a strong ability to resist noise.

[0049] , in, It is the standardized key spatiotemporal feature, which is used to improve signal sparsity and robustness and enhance the anti-interference ability of deep learning models.

[0050] The standardized key spatiotemporal features are expanded into high-dimensional feature vectors, which contain the comprehensive characteristics of the target at different azimuths and time scales.

[0051] , in, It is a high-dimensional feature vector, representing the spatiotemporal characteristics of the target; is the total dimension of the target, including spatiotemporal dynamic information; is the target dimension index, ; Used to expand the standardized key spatiotemporal features to different azimuths.

[0052] Then, data enhancement is performed through rotation, scaling, noise, etc. to expand the training samples. After data enhancement, normalization is performed again to ensure the consistency of spatiotemporal features and target feature signals.

[0053] , in, is the final signal after enhancement, including noise perturbation and rotation; , is a random noise parameter used to simulate the real environment; Used to simulate the azimuth change of the signal; Used to simulate the time periodic changes of signals.

[0054] Finally, construct the training dataset : , in, It is the label of the target, including information such as target type, trajectory, speed, etc.

[0055] S3. Generate spatiotemporal feature fusion data through spatiotemporal feature adaptive optimization algorithm to perform target classification.

[0056] In order to better classify targets, a spatiotemporal feature adaptive optimization algorithm is designed to combine temporal and spatial features to enhance the accuracy of target classification.

[0057] By fusing spatiotemporal features, the temporal information and spatial information of the target can be effectively interacted. Assuming that the spatiotemporal features of the target Included Time characteristics and These features are fused in a weighted manner.

[0058] , in, Represents the data after the fusion of spatiotemporal features; A function representing the spatiotemporal feature fusion operation; The target A time characteristic, usually including speed, acceleration, etc.; The target spatial characteristics, usually including scattering intensity, azimuth, elevation, etc.; and They are the adaptive weights of temporal and spatial features, respectively, obtained through training; Represents an element-by-element dot product operation; Represents the product operation of spatial features, adding more complex nonlinear transformations; The spatial features are normalized to avoid numerical instability.

[0059] In order to further optimize the fusion effect of spatiotemporal features, an adaptive weighting mechanism is introduced to learn how to weight the contribution of each feature by optimizing the weights.

[0060] , , in, The target time characteristics; The target time characteristics; The L2 norm of the target's spatiotemporal features. The adaptive weight formula of the temporal features uses an exponential function to perform nonlinear optimization on the weighting of the temporal features. The adaptive weight formula of the spatial features uses the Euclidean distance to calculate the similarity between spatial features as the basis for weighting.

[0061] Data after fusion of spatiotemporal features , using deep learning methods for target classification. Assume that a multi-layer perceptron (MLP) is used to complete the classification task, where the output of each layer is transformed nonlinearly through an activation function.

[0062] Assume that the input of the multilayer perceptron is , the output of each layer can be expressed as .

[0063] in, It is The output of the layer; It is The output of the layer; and They are Layer weights and biases; Is the activation function. Common activation functions include ReLU, Sigmoid or Tanh.

[0064] In the last layer of the multi-layer perceptron, the probability of the target belonging to different categories is obtained.

[0065] , in, It is the data after given spatiotemporal feature fusion Post-target category The predicted probability of and are the weights and biases of the output layer; is the number of categories; It is The output of the layer; Indicates In the layer, Output of categories; is the index of the category; Refers to the last layer of a multilayer perceptron.

[0066] To train the target classification model, i.e., the multi-layer perceptron, the cross entropy loss function is used to measure the difference between the output of the target classification model and the true label.

[0067] , in, is the cross entropy loss function, which is used to measure the difference between the model output and the true label; is the true label of the target; Indicates that given the input features After that, the model predicts samples belong to the target class probability; is the number of samples, that is, the total number of samples in the training dataset.

[0068] At the same time, in order to avoid overfitting, a regularization term is introduced, and L2 regularization is used to constrain the parameters of the target classification model: , in, is the L2 regularization term, used to prevent overfitting; is the regularization coefficient, which is used to control the influence of the regularization term; is the L2 norm of the weight matrix, which is used to penalize large weights.

[0069] The final total loss function for: , By minimizing the total loss function , able to train target classification models.

[0070] Use the gradient descent method to optimize the target classification model and update the weights and biases. The gradient is calculated using the chain rule, and the parameter update formula is: , , in, is the learning rate; and They are respectively the loss functions about and The partial derivative of .

[0071] S4. Input the data after fusion of spatiotemporal features into the target recognition model, output the prediction results and confidence values ​​of the target category, and optimize the target recognition model.

[0072] The data obtained from S3 after the fusion of spatiotemporal features is input into a deep learning model. Through a data stream processing system such as Apache Kafka or Apache Flink, real-time data from devices such as sensors and cameras is received, and the real-time data is input into the target recognition model in real time, such as a multi-layer perceptron, to generate the target category and its confidence at the current moment.

[0073] The model outputs the prediction result of the target category and their corresponding confidence values, such as classification probabilities.

[0074] , in, yes The final input features after spatiotemporal feature optimization, is the inference function of the deep learning model.

[0075] The identified targets and confidence levels are used to further optimize the target recognition model through a feedback mechanism. The feedback mechanism consists of two parts: one is error correction based on the current recognition results, and the other is dynamic adjustment of the target recognition model based on environmental changes.

[0076] Real-time monitoring of the error between the target recognition result and the actual situation. The actual state of the target can be monitored by sensors, or the recognition accuracy can be confirmed through manual feedback. If the system detects an error (e.g., misclassification, low confidence, etc.), the feedback mechanism will be triggered immediately.

[0077] If a recognition error is found, an incremental learning method is used to dynamically adjust the object recognition model. The object recognition model is fine-tuned using newly acquired data to reduce future errors.

[0078] The formula for the incremental learning method is: , in, are the adjusted model parameters; are model parameters; is the learning rate; is the gradient of the loss function; are input features, is the target label; Represents the loss function, which is determined according to the specific implementation scenario.

[0079] Based on the feedback error, the target recognition model is updated in an adaptive manner. For example, the learning rate, number of layers, regularization and other hyperparameters of the target recognition model can be adjusted according to the feedback so that the target recognition model can better cope with changes in target features. While optimizing the feedback, the performance of the target recognition model is evaluated in real time to ensure computational efficiency and accuracy in real-time scenarios. Evaluation indicators such as precision, recall and F1-score can be used to dynamically monitor the performance of target recognition.

[0080] In order to adapt to different environmental changes and target features, the dynamic windowing mechanism can adjust the recognition window size according to the actual situation. This process dynamically adjusts the size of the input data window according to the target's movement speed, scale change and environmental noise. By monitoring the target's movement speed, size change and other information, the size of the input window is dynamically adjusted to ensure the accuracy of feature extraction.

[0081] Window size update: , in, is the current window size, is the window size at the previous moment, The window adjustment amount.

[0082] If there are major changes in the environment (such as changes in lighting or background noise), the threshold can be set using expert experience to automatically trigger window adjustment.

[0083] Change Detection: , in, is the current environment status (such as light, noise, etc.); is the environmental state at the previous moment; For the amount of change.

[0084] According to the target's movement speed and size, convolution kernels of different scales are automatically selected for processing or windows of appropriate sizes are used in LSTM to capture dynamic features in time to ensure that the target can be effectively identified at different scales.

[0085] In summary, a ship radar recognition method based on deep learning has been completed.

[0086] The order of the embodiments of the invention is for description only and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0087] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A ship radar identification method based on deep learning, characterized in that: The following steps are involved: S1. De-noising and filtering the echo signal reflected by the target, and generating spatial features and filtered target feature signals based on the de-noising and filtered signals; Performing enhancement and normalization processing on the filtered target feature signal to obtain a normalized target feature signal; S2. Performing spatiotemporal decomposition on the normalized target feature signal to obtain a spatiotemporal matrix, and obtaining a fused spatiotemporal signal matrix through nonlinear fusion; performing windowing processing on the fused spatiotemporal signal matrix to obtain a windowed signal, and optimizing the spatiotemporal features of the windowed signal, and finally constructing a training data set; S3. Generate data after fusion of spatiotemporal features through spatiotemporal feature adaptive optimization algorithm to perform target classification; S4. Input the data after fusion of spatiotemporal features into the target recognition model, output the prediction results and confidence values ​​of the target category, and optimize the target recognition model.

2. The ship radar identification method based on deep learning according to claim 1 is characterized in that: The S2 specifically includes: Through the improved space-time fluctuation decomposition and fusion algorithm, the normalized target feature signal is decomposed into a space-time matrix, and the formula is as follows: ， in, is a spatiotemporal matrix used to decompose the normalized target feature signal at time and azimuth Frequency characteristics on ; It is at the moment Normalized target feature signal; is the wavelet basis matrix; is the translation parameter of the wavelet; is the scale parameter of the wavelet; is the frequency variable of the normalized target feature signal; is the direction weight; is a time variable; Is an imaginary unit.

3. The ship radar identification method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: In the process of implementing windowing processing on the fused space-time signal matrix, an adaptive dynamic windowing algorithm is used to define an adaptive windowing function. The specific formula is: ， in, is the adaptive windowing function; is the square gradient of the fused spatiotemporal signal matrix; is a discrete time index; is the azimuth; Adjust parameters for the window; is the fused spatiotemporal signal matrix.

4. The ship radar identification method based on deep learning according to claim 1, characterized in that: The S2 specifically includes: In the process of optimizing the spatiotemporal features of the windowed signal, the key spatiotemporal features of the windowed signal are extracted and standardized to obtain the standardized key spatiotemporal features, and the standardized key spatiotemporal features are expanded into a high-dimensional feature vector; based on the high-dimensional feature vector and the target label, a training data set is constructed.

5. The ship radar identification method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: In the process of implementing the spatiotemporal feature adaptive optimization algorithm, the spatiotemporal features of the target are fused in a weighted manner. The specific formula is as follows: ， in, Represents the data after the fusion of spatiotemporal features; A function representing the spatiotemporal feature fusion operation; is a high-dimensional feature vector; The target time characteristics; The target spatial features; and are the adaptive weights of temporal and spatial features, respectively; Represents an element-by-element dot product operation; Represents the product operation on spatial features; is the total number of time features; is the total number of spatial features.

6. The ship radar identification method based on deep learning according to claim 5 is characterized in that: The S3 specifically includes: An adaptive weighting mechanism is introduced to optimize the adaptive weights of temporal and spatial features.

7. The ship radar identification method based on deep learning according to claim 5, characterized in that: The S3 specifically includes: Based on the data after the fusion of spatiotemporal features, a multi-layer perceptron is used for classification to obtain the probability that the target belongs to different categories; at the same time, a regularization term is introduced to constrain the parameters of the multi-layer perceptron; based on the regularization term and the cross entropy loss function, a total loss function is generated, and the multi-layer perceptron is trained by minimizing the total loss function.

8. The ship radar identification method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: The target recognition model is optimized through a feedback mechanism, which consists of two parts: one is error correction based on the current recognition results, and the other is dynamic adjustment of the target recognition model based on environmental changes.

Images