Spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation

Abstract

The invention discloses a spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation, and relates to the technical field of image processing, and the method comprises the steps: obtaining a to-be-detected ISAR image sequence; grouping a to-be-detected ISAR image sequence according to adjacent frames, dividing the to-be-detected ISAR image sequence into image pair samples, and marking serial numbers for the image pair samples; according to a forward sequence, inputting the image pair samples into the trained multi-scale attention network for rotation parameter estimation to obtain a forward predicted angular sequence; according to a reverse sequence, inputting the image pair samples into the trained multi-scale attention network for rotation parameter estimation to obtain a reverse predicted angular sequence; and according to the forward predicted angle sequence and the reverse predicted angle sequence, calculating an anomaly score of each frame of the to-be-detected ISAR image sequence, comparing the anomaly score with a detection threshold, and taking the to-be-detected ISAR image frame exceeding the detection threshold as an anomaly point to realize spacecraft attitude anomaly detection. According to the invention, wider application scenes can be realized.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for detecting spacecraft attitude anomalies based on ISAR image sequence rotation parameter estimation. Background Technology

[0002] Spacecraft attitude anomaly detection can sense abnormal changes in the on-orbit operational status of spacecraft, which is of great significance for ensuring the safety of space assets. For a spacecraft operating normally in orbit, its attitude is generally stable relative to itself along three axes; however, when a system failure occurs, its attitude will rotate abnormally under the influence of gravity, requiring timely detection and rescue. For a failed spacecraft, its attitude is usually in a slow rotational state due to the influence of gravity gradient, but abnormal changes in rotation caused by debris collisions can affect its orbit, posing a significant threat to the operational safety of a normally operating spacecraft. Therefore, spacecraft attitude anomaly detection based on rotation parameter estimation is an effective means of monitoring the operational status of spacecraft, and is of great significance for the detection and diagnosis of spacecraft faults and early warning of collisions with runaway spacecraft.

[0003] Existing methods for spacecraft attitude anomaly detection are mainly divided into radar echo feature-based methods and image feature-based methods. The first type of method relies on typical target features of radar echoes, such as radar cross-section (RCS) and high-resolution range profile (HRRP), to perform parameter estimation, anomaly classification, or database matching to achieve attitude anomaly detection; however, the target feature information provided by radar echoes is limited, and it is greatly affected by noise, which severely limits its application in non-cooperative monitoring environments without auxiliary information. The second category of methods mostly focuses on solving the problem of instantaneous attitude inversion and binary classification of anomalous attitude states of space targets. Instantaneous attitude inversion of space targets can be achieved through simultaneous observation and imaging by multiple sensors; however, the implementation cost is high. Binary classification of anomalous attitude states of targets can be achieved by extracting and analyzing key points or linear feature flows from ISAR image sequences, i.e., identifying whether an anomalous rotation state has occurred when the target attitude is already in a normal three-axis stable state. In the latest anomaly classification schemes, researchers use a Stacked Sparse Autoencoder (SSAE) network to extract the target's geometric feature flow, and then input it into a Long-Short Time Memory (LSTM) network for sequence classification, transforming attitude anomaly detection into an ordered binary classification problem. However, existing techniques fail when facing targets in the attitude adjustment process or uncontrolled slow rotation state. Furthermore, existing binary classification schemes also face the problem of imbalanced datasets because anomalous samples are difficult to obtain in practical applications. In addition, extracting absolute geometric features of targets from ISAR images with strong sparsity, severe size distortion, and significant anisotropy is usually difficult and unrobust.

[0004] Therefore, there is an urgent need to provide a method for detecting abnormal spacecraft attitude in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation, comprising: Acquire the ISAR image sequence to be tested; The ISAR image sequence to be tested is grouped into image pairs by adjacent frames, and the image pairs are labeled with serial numbers. In the forward order, image samples are input into a trained multi-scale attention network to estimate rotation parameters, resulting in a forward-predicted angle sequence; in the reverse order, image samples are input into a trained multi-scale attention network to estimate rotation parameters, resulting in a reverse-predicted angle sequence. Based on the forward and backward predicted angle sequences, the anomaly score of each frame of the ISAR image sequence to be tested is calculated and compared with the detection threshold. The ISAR image frames to be tested that exceed the detection threshold are regarded as anomalies, so as to realize the detection of spacecraft attitude anomalies.

[0006] The beneficial effects of this invention are: This invention provides a spacecraft attitude anomaly detection method based on rotation parameter estimation of ISAR image sequences. Compared to existing technologies, this method delves deeper into the essence of target attitude changes, enabling the identification of general attitude anomalies rather than being limited to anomaly classification under specific attitudes. Furthermore, the multi-scale attention network proposed in this invention can perform angle regression tasks using only normal data for training, and can directly set anomaly detection thresholds based on normal sequences, without relying on the acquisition of anomaly samples. Moreover, the key to angle regression using the multi-scale attention network lies in extracting the relative difference features between two images. Compared to extracting the absolute geometric features of each image, the method proposed in this invention is more robust and easier to implement when dealing with ISAR images characterized by high sparsity, severe size distortion, and significant anisotropy.

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0008] Figure 1 This is a flowchart of a spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-scale attention network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a partial sample image of the constructed dataset provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of an embodiment of the present invention that utilizes rotation angle to achieve posture anomaly detection; Figure 5 This is a schematic diagram of the results of spacecraft attitude anomaly detection provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0010] Inverse Synthetic Aperture Radar (ISAR) has become an important means of space target surveillance due to its all-weather, all-time, high-resolution, and long-range imaging advantages. During ISAR observation of spacecraft, the relative motion between the spacecraft and the radar line of sight (LOS) causes continuous attitude rotation. Consequently, there is relative rotation between the spacecraft's ISAR image sequences. Whether the spacecraft is in normal orbit or has failed, once an attitude anomaly occurs, the continuity of the relative rotational motion is broken. At this point, analyzing the continuity of rotation parameters from the ISAR image sequence can enable attitude anomaly detection.

[0011] To address the aforementioned shortcomings, this invention proposes a spacecraft attitude anomaly detection method based on rotation parameter estimation from ISAR image sequences. Leveraging the continuity of rotation parameters during smooth attitude changes, this method can more effectively identify attitude anomalies, going beyond simply classifying anomalies under specific attitudes. Furthermore, a Multiscale Attention Network (MAN) is constructed, enabling the extraction of more robust features. The method can perform rotation parameter regression using only normal data for training and allows for setting anomaly detection thresholds directly based on normal sequences, independent of the acquisition of anomalous samples.

[0012] Please see Figure 1 , Figure 1 This is a flowchart of a spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation provided by an embodiment of the present invention. The spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation provided by the present invention includes: S101. Obtain the ISAR image sequence to be tested.

[0013] Specifically, in this embodiment, an ISAR image sequence is simulated when a spacecraft undergoes abnormal rotation during a period of stable attitude change. An initial target rotation vector around the axis is set to simulate a stable attitude change, and then the rotation vector is changed at certain moments in subsequent observations to simulate abnormal rotation changes.

[0014] S102. Group the ISAR image sequence to be tested into image pairs by adjacent frames, and mark the sequence number of each image pair.

[0015] Specifically, in this embodiment, the length is ISAR image sequence to be tested The samples are divided into multiple image pairs based on adjacent frames. Each image pair sample is labeled with its image pair sample number according to the image number of its first frame, for example... , ...

[0016] S103. In the forward order, the image samples are input into the trained multi-scale attention network to estimate the rotation parameters, and the forward predicted angle sequence is obtained; in the reverse order, the image samples are input into the trained multi-scale attention network to estimate the rotation parameters, and the reverse predicted angle sequence is obtained.

[0017] Specifically, in this embodiment, each image pair sample is input into a trained multi-scale attention network in forward order according to its sequence number to estimate rotation parameters and obtain a forward predicted angle sequence. The rotation parameters are estimated by inputting the data into a pre-trained multi-scale attention network in reverse order of their sequence numbers, resulting in the reverse-predicted angle sequence. The predicted corner sequence index corresponds to the original frame order of the ISAR image sequence to be tested.

[0018] In this embodiment, please refer to Figure 2 , Figure 2 This is a schematic diagram of a multi-scale attention network provided in an embodiment of the present invention. The trained multi-scale attention network is obtained by training an initial multi-scale attention network using data of a preset category as the training dataset. The multi-scale attention network includes a first concatenation layer, a feature extractor, and a multilayer perceptron (MLP). The images are preprocessed by converting the input image pairs to grayscale and cropping them to a suitable input dimension for the network. ; The first stitching layer is used to stitch the image pairs along the channel dimension, resulting in dual-channel features after shallow feature fusion. , represented as: ; in, and This represents the image pair input to the first stitching layer. Indicates a splicing operation; The feature extractor is used to process the dual-channel features fused from shallow features to obtain enhanced features with 256 channels and a fixed size of 4×4. , represented as: ; in, This indicates a feature extraction operation. The dual-channel input aims to better measure the rotational changes between two frames, driving the model to focus on learning the relative difference features of the input images. The enhanced features are flattened to obtain flattened features. , represented as:

[0019] in, Indicates the flattening operation; Multilayer perceptrons are used to perform progressive dimensionality reduction regression on the flattened features to obtain the predicted rotation angle values. ; ; in, This represents a multi-layer sensing operation.

[0020] Furthermore, the feature extractor includes an initial convolution (IC), a first multiscale convolutional attention block (MCAB), a first max pooling layer, a second multiscale convolutional attention block, a second max pooling layer, a third multiscale convolutional attention block, and an average pooling layer. The initial convolutional module includes a convolutional layer, a batch normalization (BN) layer, and a ReLU activation function. The kernel size of the convolutional layer is 3×3. The batch normalization layer is used to accelerate model convergence and alleviate gradient vanishing. The ReLU activation function can introduce nonlinear transformations to improve the model's generalization ability. The initial convolutional module is used to extract edge, texture, or corner features from the dual-channel features after shallow feature fusion. , represented as: ; in, Indicates the initial convolutional module operation; The first multi-scale convolutional attention block is used to process edge, texture, or corner features to obtain the first feature; The first max-pooling layer is used to downsample the first feature to obtain the second feature. , represented as: ; in, This indicates a max pooling operation. This represents the attention block operation in multi-scale convolution; The second multi-scale convolutional attention block is used to process the second feature to obtain the third feature; The second max-pooling layer is used to downsample the third feature to obtain the fourth feature. , represented as: ; The third multi-scale convolutional attention block is used to process the fourth feature to obtain the fifth feature; The average pooling layer is used to downsample the fifth feature to obtain the sixth feature. , represented as: .

[0021] in, This represents the average pooling operation, which is used to pool the feature map to a fixed 4×4 size, improving efficiency, robustness and generalization ability through dimensionality reduction. Furthermore, considering the high sparsity of ISAR images, varying target sizes, and lateral scaling distortion, this invention proposes MCAB. MCAB uses four parallel convolutional branches for feature extraction, employing convolutional kernels of sizes 1×1, 3×3, 5×5, and 7×7, respectively, aiming to capture features from different receptive fields more comprehensively. Like the initial convolutional layer, each branch also includes batch normalization and a ReLU activation function.

[0022] However, for network training, the importance of features varies across different scales and regions, requiring a focus on more useful parts to improve training efficiency. Attention mechanisms are widely used in deep learning, mimicking the human brain's attention mechanism to help the network focus on more useful features. In this invention, MCAB further integrates Channel Attention (CA) and Spatial Attention (SA) mechanisms, naturally complementing multi-scale convolution. This allows the network to dynamically adjust the importance of features from both channel and spatial dimensions, enhancing the model's representational ability and performance.

[0023] The first, second, and third multi-scale convolutional attention blocks have the same structure, each including a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, a second splicing layer, a channel attention module, and a spatial attention module; among them, The first convolutional module processes the input features to obtain the first sub-feature; the second convolutional module processes the input features to obtain the second sub-feature; the third convolutional module processes the input features to obtain the third sub-feature; and the fourth convolutional module processes the input features to obtain the fourth sub-feature. The second splicing layer is used to splice the first sub-feature, the second sub-feature, the third sub-feature, and the fourth sub-feature to obtain spliced ​​sub-features; The channel attention module is used to process the spliced ​​sub-features, enhancing important channels and suppressing minor channels to obtain enhanced channel features. , represented as: ; ; in, Indicates splicing sub-features, This represents the output weight vector. This represents channel-level multiplication. Indicates the weights of the dimensionality reduction layer. Indicates the compression ratio. This represents the channel statistics descriptor generated after global average pooling of the spliced ​​sub-features. Represents the ReLU activation function. This represents the weights of the dimensionality-upgrading layer; through the channel attention module, a bottleneck mechanism is formed to first reduce the dimensionality to... Restore dimensions to enhance nonlinear modeling capabilities; The spatial attention module is used to process the enhanced channel features to obtain enhanced spatial features. , represented as: ; ; in, Representing a spatial attention map, This represents element-wise multiplication. This represents the Sigmoid activation function. This represents the features generated by max pooling the enhanced channel features. This indicates the features generated by average pooling of the enhanced channel features. This represents a convolution operation with a 7×7 kernel. This indicates a splicing operation. Through the spatial attention module, the network can adaptively focus on key target regions of the image, significantly improving feature discrimination in complex backgrounds.

[0024] Furthermore, the first, second, third, and fourth convolutional modules have the same structure, all including convolutional layers, batch normalization layers, and ReLU activation functions; among them, The kernel size of the convolutional layer in the first convolutional module is 1×1, the kernel size of the convolutional layer in the second convolutional module is 3×3, the kernel size of the convolutional layer in the third convolutional module is 5×5, and the kernel size of the convolutional layer in the fourth convolutional module is 7×7.

[0025] It should be noted that MCAB uses multiple convolutional kernels with different receptive fields in parallel and applies attention mechanisms sequentially for feature fusion, which can simultaneously focus on the macroscopic structure and local details of the image, enhancing the model's robustness to changes in target size and viewpoint while maintaining network depth.

[0026] In this embodiment, training the initial multi-scale attention network includes: Obtain a dataset of ISAR image pairs with random rotation labels. The dataset includes multiple samples, each sample corresponding to an ISAR image pair. The rotation angle of the ISAR image pair is labeled, and the true labels corresponding to the ISAR image pairs are obtained. The dataset is divided into training, validation, and test datasets according to a preset ratio; optionally, the preset ratio is 8:1:1. Input a portion of the samples from the training dataset into the first... The multi-scale attention network to be trained is then trained to obtain the first... The predicted rotation angle value output during this training process; According to the The predicted rotation angle value output during the training process is the same as the value output during the training process. The true labels of the samples to be trained in the multi-scale attention network are used to calculate the mean squared error (MSE) and used as the basis for the training. Loss during each training session; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the multi-scale attention network to be trained are obtained. The multi-scale attention network to be trained is used; at the same time, at each preset validation interval, the samples in the validation dataset are input into the multi-scale attention network to be trained to obtain the output predicted rotation angle value. Combined with the true labels of the samples in the validation dataset, model selection and hyperparameter tuning are performed. This process is iterated until the number of training times or the degree of convergence meets the preset conditions, and the trained multi-scale attention network is obtained. The samples from the test dataset were input into the trained multi-scale attention network for evaluation to assess its performance; as shown in Table 1, the RMSE was within 0.26 and the MRE was within 9%. The value reached over 0.96; the test results show that the model has good performance and can robustly and accurately estimate the rotation angle.

[0027] Table 1 Evaluation results of the MAN model on the test set

[0028] Further, please see Figure 3 , Figure 3 This is a schematic diagram of a partial sample image of the constructed dataset provided in an embodiment of the present invention, which obtains a dataset of ISAR images with random corner labels, including: Multiple observations of the spacecraft were conducted to obtain the observation arc segment. Optionally, the motion of the spacecraft in orbit can be uniformly modeled as rotational motion around a fixed axis. To ensure the generalization of the network, multiple initial attitude angles and rotation vectors around the axis of the spacecraft need to be set in the centroid orbital coordinate system, and a total of 27 observations under different attitudes are conducted. For each observation arc segment, several imaging sub-aperture arc segments are randomly divided, and the time interval between the division of adjacent sub-apertures is required. The target is uniformly distributed between [0, 10]. With a fixed angular velocity of 0.01 rad / s, the angle through which the target rotates on the imaging plane is observed. They are roughly randomly distributed within the range of [0.2, 6.4]°; Imaging simulations were performed on several imaging sub-aperture arcs corresponding to the observed arc segment to obtain several ISAR image sequences; for example, for an arc segment of length... IASR image sequence Take adjacent frames as a sample, for example, frames 1 and 2 as the first set of samples. Frames 3 and 4 are used as the second group of samples. Similarly, the images in each group of samples are sorted according to the order of imaging, and the images in each group of samples are not reused. According to the above method, each observation can obtain 250 independent samples from 500 frames of ISAR images. With 27 repeated observations, a total of 6750 independent samples can be obtained. For each observation arc segment, a sequence of several ISAR image frames is obtained. Two adjacent frames are extracted as samples to construct the dataset. The dataset constructed in this embodiment fully considers the diversity of imaging viewpoints and rotation angles. The samples in the dataset are labeled, including marking the rotation angle of adjacent ISAR images in each group of samples. and imaging time interval This yields the true labels corresponding to the samples.

[0029] S104. Based on the forward-predicted angle sequence and the backward-predicted angle sequence, calculate the anomaly score for each frame of the ISAR image sequence to be tested, and compare it with the detection threshold. ISAR image frames exceeding the detection threshold are identified as anomalies to achieve spacecraft attitude anomaly detection. (See also...) Figure 4 and Figure 5 , Figure 4 This is a schematic diagram illustrating how to detect posture anomalies using rotation angles, as provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the results of spacecraft attitude anomaly detection provided in an embodiment of the present invention.

[0030] Specifically, in this embodiment, the anomaly score of each frame of the ISAR image sequence to be tested is calculated. The expression is: ; in, This indicates calculating the absolute value of the data within the parentheses. Indicates the first Frame and the Rotation angle between frames express Frame and the Rotation angle between frames This indicates the length of the ISAR image sequence to be tested.

[0031] For typical spacecraft with stable attitude changes, and The difference is very small, at which point the outlier score is considered normal. This will approach a stable straight line, corresponding to a smooth attitude change process for the spacecraft. However, when the spacecraft's attitude exhibits abnormal, unexpected changes in rotational speed, assuming from the... When the frame image begins to show anomalies, the predicted... and The mutation, calculated from this In the Frame and the The appearance of a peak at a frame corresponds to an abnormal change in the spacecraft's attitude, thus qualitatively realizing the detection of attitude anomalies under stable attitude changes.

[0032] Furthermore, the pose anomaly detection threshold is calculated from the anomaly score of the normal sequence, expressed as: ; in, This indicates that the average of the data within the parentheses is calculated. This indicates the standard deviation of the data within the parentheses.

[0033] By the sequence being tested With threshold By comparison, frame numbers exceeding the threshold are marked as anomalies. The anomaly detection error is controlled within one frame before and after, thus achieving dynamic, high-precision, and automated detection of spacecraft attitude anomalies based solely on normal data, as shown in Table 2.

[0034] Table 2 Specific Results of Spacecraft Attitude Anomaly Detection

[0035] In summary, the spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation provided by this invention has the following beneficial effects: First, construct an ISAR image pair dataset with random corner labels.

[0036] This invention performs multiple observations of the spacecraft in various attitudes and randomly divides the imaging sub-apertures. Then, for each observation arc segment, it randomly divides the imaging frame into several sub-aperture arc segments, requiring that the time interval between the division of adjacent sub-apertures be... The targets are evenly distributed within a reasonable range to ensure that the angle through which they rotate on the imaging plane is observed. The images are randomly distributed within a reasonable range. Finally, for each ISAR image sequence obtained from each observation, adjacent frames are extracted as independent dataset samples.

[0037] Second, a method for detecting spacecraft attitude anomalies based on rotation parameter estimation of ISAR image sequences.

[0038] This invention transforms the problem of attitude anomaly detection into the problem of ISAR image sequence rotation parameter estimation. Furthermore, by leveraging the continuity of rotation parameters during stable attitude changes, it differs from existing techniques in that it can identify anomalies in targets undergoing general stable attitude changes, rather than being limited to identifying rotational anomalies in specific three-axis stable targets. Moreover, this invention uses only normal data for training and directly sets the anomaly detection threshold based on normal sequences, without relying on the acquisition of anomaly samples.

[0039] Third, MAN network design.

[0040] This invention proposes a Multiscale Attention Network (MAN) for estimating rotation parameters of ISAR image sequences characterized by high sparsity, severe scaling distortion, and significant anisotropy. The core design of the network lies in the Multiscale Convolutional Attention Block (MCAB) in the feature extractor. Considering the characteristics of ISAR images, this invention improves upon existing methods by proposing the MCAB and replacing the traditional three convolutional layers with three cascaded MCABs in the feature extractor. The MCAB, by using multiple convolutional kernels with different receptive fields in parallel and sequentially applying attention mechanisms for feature fusion, can simultaneously focus on the macroscopic structure and local details of the image, enhancing the model's robustness to weak texture features, target size distortion, and viewpoint changes while maintaining network depth.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0042] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0043] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A spacecraft attitude anomaly detection method based on ISAR image sequence rotation parameter estimation, characterized in that, The method comprises the following steps: acquiring a to-be-tested ISAR image sequence; grouping the to-be-tested ISAR image sequence according to adjacent frames, dividing the image pair samples, and marking the sequence number of the image pair samples; inputting the image pair samples into a trained multi-scale attention network in a forward order to perform rotation parameter estimation, and obtaining a forward predicted angle sequence; inputting the image pair samples into the trained multi-scale attention network in a reverse order to perform rotation parameter estimation, and obtaining a reverse predicted angle sequence; calculating an anomaly score of each frame of the to-be-tested ISAR image sequence according to the forward predicted angle sequence and the reverse predicted angle sequence, and comparing the anomaly score with a detection threshold, so that the to-be-tested ISAR image frame exceeding the detection threshold is regarded as an anomaly point, thereby achieving spacecraft attitude anomaly detection.

2. The method of claim 1, wherein the ISAR image sequence rotation parameter based spacecraft attitude anomaly detection method is characterized by, calculating an anomaly score for each frame of the sequence of ISAR images under test The expression is: ； wherein, denotes the calculation of the absolute value, denotes the rotation angle between the frame and the frame image, denotes the frame and the frame image, denotes the length of the ISAR image sequence to be measured.

3. The method of claim 2, wherein the ISAR image sequence rotation parameter estimation based spacecraft attitude anomaly detection method is characterized by, computing the detection threshold The expression for the detection threshold is ； wherein, represents calculating the mean value, represents calculating the standard deviation.

4. The method of claim 1, wherein the ISAR image sequence rotation parameter based spacecraft attitude anomaly detection method is characterized by, The trained multi-scale attention network is obtained by training an initial multi-scale attention network using preset category data as a training data set; the multi-scale attention network comprises a first splicing layer, a feature extractor, and a multi-layer perception; wherein The first splicing layer is configured to splice the image pairs in the channel dimension to obtain the double-channel features after shallow feature fusion is represented as: ； wherein, and denotes a pair of images input to the first stitching layer, denotes a stitching operation; The feature extractor is configured to process the fused two-channel features of the shallow features to obtain enhanced features , is expressed as: ； wherein denotes a feature extraction operation; flattening the enhanced features to obtain flattened features is represented as: wherein represents a flattening operation; The multi-layer perception is used for performing progressive dimension reduction regression processing on the flattened features to obtain a predicted rotation angle value ; ； wherein denotes a multi-layer perception operation.

5. The method of claim 4, wherein the ISAR image sequence rotation parameter estimation based spacecraft attitude anomaly detection method is characterized by, The feature extractor comprises an initial convolution module, a first multi-scale convolution attention block, a first maximum pooling layer, a second multi-scale convolution attention block, a second maximum pooling layer, a third multi-scale convolution attention block, and an average pooling layer; the initial convolution module comprises a convolution layer, a batch normalization layer, and a ReLU activation function; wherein The initial convolution module is configured to extract edge, texture or corner point features of the fused double-channel features of the shallow layer features is expressed as: ； wherein, denotes an initial convolutional module operation; The first multi-scale convolution attention block is used for processing the edge, texture, or corner features to obtain first features; The first max-pooling layer is configured to down-sample the first feature to obtain a second feature is represented as: ； wherein, denotes a max-pooling operation, denotes a multi-scale convolutional attention block operation; The second multi-scale convolution attention block is used for processing the second features to obtain third features; The second max-pooling layer is configured to down-sample the third feature to obtain a fourth feature is represented as: ； The third multi-scale convolution attention block is used for processing the fourth features to obtain fifth features; The average pooling layer is configured to down-sample the fifth feature to obtain a sixth feature is represented as: ； wherein, denotes an average pooling operation.

6. The method of claim 5, wherein the ISAR image sequence rotation parameter based spacecraft attitude anomaly detection method is characterized by, The first multi-scale convolution attention block, the second multi-scale convolution attention block, and the third multi-scale convolution attention block have the same structure, and each comprises a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a second splicing layer, a channel attention module, and a spatial attention module; wherein The first convolution module is used for processing input features to obtain first sub-features; the second convolution module is used for processing input features to obtain second sub-features; the third convolution module is used for processing input features to obtain third sub-features; and the fourth convolution module is used for processing input features to obtain fourth sub-features; The second splicing layer is used for splicing the first sub-features, the second sub-features, the third sub-features, and the fourth sub-features to obtain spliced sub-features; The channel attention module is used for processing the splicing sub-feature to obtain an enhanced channel feature is expressed as: ； ； wherein, represents a concatenation sub-feature, represents an output weight vector, represents a channel-level multiplication, represents a dimension reduction layer weight, represents a channel statistical descriptor generated after global average pooling of the concatenation sub-feature, represents a ReLU activation function, represents a dimension increase layer weight; The spatial attention module is used for processing the enhanced channel features to obtain enhanced spatial features , is expressed as: ； ； wherein, denotes a spatial attention map, denotes an element-wise multiplication, denotes a Sigmoid activation function, denotes a feature generated by max-pooling on the enhanced channel features, denotes a feature generated by average-pooling on the enhanced channel features, denotes a convolution operation, denotes a concatenation operation.

7. The method of claim 6, wherein the ISAR image sequence rotation parameter estimation based spacecraft attitude anomaly detection method is characterized by, The first convolution module, the second convolution module, the third convolution module, and the fourth convolution module have the same structure, and each comprises a convolution layer, a batch normalization layer, and a ReLU activation function; wherein The convolution kernel size of the convolution layer in the first convolution module is 1x1, the convolution kernel size of the convolution layer in the second convolution module is 3x3, the convolution kernel size of the convolution layer in the third convolution module is 5x5, and the convolution kernel size of the convolution layer in the fourth convolution module is 7x7.

8. The method of claim 4, wherein the ISAR image sequence rotation parameter based spacecraft attitude anomaly detection method is characterized by, The training of the initial multi-scale attention network comprises the following steps: An ISAR image pair data set with random rotation angle labels is acquired, the data set including a plurality of samples, each sample corresponding to an ISAR image pair, and the rotation angle of the ISAR image pair being labeled, and a real label corresponding to the ISAR image pair is acquired; The data set is divided into a training data set, a verification data set and a test data set according to a preset ratio; inputting part of samples in the training data set into a first scale attention network to be trained to obtain a first predicted rotation angle value output in the first training process; According to the first The predicted rotation angle value output in the second training process is compared with the real label of the sample of the multi-scale attention network to be trained, the mean square error is calculated, and the mean square error is taken as the loss of the first The second training process. The second training process. According to the first loss of the second training process is back propagated to update the network parameters of the second multi-scale attention network to be trained, to obtain the second multi-scale attention network to be trained; meanwhile, every interval of a preset verification interval value, the samples in the verification data set are input into the multi-scale attention network to be trained, to obtain the output predicted rotation angle value, combined with the real label of the samples in the verification data set, model selection and hyperparameter tuning are performed, and the iteration is performed until the training times or the convergence degree meets the preset condition, to obtain the trained multi-scale attention network. The samples in the test data set are input into the trained multi-scale attention network for evaluation to evaluate the performance of the trained multi-scale attention network.

9. The method of claim 8, wherein the ISAR image sequence rotation parameter estimation based spacecraft attitude anomaly detection method is characterized by, An ISAR image pair data set with random rotation angle labels is acquired, including: A spacecraft is observed multiple times to obtain observation arcs; For each observation obtained observation arc, randomly divided into several imaging sub-aperture arc, and adjacent sub-aperture partition time interval Uniformly distributed between [0, 10]; A plurality of imaging sub-aperture arcs corresponding to the observation arcs are imaged to obtain a plurality of ISAR image sequences; For each observation obtained by the observation arc segment corresponding to a number of frame ISAR image sequence , take out two adjacent frames as samples to construct a data set ; annotating the samples in the dataset, including marking the rotation angle of adjacent frame ISAR images in each group of samples and imaging time interval to obtain the real label corresponding to the sample.

10. The method of claim 1, wherein the ISAR image sequence rotation parameter based spacecraft attitude anomaly detection method is characterized by, An ISAR image sequence to be measured is acquired, including: An ISAR image sequence of a spacecraft with a smooth change in attitude when an abnormal rotation occurs is simulated.

Images