Fuzzy deep learning single target tracking system based on generative adversarial network

Abstract

The present invention discloses a fuzzy deep learning single target tracking system based on a generative adversarial network, and relates to a fuzzy deep learning method generated by an ordinary differential equation-generative adversarial network (ODE-GAN), which is applied to the field of radar single target tracking. This method aims to solve the limitations of traditional radar target tracking algorithms when processing small amounts of data and uncertainty. The radar data is enhanced by introducing a more universal ODE-GAN generative adversarial network, and then the system is trained by deep learning, and fuzzy logic is introduced in the deep learning process to improve the adaptability and robustness of the target tracking system to different environmental conditions.

Description

Technical Field

[0001] The present invention relates to the field of radar single target tracking analysis, and more specifically, to a fuzzy deep learning single target tracking system based on a generative adversarial network. Background Art

[0002] Radar single target tracking technology is a crucial part of radar application, and its application in military, aviation, astronomy and other fields is constantly expanding. However, with the complexity of target scenes, traditional single target tracking algorithms face a series of challenges, especially in the presence of ambiguous targets or uncertain environmental conditions.

[0003] The target ambiguity problem is mainly manifested in the blurred edge of the target, unclear shape, and the mixing of the target and the background. In these complex scenarios, traditional single target tracking algorithms may find it difficult to accurately capture the characteristics of the target, resulting in increased tracking instability and errors. At the same time, environmental uncertainties, such as bad weather and strong multipath effects, also put forward higher requirements on the performance of traditional algorithms. In this context, fuzzy deep learning has become a technology that has attracted much attention, providing the system with more comprehensive perception capabilities by learning and understanding fuzzy information. The challenging problems in radar target tracking in the past, such as the complex acquisition and processing of radar data sets, the small training data sets, and the complex and inefficient feature selection process during data processing, will increase the inaccuracy of the model. Based on the traditional maneuvering target tracking algorithm, it is assumed that the target motion can be modeled by multiple predefined mathematical models. However, it is difficult to obtain enough information to predefine multiple models before tracking, so it is impossible to model the variable and uncertain maneuvers in a timely and accurate manner. In order to solve the above problems, this paper proposes a radar target tracking algorithm based on fuzzy deep learning.

[0004] In the field of deep learning, GAN is widely used in tasks such as image generation and sample synthesis. However, in the field of radar single target tracking, the combination of fuzzy information processing and deep learning is still in its early stages. The research background of this patent is based on filling the gap in this field. By introducing the synthetic data generated by ODE-GAN into the target tracking system of fuzzy deep learning, it provides the system with more comprehensive and accurate information, thereby improving its performance in complex scenarios.

[0005] Through the research of this invention, it is intended to promote the cutting-edge development of radar single target tracking technology, improve the robustness and adaptability of the system, and provide innovative solutions to target tracking problems in practical application scenarios. Summary of the invention

[0006] The present invention overcomes the defects of the prior art and proposes a fuzzy deep learning single target tracking system based on a generative adversarial network.

[0007] Fuzzy deep learning based on ordinary differential equation-generative adversarial network (ODE-GAN) realizes radar single target trajectory tracking, and the specific implementation process is divided into training stage and tracking stage. After preliminary data acquisition and preparation, radar data is collected for preprocessing, and the fuzzy deep learning network structure is designed for model training. In the training stage, the model is optimized, and the loss function is constructed by the training set trajectory and the corrected output trajectory to update the model parameters. In the testing stage, the output trajectory of the fuzzy deep learning framework is evaluated by constructing a loss function through the trajectory residual and the estimated trajectory, and the corrected trajectory is obtained and the trajectory is reconstructed to improve its robustness, thereby calculating the complete trajectory segment. The present invention realizes the full process implementation of fuzzy deep learning based on GAN generation in radar single target tracking, effectively responds to the challenges brought by target ambiguity and complex environment, and improves the robustness and tracking accuracy of the system.

[0008] A first aspect of the present invention provides a fuzzy deep learning single target tracking system based on a generative adversarial network, comprising:

[0009] In the training phase, initial radar data is collected and divided into training and test sets based on a preset ratio.

[0010] Based on the ODE differential equation and GAN network, a dynamics generation model is constructed. The training set and test set are imported into the dynamics generation model for internal law learning and training. During the learning process, a biased residual model is introduced for learning and training to obtain the potential dynamics of the trajectory, the corrected trajectory segment and the synthetic sample.

[0011] Construct a DCGAN-based induction bias generation model, import synthetic samples into the induction bias generation model for training and learning, generate data output from the dynamics generation model and the induction bias generation model, and superimpose and combine the two sets of generated data to obtain the combined output result;

[0012] Construct a fuzzy deep learning network, import the combined output results as a combined training set into the fuzzy deep learning network for fuzzy attention feature extraction and logic fusion, obtain the trajectory residual through fuzzy deep learning network training, combine the trajectory residual with the corrected trajectory segment to construct the loss function of the fuzzy deep learning network for model training;

[0013] In the tracking phase, the estimated trajectory data is obtained from the fuzzy deep learning network, the estimated trajectory data is compared with the actual trajectory data, and the estimated trajectory data is reconstructed and corrected based on the trajectory residual to obtain the complete tracking trajectory;

[0014] In the construction of the dynamics generation model based on the ODE differential equation and the GAN network, the ODE-GAN network framework is established by mixing the neural ordinary differential network into the structure of the ordinary differential equation. For an ODE, it is expressed as:

[0015]

[0016] Among them, dh(t) / dt is the specified differential equation, f(h(t), t, θ(t)) is a neural network, h(t) represents the neural network input of the specified differential equation at time t, θ(t) is the parameter of the neural network, dh(t) / dt is learned from the input data based on the neural network and feedback propagation, the neural constant differential network training process optimizes the neural network parameter θ(t) based on the training set data, t represents time, the training process updates the state through forward propagation to obtain the state at the next moment, and the forward propagation process is expressed as:

[0017]

[0018] Among them, h(t0) represents the initial state, and h(t1) represents the end state;

[0019] In the process of forward propagation, the ODESolver differential equation solver is used to solve the whole process, and the adjoint method is introduced to calculate the gradient. The parameter gradient in the network is calculated by the chain rule, and then the back propagation is performed to update the parameters. The loss function L(·) of the gradient solved by the adjoint method is expressed as:

[0020]

[0021] Where L(h(t1)) is the loss function expression, t1 and t0 represent the end time and the initial time, ODESolver is the differential equation solver, and f is the integrand expression;

[0022] In the process of optimizing the loss function L(·), the adjoint is used to calculate the gradient of the neural network. By solving the adjoint differential equation that is similar to the original differential equation but back-propagated, the adjoint formula is expressed as:

[0023]

[0024] Among them, α(t) represents the adjoint state of the hidden state h(t) at time t. The gradient can be calculated by reversing the solution in time, and the terminal state As the initial value of the solution is calculated, the gradient of the parameter θ(t) is represented by the adjoint state:

[0025]

[0026] Among them, L is the loss function.

[0027] In this solution, in the training phase, initial radar data is collected, and the initial radar data is divided into a training set and a test set based on a preset ratio, specifically:

[0028] In the training phase, initial radar data is collected from the database;

[0029] The initial radar data is preprocessed by data clarification, data outlier judgment, and data standardization;

[0030] Based on the preset ratio, the initial radar data is divided into training set and test set.

[0031] In this scheme, the dynamics generation model is constructed based on the ODE differential equation and the GAN network, and the training set and the test set are introduced into the dynamics generation model for internal law learning and training. During the learning process, the biased residual model is introduced for learning and training to obtain the potential dynamics of the trajectory, the corrected trajectory segment and the synthetic sample, specifically:

[0032] The dynamics generation model includes a discriminator and a generator, and based on input training data, simulates the state evolution of the radar target at different times to obtain the potential dynamics of the trajectory;

[0033] Generate data through the trained dynamics generation model to obtain synthetic samples;

[0034] The residual value calculation and analysis is performed based on the synthetic sample and the real sample, and the noise and deviation of the synthetic sample data set are removed and corrected through the residual value to obtain the corrected trajectory segment.

[0035] In this scheme, the DCGAN-based induction bias generation model is constructed, synthetic samples are introduced into the induction bias generation model for training and learning, the dynamics generation model and the induction bias generation model are used to generate data output, and the two sets of generated data are superimposed and combined to obtain a combined output result, specifically:

[0036] The induction deviation generation model includes a discriminator and a generator, and the synthetic samples and the corrected trajectory segments are introduced into the induction deviation generation model for training and learning until the discriminator and the generator reach a Nash equilibrium;

[0037] Integrate the synthetic samples of the dynamics generation model and the corrected trajectory segment into first sample data, and generate second sample data through the induction deviation generation model;

[0038] The first sample data and the second sample data are superimposed to obtain a combined output result.

[0039] In this scheme, the fuzzy deep learning network is constructed, and the combined output result is imported into the fuzzy deep learning network as a combined training set for fuzzy attention feature extraction and logic fusion. The trajectory residual is obtained by training the fuzzy deep learning network, and the trajectory residual is combined with the corrected trajectory segment to construct the loss function of the fuzzy deep learning network for model training, specifically:

[0040] A fuzzy deep learning network is constructed based on the fuzzy convolutional layer, pooling layer and fully connected layer network structure;

[0041] The combined output results are imported into the fuzzy deep learning network as a combined training set to perform fuzzy attention feature extraction and logic fusion learning and training, and the generated trajectory residuals are stored during the learning and training process;

[0042] The trajectory residual and the corrected trajectory segment output by the sensing deviation generation model are combined to construct the loss function, and the fuzzy deep learning network model is optimized based on the loss function and feedback parameter updating method.

[0043] In this solution, in the tracking phase, the estimated trajectory data is obtained from the fuzzy deep learning network, the estimated trajectory data is compared with the actual trajectory data, and the estimated trajectory data is reconstructed and corrected based on the trajectory residual to obtain a complete tracking trajectory, specifically:

[0044] In the tracking phase, the combined output results are imported into the optimized fuzzy deep learning network and the estimated trajectory data is output;

[0045] Comparing the estimated trajectory data with the actual trajectory data and calculating the trajectory residual, wherein the trajectory residual calculation is calculated based on a preset fuzzy mean square error loss function;

[0046] The estimated trajectory data is reconstructed and corrected based on the trajectory residual, and the obtained multiple corrected trajectory segments are spliced ​​to form a complete tracking trajectory.

[0047] The present invention discloses a fuzzy deep learning single target tracking system based on a generative adversarial network, and relates to a fuzzy deep learning method generated by an ordinary differential equation-generative adversarial network (ODE-GAN), which is applied to the field of radar single target tracking. The method aims to solve the limitations of traditional radar target tracking algorithms when processing small amounts of data and uncertainty. The radar data is enhanced by introducing a more universal ODE-GAN generative adversarial network, and then the system is trained by deep learning, and fuzzy logic is introduced in the deep learning process to improve the adaptability and robustness of the target tracking system to different environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flow chart of a fuzzy deep learning single target tracking system based on a generative adversarial network is shown in the present invention;

[0049] Figure 2 The flowchart of the ODE-GAN training process of the present invention is shown;

[0050] Figure 3 The model result flow chart of ODE-GAN of the present invention is shown;

[0051] Figure 4 A schematic diagram of the structural framework of fuzzy deep learning of the present invention is shown; DETAILED DESCRIPTION

[0052] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0053] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0054] Figure 1 A flow chart of a fuzzy deep learning single target tracking system based on a generative adversarial network is shown in the present invention.

[0055] like Figure 1 As shown, it includes the overall implementation diagram of the entire training phase and tracking phase.

[0056] A first aspect of the present invention provides a fuzzy deep learning single target tracking system based on a generative adversarial network, comprising:

[0057] In the training phase, initial radar data is collected and divided into a training set and a test set based on a preset ratio.

[0058] Based on the ODE differential equation and GAN network, a dynamics generation model is constructed. The training set and test set are imported into the dynamics generation model for internal law learning and training. During the learning process, a biased residual model is introduced for learning and training to obtain the potential dynamics of the trajectory, the corrected trajectory segment and the synthetic sample.

[0059] Construct a DCGAN-based induction bias generation model, import synthetic samples into the induction bias generation model for training and learning, generate data output from the dynamics generation model and the induction bias generation model, and superimpose and combine the two sets of generated data to obtain the combined output result;

[0060] Construct a fuzzy deep learning network, import the combined output results as a combined training set into the fuzzy deep learning network for fuzzy attention feature extraction and logic fusion, obtain the trajectory residual through fuzzy deep learning network training, combine the trajectory residual with the corrected trajectory segment to construct the loss function of the fuzzy deep learning network for model training;

[0061] In the tracking phase, the estimated trajectory data is obtained from the fuzzy deep learning network, the estimated trajectory data is compared with the actual trajectory data, and the estimated trajectory data is reconstructed and corrected based on the trajectory residual to obtain the complete tracking trajectory;

[0062] In the construction of the dynamics generation model based on the ODE differential equation and the GAN network, the ODE-GAN network framework is established by mixing the neural ordinary differential network into the structure of the ordinary differential equation. For an ODE, it is expressed as:

[0063]

[0064] Among them, dh(t) / dt is the specified differential equation, f(h(t), t, θ(t)) is a neural network, h(t) represents the neural network input of the specified differential equation at time t, θ(t) is the parameter of the neural network, dh(t) / dt is learned from the input data based on the neural network and feedback propagation, the neural constant differential network training process optimizes the neural network parameter θ(t) based on the training set data, and the training process updates the state through forward propagation to obtain the state at the next moment. The forward propagation process is expressed as:

[0065]

[0066] Among them, h(t0) represents the initial state, and h(t1) represents the end state;

[0067] In the process of forward propagation, the ODESolver differential equation solver is used to solve the whole process, and the adjoint method is introduced to calculate the gradient. The parameter gradient in the network is calculated by the chain rule, and then the back propagation is performed to update the parameters. The loss function L(·) of the gradient solved by the adjoint method is expressed as:

[0068]

[0069] Where L(h(t1)) is the loss function expression, t1 and t0 represent the end time and the initial time, ODESolver is the differential equation solver, and f is the integrand expression;

[0070] In the process of optimizing the loss function L(·), the adjoint is used to calculate the gradient of the neural network. By solving the adjoint differential equation that is similar to the original differential equation but back-propagated, the adjoint formula is expressed as:

[0071]

[0072] Among them, α(t) represents the adjoint state of the hidden state h(t) at time t. The gradient can be calculated by reversing the solution in time, and the terminal state As the initial value of the solution is calculated, the gradient of the parameter θ(t) is represented by the adjoint state:

[0073]

[0074] Among them, L is the loss function.

[0075] It should be noted that the dynamics generation model is the entire ODE-GAN network framework. The parameter θ(t) of the neural network is optimized based on the training set data, so that the prediction results of the model over the entire time interval [0, T] are closer to the real data.

[0076] It should be noted that during the forward propagation process, an ODESolver differential equation solver is used to solve the entire process. The ODESolver solver is usually regarded as a black box, and the adjoint method is used to calculate the gradient. The key to the adjoint method is that only the end state h(t1) needs to be retained as the initial value to calculate all the gradients once by calling the ODE solver. This method avoids the problems of memory explosion and uncontrollable numerical errors. The adjoint method is used to calculate and solve the adjoint equations related to the ODE, and the parameter gradients in the network are calculated by the chain rule, and then back-propagation is performed to update the parameters. The loss function L(·) of the back-propagation gradient solved by the adjoint method is expressed as:

[0077]

[0078] Where L(h(t1)) is the loss function expression, t1 and t0 represent the termination time and the initial time, ODESolver is the differential equation solver, and f is the integrand expression.

[0079] In the process of optimizing the loss function L(·), the adjoint can be used to calculate the gradient of the neural network. By solving the adjoint differential equation which is similar to the original differential equation but back-propagated, the adjoint formula can be expressed as:

[0080]

[0081] Where α(t) represents the adjoint state of the hidden state h(t) at time t. The gradient can be calculated by reversing the solution in time, using the terminal state As the initial value of the solution is calculated, the gradient of the parameter θ(t) is represented by the adjoint state:

[0082]

[0083] L is the loss function.

[0084] According to an embodiment of the present invention, in the training phase, initial radar data is collected, and the initial radar data is divided into a training set and a test set based on a preset ratio, specifically:

[0085] In the training phase, initial radar data is collected from the database;

[0086] The initial radar data is preprocessed by data clarification, data outlier judgment, and data standardization;

[0087] Based on the preset ratio, the initial radar data is divided into training set and test set.

[0088] It should be noted that the database is a database that stores existing historical radar data, and initial data that meets corresponding requirements can be exported based on certain screening conditions.

[0089] According to an embodiment of the present invention, the dynamics generation model is constructed based on the ODE differential equation and the GAN network, and the training set and the test set are introduced into the dynamics generation model for internal law learning and training. During the learning process, a biased residual model is introduced for learning and training to obtain the potential dynamics of the trajectory, the corrected trajectory segment and the synthetic sample, specifically:

[0090] The dynamics generation model includes a discriminator and a generator, and based on input training data, simulates the state evolution of the radar target at different times to obtain the potential dynamics of the trajectory;

[0091] Generate data through the trained dynamics generation model to obtain synthetic samples;

[0092] The residual value calculation and analysis is performed based on the synthetic sample and the real sample, and the noise and deviation of the synthetic sample data set are removed and corrected through the residual value to obtain the corrected trajectory segment.

[0093] It should be noted that the real samples are samples in the initial radar data. The training data refers to the training set and the test set.

[0094] According to an embodiment of the present invention, the DCGAN-based induction deviation generation model is constructed, synthetic samples are introduced into the induction deviation generation model for training and learning, the dynamics generation model and the induction deviation generation model are used to generate data output, and the two sets of generated data are superimposed and combined to obtain a combined output result, specifically:

[0095] The induction deviation generation model includes a discriminator and a generator, and the synthetic samples and the corrected trajectory segments are introduced into the induction deviation generation model for training and learning until the discriminator and the generator reach a Nash equilibrium;

[0096] Integrate the synthetic samples of the dynamics generation model and the corrected trajectory segment into first sample data, and generate second sample data through the induction deviation generation model;

[0097] The first sample data and the second sample data are superimposed to obtain a combined output result.

[0098] It should be noted that in the dynamics generation model, a dynamics generation model is trained through the ODE-GAN structure, including a discriminator and a generator. The model can simulate the state evolution of the target at different times according to the inherent rules of the training set.

[0099] Combined training set generation: Using the dynamics generation model, a biased residual model is introduced into the ODE-GAN framework to build a more universal model structure: First, the ODE-GAN model is trained to obtain the potential dynamics of the trajectory, and the residual values ​​of the synthetic samples and real samples of the ODE-GAN model are used to carefully process the data set, including removing noise and correcting bias, for training the inductive bias generation model (DCGAN). The dynamics generation model with internal rules and the inductive bias generation model are combined to obtain a combined model, and the outputs of the two models are superimposed to achieve sample combination.

[0100] Generative adversarial networks (GANs) are proposed as an alternative framework for generative training models. They consist of two network blocks: a generator (G) and a discriminator (D). The generator usually takes random noise as its input and processes it to generate output samples (such as fake images) that look similar to the data distribution. In contrast, the discriminator attempts to compare the differences between actual data samples and those generated by the generator. The main idea of ​​this framework is that it adopts an adversarial structure, that is, during the training process, the generator and the discriminator compete with each other and continuously improve the efficiency of the model until the generator can generate samples that can deceive the discriminator. This process can be regarded as a non-cooperative game that eventually reaches a Nash equilibrium.

[0101] As radar data is increasingly used, but the cost of obtaining radar datasets is high, a generative adversarial network (GAN) is used to process radar datasets to generate synthetic radar trajectory datasets to increase the labeled database and improve the data training process. The goal of GAN is to minimize the objective function by training the generator and the discriminator, which measures the generator's ability to generate a distribution p close to the real sample. dataThe objective function of GAN can be expressed as:

[0102]

[0103] In this model, D(x) is the output of the discriminator network, indicating that x comes from the true sample distribution p data G(z) is the output of the generator, which represents the sample x obtained by extracting sample z from the noise distribution p(z) and transforming it through the generator G. V is the objective function, D and G are the discriminator and generator, and E is the mathematical expectation.

[0104] The goal of the discriminator is to classify x as a real sample or a generated sample as accurately as possible, so the objective function of the discriminator is:

[0105]

[0106] The goal of the generator is to deceive the discriminator as much as possible, so that the generated sample x is mistaken by the discriminator as coming from the real sample distribution p data . Therefore, the objective function of the generator is:

[0107]

[0108] The discriminator and the generator compete with each other, so that the output of the generator gradually approaches the distribution of real data on a dynamic trajectory. During the iterative optimization process, the parameters of the generator and the discriminator are continuously adjusted through the back-propagation algorithm to achieve a dynamic balance between the two. By alternately training the discriminator and the generator, GAN can learn to generate samples that are close to the real thing.

[0109] According to an embodiment of the present invention, the fuzzy deep learning network is constructed, and the combined output result is imported into the fuzzy deep learning network as a combined training set to perform fuzzy attention feature extraction and logic fusion, and the trajectory residual is obtained by training the fuzzy deep learning network, and the trajectory residual is combined with the corrected trajectory segment to construct the loss function of the fuzzy deep learning network for model training, specifically:

[0110] A fuzzy deep learning network is constructed based on the fuzzy convolutional layer, pooling layer and fully connected layer network structure;

[0111] The combined output results are imported into the fuzzy deep learning network as a combined training set to perform fuzzy attention feature extraction and logic fusion learning and training, and the generated trajectory residuals are stored during the learning and training process;

[0112] The trajectory residual and the corrected trajectory segment output by the sensing deviation generation model are combined to construct the loss function, and the fuzzy deep learning network model is optimized based on the loss function and feedback parameter updating method.

[0113] It should be noted that the combination of these two key points (i.e., the inductive deviation generation model and the fuzzy deep learning network framework) makes the present invention have unique advantages in radar single target tracking. Through the data processing generated by GAN, the system obtains a better understanding of the target state changes, especially when facing complex environments and target ambiguity. At the same time, through the design of the fuzzy deep learning network, combined with the output of the inductive deviation generation model, real-time correction of the target trajectory is achieved, thereby improving the robustness and tracking accuracy of the system. The unique combination of these two key points brings innovative and effective solutions to the performance improvement in the field of radar single target tracking.

[0114] Figure 2 A flow chart of the ODE-GAN training process of the present invention is shown.

[0115] The present invention proposes a trajectory generation modeling method that integrates the motion dynamics model of a maneuvering target with an adversarial neural network. The training set is obtained using a preprocessed radar data set. The implementation of ODE-GAN is to learn the ODE model of radar motion trajectory based on the GAN generation model. The ODE-GAN model consists of a generation network and a discriminator network. Figure 2 Flowchart of the ODE-GAN training process.

[0116] ODE generator network: The generator network is expressed as an ordinary differential equation, and learns to generate a trajectory motion system represented by ODE. The generator network G of ODE-GAN is expressed as:

[0117] Where t is a scalar representing time, y(t) is the scalar value of the signal at time t, z is the noise vector, and θ G Denotes the network weights. The goal of the generator is to learn the dynamics of the signal y(t).

[0118] The ODE-GAN generator network is mainly used to fit the synthetic samples of radar dynamic trajectories. The ODE-Resnet model is usually trained with the help of the ordinary differential equation solver ODE-Solver. The entire trajectory process is represented as a discrete time series y = [y0...y T ], constructed by iterative equations according to the given initial condition y(0)=y0. The solution of the entire time series can be expressed as:

[0119]

[0120] in Represents an ODE generator network, y0 is the initial value of the collected radar trajectory information, ODE-Solver can be implemented using a variety of technologies, and this article chooses to use the Runge-Kutta solution method for the ODE Solver solution method. The purpose of the generator is to prevent the discriminator network from being able to distinguish between the generated samples generated by G and the real samples, and to minimize the logarithmic probability that the discriminator correctly assigns negative labels to the samples generated by G. The cross entropy loss function LOSS of the generator network is expressed as:

[0121]

[0122] in, At the same time, considering the noise distribution p z (z) and the true data distribution p data All the combined noise samples z and the corresponding real samples x0 sampled in (x0) are used to calculate the probability distribution of the sample outputs generated by the discriminator D and the generator G. The negative log-likelihood of and takes the expectation. This expected value measures the probability that the sample generated by the generator is identified as a real sample, and the goal of the generator is to minimize this expected value, that is, to generate samples close to the real sample distribution.

[0123] where G(y(t),t,z;θ G ) is the ODE generator network, which represents the log probability of assigning negative labels to the samples generated by the generator, L G θ is the loss function of the generator network. D is the network weight corresponding to the discriminator.

[0124] ODE discriminator network: The goal of the discriminator network D is to discriminate the input time series y = [y0...y T ] is from a real sample or a generated sample produced by the discriminator model G,

[0125] The discriminator network is represented as:

[0126] D(y;θ D )

[0127] The purpose of discriminator training is to maximize the correct logarithmic probability of assigning true samples and generated samples. By maximizing this loss function, the discriminator can learn to assign true samples to positive labels and assign samples generated by the generator to negative labels, thereby better distinguishing true samples from generated samples. The cross entropy loss of the discriminator is expressed as:

[0128]

[0129] Where x is from the real data distribution p data(x), D(x; θ D ) is the cross entropy loss of the discriminator to the real sample. For each noise sample z and real sample x0, the log-likelihood of the output probability distribution of the discriminator D for the noise sample and the generated sample is calculated and the expectation is taken. G(y(t),t,z;θ G ) is a sample in the generator network, D(ODESolver(G(y(t),t,z;θ G ),y0);θ D ) is the cross entropy loss of the discriminator on the samples generated by the generator.

[0130] where D(x;θ D ) is the probability of the discriminator coming from the real sample, G(y(t),t,z;θ G ) is a sample in the generator network, containing the noise input z, D(ODESolver(G(y(t),t,z;θ G ),y0);θ D ) is the cross entropy loss of the discriminator on the samples generated by the generator, L D is the discriminator loss function.

[0131] The idea of ​​general ordinary differential equations is integrated into the original ODE-GAN framework. The framework for processing radar data consists of two GAN networks. The output of the combined model after training can be expressed as:

[0132]

[0133] where G(y(t),t,z;θ G ) is the generator network of ODE-GAN, which is used to fit the dynamic synthetic samples of radar trajectories. θ (z) is the DCGAN generator network, which is used to fit the residual value between the ODE-GAN synthetic sample and the real sample to fill the gap between the two samples.

[0134] Figure 3 The model result flow chart of the ODE-GAN of the present invention is shown.

[0135] The model results of ODE-GAN are as follows Figure 3 As shown;

[0136] Through the above process, the ODE-GAN model is trained, and the cross-loss entropy of the generator and the discriminator is calculated to obtain the generative model of trajectory dynamics. The optimized ODE-GAN model synthetic samples are subtracted from the real samples to construct a training set of residual samples, and the constructed residual samples are used to train the DCGAN model. Finally, the output of the superposition model is superimposed, and the potential ECG dynamics data generated by the ODE-GAN generator is superimposed with the residual value data generated by the DCGAN generator to construct a synthetic sample that conforms to the distribution of the real sample. The output result can be expressed as:

[0137] Synthetic trajectory = G(y(t),t,z;θ G )+G(z;θ G )

[0138] G(y(t),t,z;θ G ) and G(z;θ G ) are synthetic samples generated by the ODE-GAN generator and the DCGAN generator, respectively. That is, the combined output result generated by the present invention.

[0139] Figure 4 A schematic diagram of the structural framework of fuzzy deep learning of the present invention is shown;

[0140] Although previous methods for processing data sets have many advantages, most technologies

[0141] There is a weak generalization ability. Compared with traditional neural networks, deep learning systems can capture time series data more effectively with fewer parameters and reduce computing costs. Deep learning has outstanding advantages in feature extraction, but considering the limitations of target trajectory tracking in different environments and the uncertainty factors contained in the data, the processing of rough data sets is usually solved by combining deep learning processes with fuzzy theory.

[0142] The implementation process mainly includes two contents: the implementation of the fuzzy convolution layer; the implementation of the loss function based on the fuzzy mean square error. This is very important for target tracking or other perception tasks in complex environments, because data in the real world often contains varying degrees of fuzziness and uncertainty. Therefore, this method designs and trains a fuzzy deep learning model, adds fuzzy logic on the basis of deep learning, introduces fuzzy attention feature extraction and logic fusion, and constructs a network structure such as a fuzzy convolution layer, a pooling layer, and a fully connected layer.

[0143] Implementation settings of fuzzy convolution layer: Convolution layer is used to extract features from raw radar data. Convolution operation helps to capture spatial correlation and shape information of the target. Fuzzy convolution allows to apply membership function to input features, mapping the clear value of each input element to a membership. This helps the model to better handle ambiguity and uncertainty in the input, especially in the presence of noise or fuzzy information; in addition, by introducing fuzziness in convolution operation, the model can treat input uncertainty more flexibly, which is useful for target tracking or other perception tasks in complex environments.

[0144] Since data in the real world often contains varying degrees of fuzziness and uncertainty, the patent uses the defined sigmoid fuzzy membership function to achieve feature mapping from the original radar data space to the fuzzy space when designing the fuzzy deep learning network structure. The sigmoid function is expressed as:

[0145]

[0146] Use the sigmoid fuzzy membership function to map the pixels in the input feature map to the membership values ​​in the corresponding fuzzy set, and then map them to the corresponding fuzzy set to obtain the fuzzy membership mapping result. The specific implementation steps are to map each pixel X in the input feature map X to the membership value in the corresponding fuzzy set. i,j Apply the sigmoid function to map pixel values ​​to membership values ​​between 0 and 1.

[0147] The implementation and representation of the convolution kernel in the blurred feature map is similar to the regular convolution operation. The convolution kernel is a small weight matrix used to perform a convolution operation on the input feature map to extract features or perform image processing. The convolution kernel is a two-dimensional matrix, each element represents the weight value of the corresponding position, and is used to filter the input feature map in the convolution operation. The size of the convolution kernel is usually m×n, where m and n represent the height and width of the convolution kernel respectively, and each element k i,j Represents the weight value of the convolution kernel in the i-th row and j-th column. The convolution kernel is expressed as:

[0148]

[0149] The corresponding blurred input feature map is obtained by applying sigmoid to the input feature map pixel by pixel. If the input feature map is X, the blurred input feature map is X f , expressed as

[0150]

[0151] The convolution kernel is W, and the blurred convolution is defined as:

[0152] (X f W)(i,j)=∑m ∑ n X f (i+m,j+m)·W(m,n) where i,j are the coordinates in the output feature map, and m,n are the coordinates of the convolution kernel. In the deep learning framework, the convolution kernel usually represents a tensor that contains the multiple weight values ​​of the convolution kernel. In the convolution operation, each weight of the convolution kernel is multiplied by the corresponding pixel of the input feature map, and then all the products are summed to get the result of the blurred convolution operation.

[0153] Setting of loss function of fuzzy mean square error: y pred is the predicted output of the model, y true is the true target value. When the sigmiod function is used as the fuzzy membership function, the Gaussian membership function is replaced. The Gaussian membership function can be expressed as:

[0154]

[0155] Where x represents the input value calculated by the corresponding sigmoid membership function, mean represents the center or mean of the Gaussian function, and std represents the standard deviation of the Gaussian function.

[0156] Using the modified sigmoid function as the membership function, the fuzzy mean square error (Fuzzy MES Loss) can be expressed as:

[0157]

[0158] Where N is the number of samples, y pred,i ,y true，i Represent the model prediction output and the true target value of the i-th sample respectively. σ is the sigmoid function as a Gaussian membership function, which is used to quantify the difference between the model i output value and the target value, and has two parameters, namely, mean i and standard deviation std i .

[0159] The above process sets up the implementation of fuzzy convolution in the fuzzy deep learning framework, and then sets the corresponding pooling layer and fully connected layer according to the system to build a complete fuzzy deep learning framework. The process of designing the framework also includes defining the network structure, selecting the loss function, setting the optimizer, etc.

[0160] According to an embodiment of the present invention, in the tracking stage, the estimated trajectory data is obtained from the fuzzy deep learning network, the estimated trajectory data is compared with the actual trajectory data, and the estimated trajectory data is reconstructed and corrected based on the trajectory residual to obtain a complete tracking trajectory, specifically:

[0161] In the tracking phase, the combined output results are imported into the optimized fuzzy deep learning network and the estimated trajectory data is output;

[0162] Comparing the estimated trajectory data with the actual trajectory data and calculating the trajectory residual, wherein the trajectory residual calculation is calculated based on a preset fuzzy mean square error loss function;

[0163] The estimated trajectory data is reconstructed and corrected based on the trajectory residual, and the obtained multiple corrected trajectory segments are spliced ​​to form a complete tracking trajectory.

[0164] It should be noted that in the tracking phase, the fuzzy deep learning model is used to process the input data to obtain an estimated trajectory. Then, the estimated trajectory is compared with the actual trajectory to calculate the trajectory residual. The trajectory residual represents the prediction error of the model during real-time tracking. The trajectory residual can be calculated by the difference between the estimated trajectory and the actual trajectory, using the above-mentioned fuzzy mean square error loss function.

[0165] Based on the trajectory residuals, the estimated trajectory is corrected to be closer to the actual trajectory. This can be achieved by applying some corrections to the estimated trajectory, such as adding or subtracting the trajectory residuals, to adjust the shape and position of the trajectory. The corrected trajectory segments should better match the actual trajectory, thereby improving the accuracy and stability of tracking.

[0166] The corrected trajectory segments will better reconstruct the tracking trajectory. By splicing multiple corrected trajectory segments together, a complete tracking trajectory can be obtained, which is closer to the actual trajectory. The tracking trajectory can be used in real-time tracking tasks, such as target tracking, motion analysis, etc. to obtain the target's motion path and behavior pattern. In the tracking stage, the trajectory residual is first calculated, and then the estimated trajectory is corrected according to the residual. Finally, multiple corrected trajectory segments are spliced ​​together to obtain the complete tracking trajectory.

[0167] The present invention discloses a fuzzy deep learning single target tracking system based on a generative adversarial network, and relates to a fuzzy deep learning method generated by an ordinary differential equation-generative adversarial network (ODE-GAN), which is applied to the field of radar single target tracking. The method aims to solve the limitations of traditional radar target tracking algorithms when processing small amounts of data and uncertainty. The radar data is enhanced by introducing a more universal ODE-GAN generative adversarial network, and then the system is trained by deep learning, and fuzzy logic is introduced in the deep learning process to improve the adaptability and robustness of the target tracking system to different environmental conditions.

[0168] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0169] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0170] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0171] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0172] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiment of the present invention can be embodied in the form of a software product in essence or in other words, the part that contributes to the prior art can be stored in the form of a software product, which is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A fuzzy deep learning single target tracking system based on generative adversarial network, characterized in that: include: In the training phase, initial radar data is collected and divided into training and test sets based on a preset ratio. Based on the ODE differential equation and GAN network, a dynamics generation model is constructed. The training set and test set are imported into the dynamics generation model for internal law learning and training. During the learning process, a biased residual model is introduced for learning and training to obtain the potential dynamics of the trajectory, the corrected trajectory segment and the synthetic sample. Construct a DCGAN-based induction bias generation model, import synthetic samples into the induction bias generation model for training and learning, generate data output from the dynamics generation model and the induction bias generation model, and superimpose and combine the two sets of generated data to obtain the combined output result; Construct a fuzzy deep learning network, import the combined output results as a combined training set into the fuzzy deep learning network for fuzzy attention feature extraction and logic fusion, obtain the trajectory residual through fuzzy deep learning network training, combine the trajectory residual with the corrected trajectory segment to construct the loss function of the fuzzy deep learning network for model training; In the tracking phase, the estimated trajectory data is obtained from the fuzzy deep learning network, the estimated trajectory data is compared with the actual trajectory data, and the estimated trajectory data is reconstructed and corrected based on the trajectory residual to obtain the complete tracking trajectory; In the construction of the dynamics generation model based on the ODE differential equation and the GAN network, the ODE-GAN network framework is established by mixing the neural ordinary differential network into the structure of the ordinary differential equation. For an ODE, it is expressed as: Among them, dh(t) / dt is the specified differential equation, f(h(t), t, θ(t)) is a neural network, h(t) represents the neural network input of the specified differential equation at time t, θ(t) is the parameter of the neural network, dh(t) / dt is learned from the input data based on the neural network and feedback propagation, the neural constant differential network training process optimizes the neural network parameter θ(t) based on the training set data, t represents time, the training process updates the state through forward propagation to obtain the state at the next moment, and the forward propagation process is expressed as: Among them, h(t0) represents the initial state, and h(t1) represents the end state; In the process of forward propagation, the ODESolver differential equation solver is used to solve the whole process, and the adjoint method is introduced to calculate the gradient. The parameter gradient in the network is calculated by the chain rule, and then the back propagation is performed to update the parameters. The loss function L(·) of the gradient solved by the adjoint method is expressed as: Where L(h(t1)) is the loss function expression, t1 and t0 represent the end time and the initial time, ODESolver is the differential equation solver, and f is the integrand expression; In the process of optimizing the loss function L(·), the adjoint is used to calculate the gradient of the neural network. By solving the adjoint differential equation that is similar to the original differential equation but back-propagated, the adjoint formula is expressed as: Among them, α(t) represents the adjoint state of the hidden state h(t) at time t. The gradient can be calculated by reversing the solution in time, and the terminal state As the initial value of the solution is calculated, the gradient of the parameter θ(t) is represented by the adjoint state: Among them, L is the loss function.

2. According to the fuzzy deep learning single target tracking system based on generative adversarial network in claim 1, it is characterized in that: In the training phase, initial radar data is collected and divided into a training set and a test set based on a preset ratio, specifically: In the training phase, initial radar data is collected from the database; The initial radar data is preprocessed by data clarification, data outlier judgment, and data standardization; Based on the preset ratio, the initial radar data is divided into training set and test set.

3. According to the fuzzy deep learning single target tracking system based on generative adversarial network in claim 1, it is characterized in that: The dynamics generation model is constructed based on the ODE differential equation and the GAN network, and the training set and the test set are introduced into the dynamics generation model for internal law learning and training. During the learning process, the biased residual model is introduced for learning and training to obtain the potential dynamics of the trajectory, the corrected trajectory segment and the synthetic sample, specifically: The dynamics generation model includes a discriminator and a generator, and based on input training data, simulates the state evolution of the radar target at different times to obtain the potential dynamics of the trajectory; Generate data through the trained dynamics generation model to obtain synthetic samples; The residual value calculation and analysis is performed based on the synthetic sample and the real sample, and the noise and deviation of the synthetic sample data set are removed and corrected through the residual value to obtain the corrected trajectory segment.

4. The fuzzy deep learning single target tracking system based on generative adversarial network according to claim 1 is characterized in that: The DCGAN-based induction bias generation model is constructed, synthetic samples are introduced into the induction bias generation model for training and learning, the dynamics generation model and the induction bias generation model are used to generate data output, and the two sets of generated data are superimposed and combined to obtain a combined output result, specifically: The induction deviation generation model includes a discriminator and a generator, and the synthetic samples and the corrected trajectory segments are introduced into the induction deviation generation model for training and learning until the discriminator and the generator reach a Nash equilibrium; Integrate the synthetic samples of the dynamics generation model and the corrected trajectory segment into first sample data, and generate second sample data through the induction deviation generation model; The first sample data and the second sample data are superimposed to obtain a combined output result.

5. The fuzzy deep learning single target tracking system based on generative adversarial network according to claim 1 is characterized in that: The fuzzy deep learning network is constructed, and the combined output result is imported into the fuzzy deep learning network as a combined training set to perform fuzzy attention feature extraction and logic fusion. The trajectory residual is obtained through fuzzy deep learning network training, and the trajectory residual is combined with the corrected trajectory segment to construct the loss function of the fuzzy deep learning network for model training, specifically: A fuzzy deep learning network is constructed based on the fuzzy convolutional layer, pooling layer and fully connected layer network structure; The combined output results are imported into the fuzzy deep learning network as a combined training set to perform fuzzy attention feature extraction and logic fusion learning and training, and the generated trajectory residuals are stored during the learning and training process; The trajectory residual and the corrected trajectory segment output by the sensing deviation generation model are combined to construct the loss function, and the fuzzy deep learning network model is optimized based on the loss function and feedback parameter updating method.

6. The fuzzy deep learning single target tracking system based on generative adversarial network according to claim 1 is characterized in that: In the tracking stage, the estimated trajectory data is obtained from the fuzzy deep learning network, the estimated trajectory data is compared with the actual trajectory data, and the estimated trajectory data is reconstructed and corrected based on the trajectory residual to obtain the complete tracking trajectory, specifically: In the tracking phase, the combined output results are imported into the optimized fuzzy deep learning network and the estimated trajectory data is output; Comparing the estimated trajectory data with the actual trajectory data and calculating the trajectory residual, wherein the trajectory residual calculation is calculated based on a preset fuzzy mean square error loss function; The estimated trajectory data is reconstructed and corrected based on the trajectory residual, and the obtained multiple corrected trajectory segments are spliced ​​to form a complete tracking trajectory.

Images