Intelligent discrimination method and device for spatial target balance mode

Abstract

The application provides a space target balance mode intelligent discrimination method and device, including obtaining inverse synthetic aperture radar image data sets and preprocessing; a convolutional neural network and a long short-term memory network fusion model are constructed, the fusion model includes a feature extraction module, a time sequence modeling module and a classification module; the fusion model is trained based on the preprocessed inverse synthetic aperture radar image data sets; the inverse synthetic aperture radar image to be identified is preprocessed and input into the trained fusion model, and the fusion model outputs the balance mode discrimination result corresponding to the inverse synthetic aperture radar image. The balance mode discrimination is performed through the constructed convolutional neural network and long short-term memory network fusion model, the recognition accuracy and robustness can be improved; and the fusion model is trained based on the preprocessed inverse synthetic aperture radar image data sets, so that the fusion model can realize efficient and accurate balance state evaluation in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a method and apparatus for intelligent discrimination of spatial target balance patterns. Background Technology

[0002] With the rapid development of aerospace technology and intelligent sensing technology, space target monitoring has become a core component of aerospace mission planning and orbital safety management. ISAR (Inverse Synthetic Aperture Radar), as a typical microwave imaging technology, has all-weather, all-day operation capability and is less affected by environmental factors such as clouds. It has become a key technical means for space target monitoring, and its imaging results provide basic data support for target attitude recognition.

[0003] With increasing competition for low Earth orbit space resources, space target status assessment has become a core technology for ensuring the safe and efficient operation of space activities. Space target attitude recognition based on ISAR imagery aims to provide accurate assessments of target status by analyzing the dynamic characteristics of space targets, thereby determining and monitoring their operational status.

[0004] Satellite attitude control precision has improved from a few degrees in the early days to the arcsecond level in modern times. While this high-precision adjustment meets the requirements of complex missions such as Earth observation and communication networking, it also has a significant impact on status assessment. Typically, satellites frequently adjust their operational attitude according to mission requirements, changing their pitch angle or rotation axis, etc. These changes can lead to alterations in reflective properties and visible components, resulting in biases in information acquired by ground radar stations and blurred images of space targets generated through inversion. This poses a serious challenge to the status assessment of space targets.

[0005] Traditional methods for determining the status of space targets based on ISAR images are mostly based on fixed rules or simple models, making it difficult to adapt to image changes caused by frequent satellite attitude adjustments. When a satellite changes its pitch angle or rotation axis, the reflective properties and visible components of the target change. Traditional methods cannot dynamically capture these changes and cannot accurately infer the target attitude from biased radar information. Furthermore, in terms of processing efficiency, traditional methods are typically computationally complex and slow, making it difficult to meet the real-time requirements of space missions for target status assessment and to provide timely and effective decision-making support for orbital safety management and space mission planning.

[0006] Therefore, there is an urgent need for an attitude recognition method that can adapt to the characteristics of ISAR images and take into account both local details and global correlations, so as to improve the target monitoring accuracy in complex space environments and meet the urgent needs of space missions for real-time performance and reliability. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and apparatus for intelligently identifying the balance patterns of space targets.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] On the one hand, the present invention provides an intelligent discrimination method for the balance pattern of space targets, comprising the following steps:

[0010] The inverse synthetic aperture radar (ISAR) image dataset is acquired and preprocessed. The IAR image dataset uses time-series correlated IAR image sequences as the basic sample unit. The sample data in the IAR image dataset is stored in sequence folders corresponding to different balance modes according to the corresponding balance modes.

[0011] A fusion model of convolutional neural network and long short-term memory network is constructed. The fusion model includes a feature extraction module, a temporal modeling module and a classification module. The feature extraction module uses convolutional neural network to extract spatial features of single frame images. The temporal modeling module uses long short-term memory network to capture temporal correlation features of sequences. The classification module outputs the category probability distribution of spatial target balance patterns.

[0012] The fusion model was trained based on the preprocessed inverse synthetic aperture radar image dataset;

[0013] After preprocessing, the inverse synthetic aperture radar image to be identified is input into the trained fusion model, which outputs the balanced mode discrimination result corresponding to the inverse synthetic aperture radar image.

[0014] Furthermore, the inverse synthetic aperture radar image dataset is obtained according to the following steps:

[0015] Simulations were performed on publicly available satellite models based on the characteristics of different equilibrium modes to obtain sample data corresponding to different equilibrium modes.

[0016] Construct sequence folders corresponding to different balancing modes, and store sample data into the corresponding sequence folders according to the balancing mode; by recursively traversing the sequence folders, extract the complete path of all sequence folders, use the name of the sequence folder to which the sample data belongs as the label of the sample data, and establish a mapping between sample data path and category label; thus obtaining the inverse synthetic aperture radar image dataset.

[0017] Furthermore, the balance modes include triaxial stabilization, spin stabilization, and free roll.

[0018] Furthermore, the preprocessing includes the following steps:

[0019] Perform consistent cropping on the sample data;

[0020] The cropped sample data is processed to grayscale, compressing the three-channel color image into a single-channel grayscale image.

[0021] Furthermore, the feature extraction module includes a sequence folding layer, a multi-layer convolutional block, and a sequence unrolling layer; the sequence folding layer converts the preprocessed sample data into batch images; the multi-layer convolutional block is cascaded to extract batch image features; and the sequence unrolling layer restores the batch image features to a temporal structure.

[0022] Furthermore, each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max pooling layer in sequence.

[0023] Furthermore, the temporal modeling module includes a feature flattening layer, a recurrent neural network layer, and a temporal pooling layer; the feature flattening layer converts the output of the received feature extraction module into a one-dimensional feature vector; the recurrent neural network layer captures the dynamic correlation between frames through a gating mechanism, learns temporal features, and completes the one-dimensional feature vector temporal modeling; the temporal pooling layer extracts the overall features of the sequence.

[0024] Furthermore, the classification module includes a fully connected layer, a softmax layer, and a classification layer; the fully connected layer maps the output of the temporal modeling module to an output space with a consistent number of spatial target balance mode categories; the softmax layer normalizes the output value of the fully connected layer into a probability distribution; the classification layer determines the balance mode category to which the sample data belongs based on the maximum probability criterion, and calculates the difference between the prediction result and the true label using the cross-entropy loss function.

[0025] Furthermore, the fusion model is trained based on the preprocessed inverse synthetic aperture radar image dataset, including:

[0026] The preprocessed inverse synthetic aperture radar image dataset was divided into a training set and a validation set in an 8:2 ratio, and the distribution of balanced pattern categories in the training set and the validation set was made consistent by stratified sampling and data shuffling.

[0027] The Adam optimizer is used to minimize the cross-entropy loss function to achieve iterative updates of the fusion model parameters. After each preset number of iterations, the performance of the fusion model is evaluated on the validation set, and the accuracy and loss value are calculated until the cross-entropy loss function is minimized, thus obtaining the trained fusion model.

[0028] On the other hand, the present invention provides a space target balance pattern intelligent discrimination device, comprising:

[0029] The first module is used to acquire and preprocess the inverse synthetic aperture radar (ISAR) image dataset. The IAR image dataset uses time-series associated IAR image sequences as the basic sample unit. The sample data in the IAR image dataset are stored in sequence folders corresponding to different balance modes according to the corresponding balance modes.

[0030] The second module is used to construct a fusion model of convolutional neural network and long short-term memory network. The fusion model includes a feature extraction module, a temporal modeling module and a classification module. The feature extraction module uses a convolutional neural network to extract spatial features of a single frame image. The temporal modeling module uses a long short-term memory network to capture temporal correlation features of sequences. The classification module outputs the category probability distribution of the spatial target balance pattern.

[0031] The third module is used to train the fusion model based on the preprocessed inverse synthetic aperture radar image dataset;

[0032] The fourth module is used to preprocess the inverse synthetic aperture radar image to be identified and input it into the trained fusion model. The fusion model outputs the balanced mode discrimination result corresponding to the inverse synthetic aperture radar image.

[0033] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0034] The present invention provides a method and apparatus for intelligent discrimination of spatial target balance patterns. By storing sample data in sequence folders corresponding to different balance patterns, the sample data can be matched with subsequent balance pattern recognition processing. By training a fusion model of a convolutional neural network and a long short-term memory network using a preprocessed inverse synthetic aperture radar (INS) image dataset, the invention can discriminate the balance patterns corresponding to INS images. Preprocessing the INS image dataset improves the training effect. The fusion model of the convolutional neural network and the long short-term memory network uses a convolutional neural network to mine spatial features and combines it with a long short-term memory network to improve recognition accuracy and robustness. This enables the present invention to achieve efficient and accurate evaluation of spatial target balance patterns in complex scenarios. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0036] Figure 1A flowchart of a method for intelligently identifying the balance pattern of spatial targets, provided in one embodiment;

[0037] Figure 2 A structural diagram of a fusion model provided in one embodiment;

[0038] Figure 3 A visualization of the training and validation curves of a fusion model provided in one embodiment, wherein, Figure 3 (a) is the validation accuracy curve of the fusion model. Figure 3 (b) is a graph of the loss value of the fusion model. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] Reference Figure 1 One embodiment provides a method for intelligently identifying the balance pattern of space targets, including the following steps:

[0041] The inverse synthetic aperture radar (ISAR) image dataset is acquired and preprocessed. The IAR image dataset uses time-series correlated IAR image sequences as the basic sample unit. The sample data in the IAR image dataset is stored in sequence folders corresponding to different balance modes according to the corresponding balance modes.

[0042] A fusion model of convolutional neural network and long short-term memory network is constructed. The fusion model includes a feature extraction module, a temporal modeling module and a classification module. The feature extraction module uses convolutional neural network to extract spatial features of single frame images. The temporal modeling module uses long short-term memory network to capture temporal correlation features of sequences. The classification module outputs the category probability distribution of spatial target balance patterns.

[0043] The fusion model was trained based on the preprocessed inverse synthetic aperture radar image dataset;

[0044] After preprocessing, the inverse synthetic aperture radar image to be identified is input into the trained fusion model, which outputs the balanced mode discrimination result corresponding to the inverse synthetic aperture radar image.

[0045] In a preferred embodiment, the inverse synthetic aperture radar image dataset is obtained according to the following steps:

[0046] Simulations were performed on publicly available satellite models based on the characteristics of different equilibrium modes to obtain sample data corresponding to different equilibrium modes.

[0047] Construct sequence folders corresponding to different balancing modes, and store sample data into the corresponding sequence folders according to the balancing mode; by recursively traversing the sequence folders, extract the complete path of all sequence folders, use the name of the sequence folder to which the sample data belongs as the label of the sample data, and establish a mapping between sample data path and category label; thus obtaining the inverse synthetic aperture radar image dataset.

[0048] As can be seen, each balanced mode category corresponds to at least one sequence folder, and each sequence folder contains multiple inverse synthetic aperture radar images (i.e., sample data) in a continuous time sequence. When obtaining sample data through simulation, the image acquisition frame rate is set according to the motion characteristics of the space target to ensure that there are observable attitude changes between adjacent frames.

[0049] Because the sequence folder structure strictly follows the three-level structure of "root directory / category / sequence", by establishing a mapping between sample data path and category label, a seamless association between sample data and category label can be achieved through an automatic path resolution mechanism, ensuring the efficiency and scalability of the data loading and training process.

[0050] The balance modes include three-axis stabilization, spin stabilization, and free tumbling. Specifically, three-axis stabilization refers to a state in which a space target can maintain a stable attitude with reference to an independent axis, which in this embodiment is the roll axis, pitch axis, or yaw axis; spin stabilization refers to a state in which a space target maintains a stable attitude through rotation. In this state, the space target continuously rotates around a specific axis (generally a principal axis with a high moment of inertia), using the generated angular momentum to resist external disturbances, thereby maintaining the object in a relatively stable attitude; free tumbling refers to a state in which a space target is not subject to any external control or constraint and rotates freely around its center of gravity. In this state, the attitude change of the space target depends entirely on its internal inertial characteristics and the forces of the external environment.

[0051] The preprocessing includes the following steps:

[0052] Perform consistent cropping on the sample data;

[0053] The cropped sample data is processed to grayscale, compressing the three-channel color image into a single-channel grayscale image.

[0054] By performing consistent cropping on the sample data, we ensure that the cropped sample data contains all feature information and makes the image size input to the fusion model consistent, thus eliminating the difference in the size of sample data of different spatial targets.

[0055] The cropped sample data is then processed to grayscale, compressing the three-channel color image into a single-channel grayscale image, eliminating color information redundancy, while preserving the spatial target shape and texture features.

[0056] The above preprocessing effectively removes redundant content without spatial target features at the edges, eliminates useless information in the sample data, and improves the accuracy of subsequent recognition. By compressing the three-channel color image into a single-channel grayscale image through grayscale processing, the convolution operation during training only needs to be performed on one channel, which greatly reduces the amount of computation, reduces the training time of the fusion model, and speeds up the inference speed of the fusion model.

[0057] Reference Figure 2 In a preferred embodiment, the feature extraction module includes a sequence folding layer, multiple convolutional blocks, and a sequence unrolling layer; the sequence folding layer converts the preprocessed sample data into batch images; the multiple convolutional blocks are cascaded to extract features from the batch images; and the sequence unrolling layer restores the batch image features to a temporal structure. Each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max pooling layer in sequence.

[0058] The sequence folding layer converts preprocessed sample data into batch images to adapt to parallel processing; multiple convolutional blocks are cascaded to extract batch image features, with each convolutional kernel having a predetermined size and number of channels, and padding operations maintaining the consistency of feature map size; each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max pooling layer. The batch normalization layer accelerates model convergence and suppresses overfitting by learning the mean and variance parameters of the data, while the ReLU activation layer introduces nonlinearity, enabling the model to learn complex feature relationships to enhance feature representation capabilities; the max pooling layer achieves feature dimensionality reduction and key feature selection, effectively reducing the number of model parameters; and the sequence unfolding layer restores the batch image features to a temporal structure, providing a feature sequence containing spatial details and inter-frame correlations for subsequent temporal modeling modules.

[0059] The temporal modeling module includes a feature flattening layer, a recurrent neural network layer, and a temporal pooling layer. The feature flattening layer converts the output of the received feature extraction module into a one-dimensional feature vector. The recurrent neural network layer captures the dynamic correlation between frames through a gating mechanism, learns temporal features, and completes the temporal modeling of the one-dimensional feature vector. The temporal pooling layer extracts the overall features of the sequence.

[0060] Temporal pooling layers average the output features across all time steps to extract overall sequence features, reducing temporal dimensionality and enhancing the model's robustness to local perturbations. In this embodiment, the temporal features learned by the recurrent neural network layers are rotation-related temporal features, including periodic contour changes and feature point trajectories.

[0061] The classification module comprises a fully connected layer, a softmax layer, and a classification layer. The fully connected layer maps the output of the temporal modeling module to an output space with a consistent number of balanced pattern categories, achieving a non-linear transformation from the feature space to the category space. The softmax layer normalizes the output values ​​of the fully connected layer into a probability distribution, ensuring that the sum of the probabilities of each category is 1 and that they are mutually exclusive. The classification layer determines the balanced pattern category to which the sample data belongs based on the maximum probability criterion and calculates the difference between the predicted result and the true label using the cross-entropy loss function. Calculating the difference between the predicted result and the true label using the cross-entropy loss function provides an optimization target for subsequent model training.

[0062] The fusion model is trained based on the preprocessed inverse synthetic aperture radar image dataset, including:

[0063] The preprocessed inverse synthetic aperture radar image dataset was divided into a training set and a validation set in an 8:2 ratio, and the distribution of balanced pattern categories in the training set and the validation set was made consistent by stratified sampling and data shuffling.

[0064] The Adam optimizer is used to minimize the cross-entropy loss function to achieve iterative updates of the fusion model parameters. After each preset number of iterations, the performance of the fusion model is evaluated on the validation set, and the accuracy and loss value are calculated until the cross-entropy loss function is minimized, thus obtaining the trained fusion model.

[0065] In one embodiment, the preprocessed sample data is a 300×300 pixel single-channel grayscale image. Experiments have shown that this size preserves the target details in the sample data while avoiding a surge in the computational load of the fusion model due to excessively large image size.

[0066] The sequence folding layer transforms the preprocessed sample data into batch images; specifically, the input image sequence has a dimension of [missing information]. (Height, width, number of channels, time step), converted through sequence folding operation Batch format ( (where the batch size is specified) so that the convolutional neural network can perform parallel feature extraction on each frame of the image. The folding operation can be represented as:

[0067] ;

[0068] in, Indicates the first Frame image, after being folded, forms a size of Batch data.

[0069] Multi-layer convolutional blocks are cascaded to extract batch image features. Specifically, in this embodiment, it includes a first convolutional block and a second convolutional block. The first convolutional block is configured with a 3×3 convolutional kernel and 16 output channels, while the second convolutional block is configured with a 3×3 convolutional kernel and 32 output channels. By setting an all-zero padding strategy, the feature image size after the convolution operation is ensured to remain consistent with the input, avoiding information loss due to dimensionality reduction and thus fully preserving the spatial structural features of the target. Each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max pooling layer. The max pooling layer uses a 2×2 pooling window with a stride of 2 for downsampling, reducing the feature map size to half its original size. By retaining the maximum value in local regions, feature dimensionality reduction and key feature selection are achieved, reducing the number of model parameters while improving the robustness and generalization ability of the features.

[0070] The sequence unrolling layer restores the temporal structure of batch image features. The unrolling operation is the inverse process of the folding operation, which can reduce the dimension to... Batch features reorganized into The temporal feature sequence provides a feature sequence containing spatial details and inter-frame correlations for subsequent temporal modeling modules. The unfolding operation can be represented as follows:

[0071] ;

[0072] in, Indicates the first batch after batch processing Each feature map, when unfolded, is restored to a time step. The corresponding feature sequence.

[0073] The feature flattening layer converts the output of the received feature extraction module into a one-dimensional feature vector and inputs it into the recurrent neural network layer for temporal modeling. The recurrent neural network layer captures the dynamic correlation between frames through a gating mechanism, learns temporal features, and completes the temporal modeling of the one-dimensional feature vector. Specifically, the gating mechanism of the recurrent neural network layer includes a forget gate, an input gate, and an output gate, which can selectively retain long-term dependency information and effectively capture temporal features such as the rotational change characteristics of the rotating target and the motion trajectory of feature points. The recurrent neural network layer is configured with 128 hidden units and set to sequence output mode to output the hidden state of each time step. The forget gate bias is initialized to 1 to enhance long-term memory ability, and the recurrent weights are initialized with a narrow normal distribution to optimize the learning of periodic patterns.

[0074] After completing the one-dimensional feature vector temporal modeling, a temporal pooling layer is used to average the output features of all time steps to extract the overall sequence features, reducing the temporal dimensionality and enhancing the model's robustness to local perturbations; specifically, the output of the recurrent neural network layer... The feature matrix of dimension is calculated by averaging along the time step dimension to obtain... The fixed-length feature vector can be represented as follows:

[0075] ;

[0076] in, For the first The hidden state of the time step. This is the feature vector after global averaging. Through the above processing, the model compresses the temporal dynamic features into a global representation, reducing sensitivity to local noise while preserving the overall pattern information of the sequence.

[0077] The fully connected layer flattens the input feature map into a one-dimensional vector, and maps it to a three-dimensional space (i.e., the space corresponding to the three equilibrium state categories of the target object) through matrix multiplication. The specific process can be expressed as follows:

[0078] ;

[0079] in, This is the vector mapped to three-dimensional space; For flattened one-dimensional vectors, This is the weight matrix. This is the bias vector.

[0080] The Softmax layer normalizes the output values ​​of the fully connected layer into a probability distribution, ensuring that the sum of the probabilities of each class is 1 and that they are mutually exclusive. The specific process can be expressed as follows:

[0081] ;

[0082] in, This is the output value of the fully connected layer.

[0083] The classification layer determines the balanced pattern category to which the sample data belongs based on the maximum probability criterion, and calculates the difference between the predicted result and the true label using the cross-entropy loss function. Based on the cross-entropy loss function, the classification layer adjusts the model parameters through backpropagation to minimize the difference between the prediction and the true distribution. The specific process can be represented as follows:

[0084] ;

[0085] in, For the sample size, For the first The true label of the sample (when the sample belongs to the first) The value is 1 if the condition is met, otherwise it is 0. For the first The probability value predicted by the model for each sample.

[0086] The preprocessed inverse synthetic aperture radar image dataset is divided into a training set and a validation set in an 8:2 ratio. Stratified sampling and data shuffling are used to ensure that the distribution of balanced pattern categories in the training set and the validation set is consistent. Specifically, during the partitioning process, the sample index of each category is randomly shuffled, and then a corresponding number of samples are selected and added to the training set and the validation set respectively.

[0087] Training employs the Adam optimizer to update the parameters of the fusion model. The learning rate uses a dynamic adjustment algorithm with a piecewise decay strategy to balance the convergence speed and accuracy of the fusion model, thereby improving its discrimination accuracy. The learning rate is multiplied by 0.5 every 10 training epochs. The specific process can be represented as follows:

[0088] ;

[0089] in, This is the number of training cycles, with an initial learning rate of 0.0001.

[0090] During training, preprocessed sample data is loaded in batches, with a batch size of 16. Stratified sampling and data shuffling ensure consistent class distribution across the training and validation sets, effectively addressing class imbalance. Every 5 iterations, the fusion model's performance is evaluated on the validation set, and accuracy and loss are calculated to monitor its generalization ability. The maximum training epochs are set to 30. Simultaneously, a training progress curve is plotted in real-time to visually display the trends in loss and accuracy across the training and validation sets; refer to... Figure 3 As can be seen, after 30 rounds of training, with 12 iterations per round, and a total of 360 iterations, both the validation set accuracy and the loss value tend to stabilize. Figure 3 (a) and Figure 3 As shown in (b), the fusion model eventually converges and has a high validation set accuracy of 97.92%. Testing demonstrates that the fusion model described in this invention can efficiently and accurately evaluate the equilibrium patterns of space targets.

[0091] In one embodiment, the present invention provides a space target balance pattern intelligent discrimination device, comprising:

[0092] The first module is used to acquire and preprocess the inverse synthetic aperture radar (ISAR) image dataset. The IAR image dataset uses time-series associated IAR image sequences as the basic sample unit. The sample data in the IAR image dataset are stored in sequence folders corresponding to different balance modes according to the corresponding balance modes.

[0093] The second module is used to construct a fusion model of convolutional neural network and long short-term memory network. The fusion model includes a feature extraction module, a temporal modeling module and a classification module. The feature extraction module uses a convolutional neural network to extract spatial features of a single frame image. The temporal modeling module uses a long short-term memory network to capture temporal correlation features of sequences. The classification module outputs the category probability distribution of the spatial target balance pattern.

[0094] The third module is used to train the fusion model based on the preprocessed inverse synthetic aperture radar image dataset;

[0095] The fourth module is used to preprocess the inverse synthetic aperture radar image to be identified and input it into the trained fusion model. The fusion model outputs the balanced mode discrimination result corresponding to the inverse synthetic aperture radar image.

[0096] Matters not covered in this invention are common knowledge.

[0097] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligently identifying the balance pattern of space targets, characterized in that, Includes the following steps: The inverse synthetic aperture radar (ISAR) image dataset is acquired and preprocessed. The IAR image dataset uses time-series correlated IAR image sequences as the basic sample unit. The sample data in the IAR image dataset is stored in sequence folders corresponding to different balance modes according to the corresponding balance modes. A fusion model of convolutional neural network and long short-term memory network is constructed. The fusion model includes a feature extraction module, a temporal modeling module and a classification module. The feature extraction module uses convolutional neural network to extract spatial features of single frame images. The temporal modeling module uses long short-term memory network to capture temporal correlation features of sequences. The classification module outputs the category probability distribution of spatial target balance patterns. The feature extraction module includes a sequence folding layer, a multi-layer convolutional block, and a sequence unrolling layer. The sequence folding layer converts the preprocessed sample data into batch images. The multi-layer convolutional block is cascaded to extract features from the batch images. The sequence unrolling layer restores the batch image features to a temporal structure. Each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max pooling layer. The temporal modeling module includes a feature flattening layer, a recurrent neural network layer, and a temporal pooling layer. The feature flattening layer converts the output of the received feature extraction module into a one-dimensional feature vector. The recurrent neural network layer captures the dynamic correlation between frames through a gating mechanism, learns temporal features, and completes the temporal modeling of the one-dimensional feature vector. The temporal pooling layer extracts the overall features of the sequence. The fusion model was trained based on the preprocessed inverse synthetic aperture radar image dataset; After preprocessing, the inverse synthetic aperture radar image to be identified is input into the trained fusion model, which outputs the balanced mode discrimination result corresponding to the inverse synthetic aperture radar image.

2. The intelligent discrimination method for space target balance patterns as described in claim 1, characterized in that, The inverse synthetic aperture radar image dataset is obtained according to the following steps: Simulations were performed on publicly available satellite models based on the characteristics of different equilibrium modes to obtain sample data corresponding to different equilibrium modes. Construct sequence folders corresponding to different balance modes, and store sample data into the corresponding sequence folders according to the balance mode; By recursively traversing the sequence folders, the complete paths of all sequence folders are extracted, and the names of the sequence folders to which the sample data belongs are used as the labels of the sample data to establish a mapping between sample data paths and category labels; thus, the inverse synthetic aperture radar image dataset is obtained.

3. The intelligent discrimination method for space target balance patterns as described in claim 2, characterized in that, The balance modes include triaxial stabilization, spin stabilization, and free roll.

4. The intelligent discrimination method for space target balance patterns as described in claim 1, characterized in that, The preprocessing includes the following steps: Perform consistent cropping on the sample data; The cropped sample data is processed to grayscale, compressing the three-channel color image into a single-channel grayscale image.

5. The intelligent discrimination method for space target balance patterns as described in claim 1, characterized in that, The classification module includes a fully connected layer, a softmax layer, and a classification layer. The fully connected layer maps the output of the temporal modeling module to an output space with a consistent number of spatial target balance mode categories. The softmax layer normalizes the output values ​​of the fully connected layer into a probability distribution. The classification layer determines the balance mode category to which the sample data belongs based on the maximum probability criterion and calculates the difference between the prediction result and the true label using the cross-entropy loss function.

6. The intelligent discrimination method for space target balance patterns as described in claim 5, characterized in that, The fusion model was trained based on the preprocessed inverse synthetic aperture radar image dataset, including: The preprocessed inverse synthetic aperture radar image dataset was divided into a training set and a validation set in an 8:2 ratio, and the distribution of balanced pattern categories in the training set and the validation set was made consistent by stratified sampling and data shuffling. The Adam optimizer is used to minimize the cross-entropy loss function to achieve iterative updates of the fusion model parameters. After each preset number of iterations, the performance of the fusion model is evaluated on the validation set, and the accuracy and loss value are calculated until the cross-entropy loss function is minimized, thus obtaining the trained fusion model.

7. A space target balance mode intelligent discrimination device, characterized in that, include: The first module is used to acquire and preprocess the inverse synthetic aperture radar (ISAR) image dataset. The IAR image dataset uses time-series associated IAR image sequences as the basic sample unit. The sample data in the IAR image dataset are stored in sequence folders corresponding to different balance modes according to the corresponding balance modes. The second module is used to construct a fusion model of convolutional neural networks and long short-term memory networks. This fusion model includes a feature extraction module, a temporal modeling module, and a classification module. The feature extraction module uses a convolutional neural network to extract spatial features from a single frame image. The temporal modeling module uses a long short-term memory network to capture temporal correlation features of sequences. The classification module outputs the class probability distribution of spatial target balance patterns. The feature extraction module includes a sequence folding layer, multiple convolutional blocks, and a sequence unrolling layer. The sequence folding layer converts preprocessed sample data into batch images. Multiple convolutional blocks are cascaded to extract batch image features. The sequence unrolling layer restores the batch image features to a temporal structure. Each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max-pooling layer. The temporal modeling module includes a feature flattening layer, a recurrent neural network layer, and a temporal pooling layer. The feature flattening layer converts the output of the received feature extraction module into a one-dimensional feature vector. The recurrent neural network layer captures the dynamic correlation between frames through a gating mechanism, learns temporal features, and completes the temporal modeling of the one-dimensional feature vector. The temporal pooling layer extracts the overall features of the sequence. The third module is used to train the fusion model based on the preprocessed inverse synthetic aperture radar image dataset; The fourth module is used to preprocess the inverse synthetic aperture radar image to be identified and input it into the trained fusion model. The fusion model outputs the balanced mode discrimination result corresponding to the inverse synthetic aperture radar image.

Images