Enhancing Anomaly Detection in Sensor Time Series Data using Machine Learning Model Optimization

Launch vehicles and satellites produce a substantial amount of sensor data primarily indented for monitoring and control applications. Human expertise plays an important role in interpreting the sensor data, detecting the anomalies, and making decisions to safeguard the underlying system, entails inherent operational risks. To reduce costs, automation using machine learning algorithms to enhance the identification of anomalies are presently employed, but it exhibits inaccuracies in its detection capabilities. Our investigation highlights the effectiveness of deep learning (DL) methods, including optimized architectures of ‘Long Short-Term Memory (LSTM) networks’, in detecting anomalies, improving performance, and reducing computation time. In the scenario of anomaly detection, various techniques including regularization methods such as early stopping, additional dropout, and input dimensionality reduction are employed. The proposed architecture is realized and general effectiveness of the different architectures are compared using publicly accessible expert-labeled telemetry datasets. The performance of the baseline improves with the optimization approaches, as seen in the experimental results.