A photometric curve anomaly detection method based on long short-term memory network

Abstract

The application provides a photometric curve anomaly detection method based on a long short-term memory network and relates to the technical field of intelligent perception of space targets. In view of the problems of dependence on labeled data, noise sensitivity and insufficient local anomaly capturing capacity in the prior art, a high-precision anomaly recognition is realized through the construction of an unsupervised learning framework. The long short-term memory network is used for modeling the time sequence dependence relationship, and a time attention mechanism is introduced to dynamically focus on key time steps, so that the anomaly detection precision of the model is improved. The application provides an efficient solution for space target abnormality perception and is suitable for real-time monitoring of complex scenes such as geosynchronous orbits and earth-moon space.

Description

TECHNICAL FIELD

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

[0002] With the rapid development of space technology, the number of space targets is increasing, space activities are becoming more frequent, and the complexity and dynamics of the space environment are significantly improving. Space situation awareness, as an important means to ensure the safe operation of spacecraft, avoid orbital collisions, and improve the reliability of space missions, has attracted widespread attention. In particular, in key orbital environments such as geosynchronous orbit, high-orbit region, and lunar space, high-precision perception and abnormal behavior monitoring of space targets have become an important direction of current space safety research.

[0003] Existing space target monitoring methods mostly rely on active remote sensing technologies such as radar and laser ranging, but they are costly, have poor real-time performance, and are difficult to deploy on a large scale in the entire orbit. In contrast, methods based on photometric curves (i.e., the brightness variation curve of a target) have the advantages of being passive, low-cost, and widely covered, and have become a hot research topic in recent years. By modeling the temporal characteristics of the photometric curve of a target, its attitude change, structural anomaly, or abnormal operating state can be indirectly inferred, and thus space target behavior recognition and anomaly detection can be achieved.

[0004] Existing photometric curve anomaly detection methods can be mainly divided into two categories: traditional machine learning methods and deep learning methods. Traditional machine learning methods usually need to extract features from the photometric curve first, and then complete anomaly detection through clustering, regression, classification, etc. based on the extracted features, but these methods are too dependent on the quality of handcrafted features, and their generalization and robustness cannot be guaranteed. With the development of deep learning technology, deep learning methods such as convolutional neural networks and recurrent neural networks have shown good performance in photometric curve anomaly detection. However, current anomaly detection methods based on photometric curves still face many difficult problems, such as the dependence on a large amount of labeled data. Many methods need to be trained based on artificially labeled abnormal samples, but in practice, abnormal data is scarce and difficult to obtain, which limits the generalization ability of the model. In addition, the photometric curves actually collected often have measurement errors, environmental interference, and other factors, and there is serious noise interference. Conventional methods are difficult to accurately distinguish between noise and abnormal signals, and are prone to false positives.

[0005] In summary, existing methods are difficult to achieve high-precision anomaly detection in scenarios with no abnormal annotations and severe noise interference. SUMMARY

[0006] The present application aims at the problems of noise sensitivity and insufficient local anomaly capturing ability in existing photometric curve anomaly detection methods, and proposes a photometric curve anomaly detection method based on a long short-term memory network. By constructing an unsupervised learning framework, the model is trained only with normal samples, and the unknown anomalies are accurately identified. The method combines the long short-term memory network with strong time series modeling capability and the time attention mechanism with dynamic focusing ability at key time points, significantly improving the accuracy and robustness of photometric curve anomaly detection in an unsupervised background.

[0007] To achieve the above purpose, the present application adopts the following technical scheme: a photometric curve anomaly detection method based on a long short-term memory network, comprising the following steps:

[0008] Step (1) divides the photometric curve simulation data set containing normal and abnormal data into a training set and a test set, and pre-processes the training set and the test set; wherein the training set only contains normal data, and each photometric curve data in the test set contains at least one abnormal point;

[0009] Step (2) divides the photometric curve into a fixed-length history sequence by using a sliding window to construct an input sample;

[0010] Step (3) uses a long short-term memory network to model the time series dependence of data in the current window to obtain the hidden state of each time step;

[0011] Step (4) uses a time 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 magnitude value at the next time corresponding to the current window is obtained;

[0012] Step (5) calculates the prediction error according to the prediction of the magnitude value by the model and the actual observation value, and determines the threshold based on the statistical distribution of the error of the training set to determine the abnormal event.

[0013] Compared with the prior art, the present application has the following advantages:

[0014] (1) No need for abnormal labeling, high practicality: the present application uses an unsupervised learning framework, and only uses normal photometric curve data for training, without relying on manually labeled abnormal samples, which is suitable for application scenarios where abnormal data is scarce or cannot be labeled in actual space monitoring, and has stronger generalization ability and practicality.

[0015] (2) Strong robustness, superior noise resistance: in view of the observation error and environmental noise problems commonly existing in real photometric curves, the present application enhances the attention ability of the model to abnormal points by introducing a time attention mechanism, effectively improves the identification accuracy of the model to abnormal behaviors in a high-noise background, and reduces the risk of misjudgment and omission.

[0016] (3) Strong timing modeling capability, adaptive to complex dynamic characteristics: The present application models the time dependence of the photometric curve based on the long short-term memory network, and further strengthens the learning ability of the model to key time steps in combination with the attention mechanism, which significantly improves the response capability to complex dynamic processes (such as rapid posture change, orbit adjustment, etc.).

[0017] (4) High detection accuracy and high calculation efficiency: The experimental results show that the present application can realize high-precision detection of abnormal events in the photometric curve of a space target under the premise of ensuring real-time performance, and has strong engineering application potential and deployment value. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is the overall flowchart of a photometric curve anomaly detection method based on a long short-term memory network of the present application. DETAILED DESCRIPTION

[0019] The present application will be described in detail below in combination with the drawings and examples.

[0020] As shown in Figure 1 , the photometric curve anomaly detection method based on the long short-term memory network of the present application comprises the following steps:

[0021] Step (1) divides the photometric curve simulation data set containing normal and abnormal data into a training set and a test set, and pre-processes the training set and the test set; wherein the training set only contains normal data, and each photometric curve data in the test set contains at least one abnormal point;

[0022] Step (2) divides the photometric curve into a fixed-length history sequence using a sliding window to construct an input sample;

[0023] In the step (2), the original photometric sequence is segmented using a sliding window method according to the timing characteristics of the space target photometric curve to construct a structured model input sample. Given the original photometric curve sequence:

[0024] ,

[0025] Define the sliding window length as W=20, and each input sample is composed of a history sequence and the corresponding prediction target :

[0026] ,

[0027] ,

[0028] Input samples constructed using a sliding window possess a local temporal structure, which helps to capture the trend and fluctuation characteristics in the photometric curves of spatial targets.

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

[0030] In step (3), the Long Short-Term Memory network contains two LSTM layers, each containing 64 hidden units, with an input dimension of 1 and an output dimension of 64. The Long Short-Term Memory network is used to model the temporal dependencies of the data in the current window to obtain the hidden state at each time step.

[0031] Specifically, given the input sequence:

[0032] ,

[0033] Long Short-Term Memory (LSTM) networks process each time step as follows:

[0034] 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, This determines whether this information is forgotten or retained. It is the hidden state of the previous time step. It is the weight matrix of the forget gate. It is the bias vector of the forget gate. It 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, For activation function, and It is the corresponding weight matrix in the input gate. and This is the corresponding bias vector. Based on the results obtained from the forget gate and the input gate, the cell state at the current time step will be updated:

[0041] ,

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

[0043] Finally, there's the output gate, which determines how much information from the current time step's cell state is output as a hidden state, serving as the output for the current time step and being passed to the next time step.

[0044] ,

[0045] ,

[0046] in, This determines how much information from the current cell state is output. The hidden state at the current time step. Here is the weight matrix of the output gate. This is the bias vector for the output gate.

[0047] The hidden state obtained at each time step by the Long Short-Term Memory network will be passed to the subsequent temporal attention module to further improve prediction accuracy.

[0048] Step (4) Calculate the attention score for the hidden state of each time step using the time attention module, multiply the hidden state of each time step by the attention score, and then pass it through the output module to obtain the prediction of the star magnitude value of the next time step corresponding to the current window.

[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, and the fully connected layer has an input dimension of 64 and an output dimension of 1.

[0050] For the hidden state of the output of the Long Short-Term Memory network at each time step The time-based attention module will calculate an attention score:

[0051] ,

[0052] in, and These are the weight matrix and bias vector corresponding to the first fully connected layer. and These are the weight matrix and bias vector corresponding to the second fully connected layer. and The activation function is used to score the attention. Then in hidden state Multiply the results and sum them over all time steps to obtain the eigenvector C:

[0053] ,

[0054] Finally, after the feature vector C is passed through the output module, the prediction of the star magnitude value for the next time step corresponding to the current window can be obtained:

[0055] ,

[0056] ,

[0057] ,

[0058] Here, R is a mask vector with the same shape as C, and each element in R is randomly sampled from a Bernoulli distribution with probability p; the mask R is multiplied element-wise by the feature vector C and then processed by a scaling factor. After scaling, the feature vector is obtained. The purpose of scaling is to ensure numerical continuity during training and testing. This masking operation randomly discards a certain proportion of p elements in the feature vector C, thereby reducing the risk of overfitting. Feature vector Finally, a weight matrix is ​​used as follows: The bias vector is The fully connected layer yields the final star magnitude prediction output. .

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

[0060] In step (5), the prediction error (absolute value) is calculated by comparing the model's prediction of the magnitude with the actual observed value:

[0061] ,

[0062] To establish an anomaly detection threshold, 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 model's ability to model on normal data and the reasonable fluctuation range of data in a normal photometric curve. Photometric fluctuations exceeding this range are considered potentially anomalies. Specifically, let the mean of the model's prediction error on the training set be... The variance is The threshold for anomaly detection is... Defined as follows:

[0063] ,

[0064] Here, k is an adjustable hyperparameter that controls the sensitivity of anomaly detection. The larger the value of k, the more conservative the model is in detecting anomalies, only identifying the most prominent ones. Conversely, the smaller the value of k, the higher the sensitivity of anomaly detection, but the higher the risk of false alarms. In practical applications, k can be adjusted according to task requirements to strike a balance between missed detections and false alarms.

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

[0066] ,

[0067] Then that point will be judged as abnormal.

[0068] Example:

[0069] All programs are based on the Python language and implemented in the PyTorch framework, using a single NVIDIA GeForce RTX3090 GPU for training and testing.

[0070] The specific steps of this invention include:

[0071] Step (1) Divide the photometric curve simulation dataset containing normal and abnormal data into training set and test set, and preprocess the training set and test set; the training set contains only normal data, and each photometric curve data in the test set contains at least one abnormal point;

[0072] Step (2) uses a sliding window to segment the photometric curve into a fixed-length historical sequence to construct the input sample;

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

[0074] Step (4) Calculate the attention score for the hidden state of each time step using the time attention module, multiply the hidden state of each time step by the attention score, and then pass it through the output module to obtain the prediction of the star magnitude value of the next time step corresponding to the current window.

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

[0076] Compared with existing technologies, this invention utilizes a Long Short-Term Memory (LSTM) network to model the temporal dependencies in the photometric curve, and adaptively assigns different levels of attention to different time steps through a temporal attention module. This allows 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]

[0079] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for detecting anomalies in photometric curves based on long short-term memory networks, characterized in that, Includes the following steps: Step (1) Divide the photometric curve simulation dataset containing normal and abnormal data into training set and test set, and preprocess the training set and test set; The training set contains only normal data, while each photometric curve in the test set contains at least one outlier. Step (2) uses a sliding window to segment the photometric curve into a fixed-length historical sequence to construct the input sample; Step (3) Use a long short-term memory network to model the temporal dependencies of the data in the current window and obtain the hidden state at each time step; Step (4) Calculate the attention score for the hidden state of each time step using the time attention module, multiply the hidden state of each time step by the attention score, and then pass it through the output module to obtain the prediction of the star magnitude value of the next time step corresponding to the current window. Step (5) Calculate the prediction error based on the model's prediction of star magnitude and the actual observed value, and determine the threshold based on the statistical distribution of the training set error to identify abnormal events; In step (3), the Long Short-Term Memory network contains two LSTM layers, each containing 64 hidden units, with an input dimension of 1 and an output dimension of 64; the Long Short-Term Memory network is used to model the temporal dependencies of the data in the current window to obtain the hidden state at each time step; Given the input sequence: ， Long Short-Term Memory (LSTM) networks process each time step as follows: First is the forget gate; the forget gate determines how much historical information is saved or forgotten at the current time step: ， in, This determines whether this information is forgotten or retained. It is the hidden state of the previous time step. It is the weight matrix of the forget gate. It is the bias vector of the forget gate. It 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, For activation function, and It is the corresponding weight matrix in the input gate. and This is the corresponding bias vector. Based on the results obtained from the forget gate and the input gate, the cell state at the current time step will be updated: ， in, This represents the amount of new information introduced into the cell state at the current time step. This represents the candidate cell status information; Finally, there's the output gate, which determines how much information from the current time step's cell state is output as a hidden state, serving as the output for the current time step and being passed to the next time step. ， ， in, This determines how much information from the current cell state is output. The hidden state at the current time step. Here is the weight matrix of the output gate. This is the bias vector for the output gate; 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. For the hidden state at each time step of the Long Short-Term Memory network output An attention score is calculated using the time-based attention module: ， in, and These are the weight matrix and bias vector corresponding to the first fully connected layer. and These are the weight matrix and bias vector corresponding to the second fully connected layer. and For the activation function, the attention score is... Then in hidden state Multiply the results and sum them over all time steps to obtain the eigenvector C: ， Finally, after the feature vector C is passed through the output module, the prediction of the star magnitude value for the next time step corresponding to the current window can be obtained: ， ， ， Here, R is a mask vector with the same shape as C, and each element in R is randomly sampled from a Bernoulli distribution with probability p; the mask R is multiplied element-wise by the feature vector C and then processed by a scaling factor. After scaling, the feature vector is obtained. eigenvectors Finally, a weight matrix is ​​used as follows: The bias vector is The fully connected layer yields the final star magnitude prediction output. .

2. The photometric curve anomaly detection method based on a long short-term memory network according to claim 1, characterized in that: In step (2), considering the temporal characteristics of the photometric curve, the original photometric sequence is segmented using a sliding window method to construct a structured model input sample, given the original photometric curve sequence: ， Define a sliding window, where each input sample consists of a historical sequence within a window. With the corresponding prediction target constitute: ， ， Input samples constructed using a sliding window possess a local temporal structure, which helps to capture the trend and fluctuation characteristics in the photometric curves of spatial targets.

3. The photometric curve anomaly detection method based on a long short-term memory network according to claim 1, characterized in that: In step (5), the mean of the prediction error of the model on the training set is The variance is The threshold for anomaly detection is... Defined as follows: ， Here, k is an adjustable hyperparameter.

4. The photometric curve anomaly detection method based on a long short-term memory network according to claim 3, characterized in that: The prediction error for a certain test data point exceeds the set threshold, that is: ， This point was identified as abnormal.

5. The photometric curve anomaly detection method based on a long short-term memory network according to claim 2, characterized in that: The sliding window length is W=20.

Images