In order to perform data-based approach anomaly detection using machine learning, supervised learning, semi-supervised learning, and unsupervised learning method are applied in consideration of labeling related to normal and abnormal state of utilization data. This paper analyzes the feasibility of applying the machine learning-based anomaly detection method through the case study for each learning method by using the multidimensional time series data of satellite attitude control system and multi drone system among aerospace systems.
As a labeled data example, we used the data set created in the normal and abnormal state through LUNASIM, the attitude control simulator of Korea Pathfinder Lunar Orbiter (KPLO) under development by Korea Aerospace Research Institute. In fact, most satellite systems operate normally, so the number of samples of normal and abnormal data sets that can be obtained is not the same or equivalent. Therefore, after learning the generation model using only normal data, we applied unsupervised learning method that distinguishes between normal and abnormal by measuring reconstruction errors when new data that is not used for training is input. Using the deep learning technique (1D-CNN, Convolutional Neural Network), we extracted the feature of multidimensional time series data of satellite attitude control system and applied artificial neural network learning algorithm based on two generative models (VAE and GANomaly). In addition, prior to training the neural network, hyperparameter optimization algorithm based on Bayesian optimization was added to determine appropriate hyperparameters to automatically obtain a well-generalized neural network for the target task using AUROC (Area Under ROC Curve), a performance indicator. As a result, in various satellite operating modes that constitute the full scenario in which the satellite operates, it was confirmed that it could detect anomalies within the satellite self-control system and identify anomaly types related to mechanical and control logic.
As unlabeled data, we used a swarm flight test data set conducted by Korea Aerospace Research Institute. Data characteristics suitable for the purpose of identifying and predicting anomalies among a large number of aircraft operating in clusters were derived, and K-mean clustering of the unsupervised learning methods was used after reducing the data dimension through principal component analysis (PCA). Based on the results identified, supervised learning based binary classification model was studied using labeled normal and abnormal data sets.
In cases where the number of normal and abnormal objects is significantly different (class imbalanced circumstance) the binary classification performance is reduced, and in order to solve the problem, normal and abnormal data are sampled in a ratio of one to one when creating a batch for the stochastic optimization. Using test data not used for learning, the diagnostic algorithm produced the results of anomaly detection, which was verified by the actual flight test data analysis method, to ensure that the aircraft with abnormal behavior in the cluster is correctly identified.
In order to enhance reliability in aerospace systems and their operations, appropriate decisions need to be made depending on the circumstances, taking into account the level of their impact on the overall system, even if they detect and identify anomalies within the system. To this end, a simple idea of quantified decision-making methods based on an anomaly detection was proposed. This is meaningful in suggesting future research directions for the expansion and utilization of anomaly detection techniques.