Luminosity curve anomaly detection method based on long and short term memory network

Abstract

The invention provides a luminosity curve anomaly detection method based on a long and short term memory network, and relates to the technical field of space target intelligent perception. In view of the problems of dependence on labeled data, noise sensitivity, insufficient local anomaly capture capability and the like in the prior art, high-precision anomaly recognition is realized by constructing an unsupervised learning framework. According to the method, the time sequence dependency relationship is modeled through the long and short term memory network, and the time attention mechanism is introduced to dynamically focus the key time step, so that the anomaly detection precision of the model is improved. The invention provides an efficient solution for sensing the abnormal movement of the space target, and is suitable for real-time monitoring of complex scenes such as a geosynchronous orbit and earth-moon space.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent perception of space targets, and in particular to a light curve anomaly detection method based on a long short-term memory network. Background Art

[0002] With the rapid development of space technology, the number of space targets continues to increase, space activities are becoming more frequent, and the complexity and dynamics of the space environment have significantly increased. Space situational awareness, as a key means of ensuring the safe operation of spacecraft, avoiding orbital collisions, and improving the reliability of space missions, has garnered widespread attention. In particular, high-precision perception of space targets and monitoring of abnormal behavior in critical orbital environments such as geosynchronous orbit, high-orbit regions, and cis-lunar space have become important areas of current space security research.

[0003] Existing methods for monitoring space targets rely heavily on active remote sensing technologies like radar and laser ranging, but these technologies are costly, lack real-time performance, and are difficult to deploy on a large scale across the entire orbital range. In contrast, monitoring methods based on light curves (i.e., the curves of target brightness variations) offer the advantages of being passive, low-cost, and providing wide coverage, making them a hot topic of research in recent years. By modeling the temporal characteristics of a target's light curve, it is possible to indirectly infer its attitude changes, structural anomalies, or operational abnormalities, thereby enabling the identification and detection of space target behavior.

[0004] Existing light curve anomaly detection methods can be primarily categorized into two main categories: traditional machine learning and deep learning. Traditional machine learning methods typically require feature extraction from the light curve. These methods then perform anomaly detection based on these features through clustering, regression, and classification. However, these methods rely heavily on the quality of handcrafted features, making generalization and robustness difficult to guarantee. With the advancement of deep learning technology, deep learning methods such as convolutional neural networks and recurrent neural networks have demonstrated promising performance in light curve anomaly detection. However, current light curve-based anomaly detection methods still face numerous challenges. For example, they rely on large amounts of labeled data. Many methods require training based on manually labeled anomaly samples. However, in practice, anomaly data is scarce and difficult to obtain, limiting the model's generalization capabilities. Furthermore, actual light curves collected often suffer from significant noise, including measurement errors and environmental interference. Conventional methods struggle to accurately distinguish between noise and anomaly signals, leading to misjudgments.

[0005] In short, existing methods find it difficult to achieve high-precision anomaly detection when faced with scenarios where there are no anomaly annotations and severe noise interference. Summary of the Invention

[0006] This paper addresses the noise sensitivity and inadequate ability to capture local anomalies in existing light curve anomaly detection methods. By constructing an unsupervised learning framework, this method relies solely on training models with normal samples to accurately identify unknown anomalies. By combining the robust time series modeling capabilities of the long short-term memory network with a temporal attention mechanism that dynamically focuses on key time points, this method significantly improves the accuracy and robustness of light curve anomaly detection in an unsupervised setting.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a light curve anomaly detection method based on a long short-term memory network, comprising the following steps:

[0008] Step (1) dividing the light curve simulation data set containing normal and abnormal data into a training set and a test set, and preprocessing the training set and the test set; wherein the training set only contains normal data, and each light curve data in the test set contains at least one abnormal point;

[0009] Step (2) uses a sliding window to divide the light curve into a fixed-length history sequence to construct the input sample;

[0010] Step (3) Use the long short-term memory network to model the temporal dependency of the data in the current window and obtain the hidden state of each time step;

[0011] Step (4) uses the temporal attention module to calculate the attention score for the hidden state of each time step, and multiplies the hidden state of each time step by the attention score. After passing through the output module, the prediction of the star magnitude value of the current window corresponding to the next moment is obtained;

[0012] Step (5) calculates the prediction error based on the model's prediction of the magnitude value and the actual observation value, and determines the threshold based on the statistical distribution of the training set error to identify abnormal events.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] (1) No need for abnormal labeling and high practicality: The present invention adopts an unsupervised learning framework and only uses normal light curve data for training. It does not need to rely on manually labeled abnormal samples. It is suitable for application scenarios in actual space monitoring where abnormal data is scarce or cannot be labeled, and has stronger generalization ability and practicality.

[0015] (2) Strong robustness and excellent noise resistance: In response to the observation errors and environmental noise problems that are common in real light curves, the present invention enhances the model's ability to focus on abnormal points by introducing a temporal attention mechanism, effectively improving the model's recognition accuracy for abnormal behaviors in a high-noise background and reducing the risk of misjudgment and missed judgment.

[0016] (3) Strong temporal modeling capability and adaptability to complex dynamic features: This invention models the temporal dependence of the light curve based on the long short-term memory network, and combines the attention mechanism to further enhance the model's learning ability for key time steps, significantly improving the ability to respond to complex dynamic processes (such as rapid attitude changes, orbit adjustments, etc.).

[0017] (4) High detection accuracy and excellent computational efficiency: Experimental results show that the method of the present invention can achieve high-precision detection of abnormal events in the light curve of space targets while ensuring real-time performance, and has strong engineering application potential and deployment value. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is an overall flow chart of a light curve anomaly detection method based on a long short-term memory network of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0020] like Figure 1 As shown, a light curve anomaly detection method based on a long short-term memory network of the present invention includes the following steps:

[0021] Step (1) dividing the light curve simulation data set containing normal and abnormal data into a training set and a test set, and preprocessing the training set and the test set; wherein the training set only contains normal data, and each light curve data in the test set contains at least one abnormal point;

[0022] Step (2) uses a sliding window to divide the light curve into a fixed-length history sequence to construct the input sample;

[0023] In step (2), according to the temporal characteristics of the space target light curve, the sliding window method is used to segment the original light sequence and construct structured model input samples. Given the original light curve sequence:

[0024] ,

[0025] The sliding window length is defined as W=20, and each input sample consists of a historical sequence within a window The corresponding prediction target constitute:

[0026] ,

[0027] ,

[0028] The input samples constructed by sliding windows have a local temporal structure, which helps to capture the trends and fluctuation characteristics in the light curve of space targets.

[0029] Step (3) Use the long short-term memory network to model the temporal dependency of the data in the current window and obtain the hidden state of each time step;

[0030] In step (3), the long short-term memory network includes two LSTM layers, each layer includes 64 hidden units, the input dimension is 1, and the output dimension is 64; the long short-term memory network is used to model the temporal dependency relationship of the data in the current window to obtain the hidden state of each time step.

[0031] Specifically, given the input sequence:

[0032] ,

[0033] The LSTM network performs the following processing for each time step:

[0034] The first is the forget gate. The forget gate determines how much historical information is saved or forgotten at the current time step:

[0035] ,

[0036] in, Determines whether this information will be forgotten or retained. is the hidden state at the previous time step, is the weight matrix of the forget gate, is the bias vector of the forget gate, is the sigmoid activation function.

[0037] Next is the input gate. The input gate controls how much new information is added to the current cell state:

[0038] ,

[0039] ,

[0040] in, is the activation function, and is the corresponding weight matrix in the input gate, and is the corresponding bias vector. According to the results obtained by the forget gate and the input gate, the cell state of the current time step will be updated:

[0041] ,

[0042] in, represents the amount of new information introduced into the cell state at the current time step, Represents the candidate unit status information.

[0043] Finally, there is the output gate, which determines how much information in the cell state at the current time step is output as a hidden state, as the output of the current time step, and passed to the next time step:

[0044] ,

[0045] ,

[0046] in, Determines how much information is output in the current unit state, is the hidden state of the current time step, is the weight matrix of the output gate, is the bias vector of the output gate.

[0047] The hidden state of each time step obtained by the long short-term memory network will be passed to the subsequent temporal attention module to further improve the prediction accuracy.

[0048] Step (4) uses the temporal attention module to calculate the attention score for the hidden state of each time step, and multiplies the hidden state of each time step by the attention score. After passing through the output module, the prediction of the star magnitude value of the current window corresponding to the next moment is obtained;

[0049] In step (4), the temporal attention module consists of a fully connected layer, a tanh activation function, a fully connected layer, and a softmax layer. The first fully connected layer has an input dimension of 64 and an output dimension of 32; the second fully connected layer has an input dimension of 32 and an output dimension of 1. The output module consists of a dropout layer and a fully connected layer. The dropout layer can reduce the risk of model overfitting. The fully connected layer has an input dimension of 64 and an output dimension of 1.

[0050] For each time step of the LSTM network output, the hidden state , after the time attention module will calculate an attention score:

[0051] ,

[0052] in, and is the weight matrix and bias vector corresponding to the first fully connected layer, and is the weight matrix and bias vector corresponding to the second fully connected layer, and is the activation function. Then with the hidden state Multiply and accumulate over all time steps to get the feature vector C:

[0053] ,

[0054] Finally, after the feature vector C passes through the output module, the prediction of the magnitude value of the current window corresponding to the next moment can be obtained:

[0055] ,

[0056] ,

[0057] ,

[0058] Among them, R is a mask vector with the same shape as C. Each element in R is randomly sampled from the Bernoulli distribution with probability p; the mask R is multiplied element by element with the feature vector C and is scaled by the factor After scaling, the feature vector is obtained The purpose of scaling is to ensure numerical continuity during training and testing. This masking operation can randomly discard a certain proportion p of elements in the feature vector C, thereby reducing the risk of overfitting the model. Feature vector Finally, a weight matrix is , the bias vector is The fully connected layer obtains the final magnitude prediction output .

[0059] Step (5) calculates the prediction error based on the model's prediction of the magnitude value and the actual observed value, and determines the threshold based on the statistical distribution of the training set error to identify abnormal events.

[0060] In step (5), the magnitude value predicted by the model is compared with the actual observed value to calculate the prediction error (take the absolute value):

[0061] ,

[0062] In order to establish the threshold for anomaly detection, the prediction error is first calculated on all samples in the training set. Since the training set consists entirely of normal data, the prediction error on the training set actually reflects the modeling ability of the model on normal data and the reasonable fluctuation range of data in the normal light curve. Luminosity fluctuations outside this range are considered to be possible anomalies. Specifically, let the mean prediction error of the model on the training set be , the variance is , then the threshold for anomaly detection is Define it as follows:

[0063] ,

[0064] k is an adjustable hyperparameter that controls the sensitivity of anomaly detection. A larger value for k makes the model more conservative in detecting anomalies, identifying only prominent outliers. Conversely, a smaller value for k increases the sensitivity of anomaly detection, but also increases the risk of false alarms. In practical applications, k can be adjusted based on task requirements, striking a balance between missed detections and false alarms.

[0065] During the testing process, the model will predict new data points and calculate the prediction error. If the prediction error of a test data point exceeds the set threshold, that is:

[0066] ,

[0067] Then the point will be judged as an anomaly.

[0068] Example:

[0069] All programs are based on Python language, implemented under the Pytorch framework, and trained and tested using a single NVIDIA GeForce RTX3090 GPU.

[0070] The specific steps of the present invention include:

[0071] Step (1) dividing the light curve simulation data set containing normal and abnormal data into a training set and a test set, and preprocessing the training set and the test set; wherein the training set only contains normal data, and each light curve data in the test set contains at least one abnormal point;

[0072] Step (2) uses a sliding window to divide the light curve into a fixed-length history sequence to construct the input sample;

[0073] Step (3) Use the long short-term memory network to model the temporal dependency of the data in the current window and obtain the hidden state of each time step;

[0074] Step (4) uses the temporal attention module to calculate the attention score for the hidden state of each time step, and multiplies the hidden state of each time step by the attention score. After passing through the output module, the prediction of the star magnitude value of the current window corresponding to the next moment is obtained;

[0075] Step (5) calculates the prediction error based on the model's prediction of the magnitude value and the actual observation value, and determines the threshold based on the statistical distribution of the training set error to identify abnormal events.

[0076] Compared to existing techniques, this method uses a long short-term memory network to model the temporal dependencies in light curves. The temporal attention module adaptively assigns different levels of attention to different time steps, allowing the model to focus on key time steps and reduce interference from factors such as noise. A comparison with other methods is shown in Table 1.

[0077] Table 1

[0078] It should be emphasized that the above are only preferred embodiments of the present invention and do not limit the present invention in any form. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A light curve anomaly detection method based on long short-term memory network, characterized in that: The following steps are involved: Step (1) dividing the light curve simulation data set containing normal and abnormal data into a training set and a test set, and preprocessing the training set and the test set; The training set contains only normal data, and each light curve data in the test set contains at least one abnormal point; Step (2) uses a sliding window to divide the light curve into a fixed-length history sequence to construct the input sample; Step (3) Use the long short-term memory network to model the temporal dependency of the data in the current window and obtain the hidden state of each time step; Step (4) uses the temporal attention module to calculate the attention score for the hidden state of each time step, and multiplies the hidden state of each time step by the attention score. After passing through the output module, the prediction of the star magnitude value of the current window corresponding to the next moment is obtained; Step (5) calculates the prediction error based on the model's prediction of the magnitude value and the actual observation value, and determines the threshold based on the statistical distribution of the training set error to identify abnormal events.

2. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 1, characterized in that: In step (2), according to the temporal characteristics of the light curve, the sliding window method is used to segment the original light curve and construct a structured model input sample. Given the original light curve sequence: ， Define a sliding window, each input sample consists of a historical sequence within a window The corresponding prediction target constitute: ， ， The input samples constructed by sliding windows have a local temporal structure, which helps to capture the trends and fluctuation characteristics in the light curve of space targets.

3. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 1, characterized in that: In step (3), the long short-term memory network includes two LSTM layers, each layer includes 64 hidden units, the input dimension is 1, and the output dimension is 64; the long short-term memory network is used to model the temporal dependency relationship of the data in the current window to obtain the hidden state of each time step.

4. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 3, characterized in that: Given an input sequence: ， The LSTM network performs the following processing for each time step: The first is the forget gate; the forget gate determines how much historical information is saved or forgotten at the current time step: ， in, Determines whether this information will be forgotten or retained. is the hidden state at the previous time step, is the weight matrix of the forget gate, is the bias vector of the forget gate, is the sigmoid activation function; Next is the input gate; the input gate controls how much new information is added to the current cell state: ， ， in, is the activation function, and is the corresponding weight matrix in the input gate, and is the corresponding bias vector. According to the results obtained by the forget gate and the input gate, the cell state of the current time step will be updated: ， in, represents the amount of new information introduced into the cell state at the current time step, Represents the candidate unit status information; Finally, there is the output gate, which determines how much information in the cell state at the current time step is output as a hidden state, as the output of the current time step, and passed to the next time step: ， ， in, Determines how much information is output in the current unit state, is the hidden state of the current time step, is the weight matrix of the output gate, is the bias vector of the output gate.

5. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 1, characterized in that: In step (4), the temporal attention module consists of a fully connected layer, a tanh activation function, a fully connected layer and a softmax layer, wherein the first fully connected layer has an input dimension of 64 and an output dimension of 32; the second fully connected layer has an input dimension of 32 and an output dimension of 1; the output module consists of a dropout layer and a fully connected layer, wherein the fully connected layer has an input dimension of 64 and an output dimension of 1.

6. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 5, characterized in that: For each time step of the LSTM network output, the hidden state , an attention score is calculated through the time attention module: ， in, and is the weight matrix and bias vector corresponding to the first fully connected layer, and is the weight matrix and bias vector corresponding to the second fully connected layer, and As the activation function, the attention score Then with the hidden state Multiply and accumulate over all time steps to get the feature vector C: ， Finally, after the feature vector C passes through the output module, the prediction of the magnitude value of the current window corresponding to the next moment can be obtained: ， ， ， Among them, R is a mask vector with the same shape as C. Each element in R is randomly sampled from the Bernoulli distribution with probability p; the mask R is multiplied element by element with the feature vector C and is scaled by the factor After scaling, the feature vector is obtained , the eigenvector Finally, a weight matrix is , the bias vector is The fully connected layer obtains the final magnitude prediction output .

7. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 1, characterized in that: In step (5), the mean prediction error of the model on the training set is , the variance is , then the threshold for anomaly detection is Defined as follows: ， Among them, k is an adjustable hyperparameter.

8. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 7, characterized in that: The prediction error of a test data point exceeds the set threshold, that is: ， This point is identified as abnormal.

9. The method for detecting anomalies in light curves based on a long short-term memory network according to claim 2, characterized in that: The sliding window length is W=20.

Images