Real-Time Detection and Diagnosis of Flight Anomalies in Small Unmanned Aerial Systems
0:00:12 Introduction: Dr. Mahbub introduces the speaker, Dr. Mohammad Nafi Al Islam, an assistant professor of computer science at the University of San Diego.
0:01:18 Research Focus: Dr. Islam's research focuses on cyber-physical systems, artificial intelligence, and small unmanned aerial systems (sUAS), with an emphasis on anomaly detection, mission-specific behavior configuration, and real-time safety analysis.
0:02:25 Presentation Overview: The talk will cover anomaly detection and diagnosis in sUAS, including deep learning and heuristic-based detectors, real-world applications, and data collection efforts.
0:04:06 Drones and Safety: The increasing use of drones in various sectors necessitates robust safety measures to prevent accidents and ensure public safety.
0:06:02 Research Motivation: Dr. Islam's research is motivated by real-world incidents where drones malfunctioned, highlighting the need for real-time anomaly detection and diagnosis.
0:09:06 Anomaly Detection Challenges: Challenges in this field include the vast amount of time-series data generated by drones, limited knowledge shared in flight controller documentation, lack of annotated datasets, and the difficulty in distinguishing anomalies from normal variations.
0:10:52 RESAM Framework: Dr. Islam introduces RESAM, a requirements elicitation process that leverages multiple data sources to specify functional requirements for deep learning-based anomaly detectors.
0:11:14 Deep Learning-Based Detectors: RESAM was used to build LSTM autoencoders for detecting vibration, attitude, and interference anomalies, outperforming rule-based methods.
0:13:25 Resource Constraints: Real-world drones have limited computational resources and power, making it challenging to deploy heavyweight deep learning models.
0:14:49 Heuristic-Based Detectors: To address resource constraints, seven heuristic-based anomaly detectors were developed, offering a lightweight alternative to deep learning.
0:15:31 System Architecture: An architecture was devised to reliably transmit data from the flight controller to the companion computer, where anomaly detectors reside, using the MAVLink and ROS protocols.
0:16:47 Simulation and Field Tests: The detectors were tested in the Gazebo simulator and during real-world flights, demonstrating their effectiveness in detecting anomalies.
0:18:04 Atmospheric Effects Consideration: Discussion regarding the consideration of atmospheric effects on drone electronics and data reliability. The current anomaly detectors do not take atmospheric data into account, but future research will.
0:22:56 CADA Framework: CADA, a framework for detecting and adapting to parameter misconfigurations, was introduced. It can revert parameters to safe values upon anomaly detection, potentially preventing crashes.
0:28:37 Resource Optimization: The need for lightweight anomaly detection methods led to the development of ADAM (Adaptive Drone Anomaly Monitor).
0:29:02 ADAM Framework: ADAM employs lightweight "canary" detectors that trigger more specialized detectors upon anomaly detection, reducing computational load.
0:30:07 Activation Sequence Tree: ADAM uses a hierarchical tree structure to activate detectors based on anomaly detection by parent nodes.
0:33:31 Evaluation of ADAM: ADAM was evaluated against a non-adaptive approach, demonstrating comparable detection accuracy with significantly reduced resource consumption.
0:36:51 Anomaly Diagnosis: The goal of diagnosis is to determine the root cause of anomalies based on detected symptoms.
0:37:55 Diagnosis Methodology: A process was developed to identify common symptoms and root causes, generate search terms, and mine association rules from forum data to recommend potential root causes.
0:39:56 Evaluation of Diagnosis Tool: The tool was evaluated on unseen forum data, demonstrating its ability to recommend correct root causes with high accuracy.
0:41:55 Case Study: A real-world flight incident was analyzed using the anomaly detection and diagnosis tools, demonstrating their ability to identify potential root causes consistent with expert opinions.
0:43:55 Data Collection Project: A longitudinal data collection project was undertaken to gather comprehensive flight data, including sensor readings, weather information, battery status, and post-flight analysis reports, to support further research.
0:46:17 Anomalous Flights: Two anomalous flights from the data collection project were described, highlighting the importance of geofence settings and the challenges of manual control during unexpected events.
0:49:02 Conclusion: Dr. Islam thanks his collaborators and invites questions from the audience.
0:49:55 Q&A: Discussion regarding the statistical measures used to assess model performance, applicability of anomaly detectors to different drone types, and the use of autoencoders for anomaly detection.