Satellite photometric curve data preprocessing and classification method

Abstract

This invention provides a method for preprocessing and classifying satellite photometric curve data. The method includes: converting phase angle-magnitude curves with discontinuous observations and uneven sampling frequencies into fixed-dimensional window function feature vectors using a frame-segmentation and windowing preprocessing strategy; employing a random erasure data augmentation strategy on these window function feature vectors to generate different photometric curve feature vectors through multiple random erasures while simulating phase observation gaps; performing positional encoding of the phase angle features on the photometric curve feature vectors to obtain photometric features containing phase information; constructing query vectors, index vectors, and key value vectors using a transformer model to extract global features; and classifying the photometric curve feature vectors using a fully connected layer. This method effectively captures both local and global feature information of the photometric curves, exhibiting superior performance compared to one-dimensional convolutional neural networks.

Description

Technical Field

[0001] This invention belongs to the field of satellite photometric observation data processing and machine learning, specifically involving a method for preprocessing and classifying satellite photometric curve data. Background Technology

[0002] Due to the physical limitations of optical imaging sensors, such as wide viewing angles and long line-of-sight distances, and interference from complex ambient light, targets tracked and captured by space situational awareness systems often exhibit low signal-to-noise ratio (SNR) point-like characteristics. Traditional target recognition algorithms based on shape and color features struggle to achieve good performance. Classifying satellite photometric curve data is an important and meaningful task in practice, with significant application value for satellite fault diagnosis and model analysis. Currently, satellite photometric curve classification is mainly done manually, which requires extensive professional knowledge and is a time-consuming and labor-intensive process. With the development of artificial intelligence technology, state classification based on the characteristics of satellite photometric curves has significant application prospects and value.

[0003] In speech time-series data processing, a common approach is to segment non-stationary, time-varying signals into multiple short-term, steady-state, time-invariant segments using frame-by-frame windowing. Frequency features are then extracted from each windowed frame. However, unlike the fixed frequency in speech signal sampling, photometric curve observations are affected by objective factors such as weather and season, leading to uneven sampling density and missing observations, making photometric curve recognition more challenging. Converting the data into fixed-length feature vectors and extracting discriminative features for classification has become a pressing technical problem. Furthermore, neural network training often relies on large amounts of labeled training data. Satellite photometric observation curves are not only scarce but also require professional annotation, making it difficult to train neural networks with existing data, often resulting in overfitting and weak model generalization. Additionally, traditional machine learning models and convolutional neural network models struggle to capture global information from photometric curves, leading to low model recognition accuracy. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method for preprocessing and classifying satellite photometric curve data. This method effectively captures the temporal information of satellite photometric curve data and standardizes the data into fixed-length feature vectors. Simultaneously, a random erasure strategy is proposed to simulate missing observation data, achieving the goal of data augmentation. Furthermore, to address the difficulty of existing models in capturing global features of photometric curves, a Transformer model with positional encoding is proposed for feature extraction and classification. This method has significant research implications for improving the classification accuracy of satellite photometric curves.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for preprocessing and classifying satellite photometric curve data includes the following steps:

[0007] Step (1) converts the phase angle-magnitude curves with discontinuous observations and uneven sampling frequencies into fixed-dimensional window function feature vectors through a frame-by-frame windowing preprocessing strategy;

[0008] Step (2) uses a random data erasure strategy to augment the fixed-dimensional window function feature vector. While simulating the lack of phase observation, different photometric curve feature vectors are generated by randomly erasing data multiple times.

[0009] Step (3) For the photometric curve feature vector, perform position encoding of the phase angle feature to obtain photometric features containing phase information;

[0010] Step (4) For the photometric features, a query vector, index vector and key value vector are constructed using a transformer model to complete the extraction of global features. Then, a fully connected layer is used to classify the photometric curve feature vector.

[0011] The beneficial effects of this invention are as follows: Addressing the problems of discontinuous photometric curve observations and uneven sampling frequencies, this invention proposes a feature extraction method involving frame-by-frame windowing of photometric curves, which can effectively extract temporal information from the photometric curves and standardize the data into fixed-length feature vectors; it proposes a random erasure strategy to achieve data augmentation, further solving the problems of limited observation data and difficulty in network convergence; it proposes position encoding of phase angle features, effectively combining phase information with photometric features, and finally uses a transformer classification model to complete the classification task of satellite target states. Attached Figure Description

[0012] Figure 1 This is a flowchart of a satellite photometric curve data preprocessing and classification method according to the present invention;

[0013] Figure 2 This is a schematic diagram of frame-by-frame windowing of the photometric curve in an embodiment of the present invention;

[0014] Figure 3 The statistical characteristics of the original satellite photometric curve and the window function after frame-by-frame windowing are given, where, Figure 3 (a) shows the original photometric curve. Figure 3 (b) is a schematic diagram of the curve mean after windowing. Figure 3 (c) is a schematic diagram of the curve variance after windowing;

[0015] Figure 4 Visualization images of the photometric curves before and after the random erasure strategy. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0017] like Figure 1 As shown, a flowchart of a satellite photometric curve data preprocessing and classification method according to the present invention is given. In step (1), the phase angle-magnitude curves with intermittent observation data and uneven sampling frequency are converted into window function feature vectors with fixed dimensions through a frame-windowing preprocessing strategy. In step (2), the above window function feature vectors are processed by random erasure data augmentation technology to simulate the missing phase observation situation, and diverse photometric curve feature vectors are generated through multiple random erasures. In step (3), the phase angle feature position encoding is implemented on the photometric curve feature vectors to ensure that the obtained photometric features contain phase information. In step (4), the global feature extraction of the encoded photometric curves is performed using the transformer model to construct query vectors, index vectors and key value vectors, and the extracted photometric curve feature vectors are finally classified through a fully connected layer. The specific implementation process of each step is described below:

[0018] 1. Frame segmentation, windowing, and feature preprocessing

[0019] like Figure 2 The diagram shows a frame-by-frame windowing scheme for the photometric curve. Specifically, for the i-th frame, let's assume a set of windows is used. The original curve is divided into frames, where s is the window shift and l is the window length. The mean μ of each window is calculated. i and variance

[0020]

[0021]

[0022] Where i∈0,1,...,K, Where N is the total number of frames, θ is the number of observation points, and θ is the total number of frames. n Let L be the phase angle at the nth point. n Let I(·) be the magnitude intensity at the nth point, and let I(·) be the indicator function.

[0023]

[0024] Where q represents the decision content, referring to θ in the above formula. nDoes it exist in window W? i If the condition is true, q is considered true; otherwise, q is considered false. In practice, a window shift of s = 0.05 and a window length of l = 1 are used, which divides the 0–180° phase angle into 3600 small windows, i.e., T = 3600. See also... Figure 3 , Figure 3 (a) shows the original satellite spectral intensity and phase angle variation curves (i.e., the original luminosity curves). After windowing, a feature vector of fixed length is obtained, as shown below. Figure 3 (b) and Figure 3 As shown in (c).

[0025] 2. Random Erasure Data Augmentation Strategy

[0026] like Figure 4 As shown, a schematic diagram of the random erasure strategy for photometric curves is presented. Due to the special nature of satellite point target photometric curves, the small training sample size is one of the main problems, and directly training on the data will lead to severe overfitting. On the other hand, in spectral phase sequence data, due to the discontinuity of observations, there are usually some phase gaps, which makes subsequent identification difficult. To address these two problems, this project proposes to use a random erasure method to augment the data, that is, randomly setting the statistics of a window function to zero. Let the interval where sample points exist in the data be [θ]. min ,θ max If the proportion of randomly erased fragments is r, then the random erasure process is as follows:

[0027]

[0028] Where, μ i 'and Let ξ represent the mean and variance of each window after random erasure, and let ξ ~ U(θ) be the random variable. min ,θ max Let θ be the initial point for selecting random erasure locations, and ~U(θ) min ,θ max ) indicates that ξ follows the interval [θ min ,θ max The uniform distribution function of ] is used. The mean μ after erasure is... i ′ and variance The concatenation is performed to obtain the photometric curve feature vector X∈R for subsequent classification. K×2 R represents the real number space. During model training, ξ will be repeatedly sampled, meaning that multiple curve data points will be erased from a single photometric observation curve. In the specific implementation, r is set to 0.3. For example... Figure 4 As shown, the left image displays the mean features extracted after windowing. Random erasure simulates the phase loss phenomenon in actual observations. Figure 4As shown in the right figure, on the one hand, it achieves the effect of simulating the lack of phase observations, and on the other hand, a satellite spectral phase sequence can generate different curves through multiple random erasures, thus achieving the purpose of data expansion.

[0029] 3. Phase position encoding

[0030] For each frame, the phase angle is position-coded (PE) to obtain the phase angle position-coded feature PE∈R. K×C :

[0031]

[0032]

[0033] Where C is the dimension of the encoded vector, and in the specific implementation, C = 256. k = 0, 1, ..., 127 is the index of the feature map channel dimension. θ∈0,1,...,3600 represents the position number corresponding to the feature within the window function, recording the magnitude of the phase angle and also representing the position number corresponding to the feature within the window function. The feature obtained from this positional encoding is divided into two parts: odd-numbered dimensions use a sine wave, and even-numbered dimensions use a cosine wave. In this way, the encoding can distribute positional information across different dimensions. The advantage of this is that it provides richer positional information and helps the subsequent transformer model capture the relative positional relationships between elements in the sequence.

[0034] 4. Photometric Curve Classification Model Based on Transformer

[0035] After feature extraction and position encoding, the feature vector X∈R of the photometric curve can be obtained. K×2 Phase angle position encoding feature PE∈R K×C Before constructing the query vector, index vector, and key value vector, a fully connected layer is used to align the star magnitude brightness features with the location encoding features:

[0036] F = FFN0(X) + PE

[0037] Wherein, FFN0 is a fully connected layer aligned with star magnitude features and position encoding features, F∈R K×C To obtain a feature vector that integrates star magnitude and phase angle features, a set of query vectors (Q) is then obtained through a convolutional layer. j Key value vector (K) j ) and index vector (V j ):

[0038] Q j =Φ j (F) K j =Ψj (F) V j =Θ j (F) j=1,2,...,h

[0039] Where h is the number of attention mechanism heads, Φ j ,Ψ j ,Θ j The query mapping function, key value mapping function, and index mapping function are all composed of fully connected layers and normalization functions to map and normalize the feature dimensions. Then, a multi-head attention mechanism is used for global feature extraction and classification.

[0040]

[0041] pred=softmax(FFN1(F+[head1,...,head h W O ))

[0042] Among them, head j Represents the j-th attention mechanism head. W is the number of channels in the attention head. O FFN1 represents a fully connected layer network for global feature extraction and classification, softmax is a normalization function, and finally the predicted value pred is constrained by the cross-entropy classification loss function.

[0043] Contents not described in detail in this specification are common knowledge to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications will be readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the inventive concept are protected.

Claims

1. A method for satellite photometric curve data preprocessing and classification, characterized in that, Includes the following steps: Step (1) converts the phase angle-magnitude curves with discontinuous observations and uneven sampling frequencies into fixed-dimensional window function feature vectors through a frame-by-frame windowing preprocessing strategy; Step (2) involves using a random erasure data augmentation strategy on the fixed-dimensional window function feature vector. This simulates missing phase observations while generating different photometric curve feature vectors through multiple random erasures; including: The phase angle interval in which there is a non-zero observation data point in the observation luminosity curve data is , the proportion of the randomly erased segment is , and the process of random erasure is as shown in the following formula: wherein, and denote the mean and variance of each window after random erasing, the random variable is the initial point for selecting the random erasing position, denotes obeys a uniform distribution function with phase angle interval , the mean and variance of each window after random erasing are spliced as the photometric curve feature vector for subsequent classification, R represents the real number space; Step (3) For the photometric curve feature vector, perform position encoding of the phase angle feature to obtain photometric features containing phase information; Step (4) For the photometric features, a query vector, index vector, and key value vector are constructed using a transformer model to complete the extraction of global features. Then, a fully connected layer is used to classify the photometric curve feature vector, including: After feature extraction and position encoding in steps (2) and (3), the photometric curve feature vector and phase angle position encoding feature are obtained. Before constructing the query vector, index vector, and key value vector, a fully connected layer is used to align the photometric curve feature vector X with the phase angle position encoding feature PE. in, This is a fully connected layer network that aligns star magnitude brightness features with location encoding features. To fuse the feature vectors of the photometric curve feature vector X and the phase angle position encoding feature PE, a set of query vectors is then obtained through a convolutional layer. Key value vector and index vector : in, For the number of attention mechanism heads, The query mapping function, key value mapping function, and index mapping function are all composed of fully connected layers and normalization functions to map and normalize the feature dimensions. Then, a multi-head attention mechanism is used for global feature extraction and classification. in, Represents the j-th attention mechanism head. It is the number of channels in the attention head. Represents a linear mapping function. This represents a fully connected network for global feature extraction and classification. The superscript T indicates transpose, softmax is the normalization function, and the final predicted value is... The predicted values ​​are constrained by the cross-entropy classification loss function.

2. The method for preprocessing and classifying satellite photometric curve data according to claim 1, characterized in that, Step (1) includes: The phase angle-magnitude curves, which are observed to be discontinuous and have uneven sampling frequencies, are divided into multiple frames, and the statistical characteristics within each window are calculated. For the i-th frame, let a set of windows be used. The phase angle-magnitude curves, which are observed discontinuously and have uneven sampling frequencies, are framed, where s is the window shift and l is the window length. The mean value of each window is calculated. and variance : in, , The total number of frames. This is the floor function, where N is the number of observed data points. Let n be the phase angle of the nth observation data point. Let n be the magnitude intensity of the star at the nth observation data point. For indicator functions: in, The content of the event, in the above formula refers to Does it exist in the window? If yes, q is considered true; otherwise, q is considered false.

3. The method for preprocessing and classifying satellite photometric curve data according to claim 1, characterized in that, Step (3) includes: Phase angle position encoding is performed on the feature vector of the photometric curve to obtain the phase angle position encoded feature. : Where C is the dimension of the position-encoded vector. , is the index of the feature map channel dimension. , represents the magnitude of the phase angle, and also represents the position number corresponding to the feature of the window function; the phase angle feature obtained by this position encoding is divided into two parts: the odd dimension uses the sin sine wave, and the even dimension uses the cosine wave.

Images