ARADISS

Adaptive Real-time Anomaly Detection & Identification in Sensored Systems

On Earth, in space, and across industries, there is a growing variety of manned and unmanned vehicles that use batteries as their primary power source. These vehicles are composed of a wide array of interacting components and sensors that are needed to safely execute their respective missions. As their interactions, complexity, and numbers increase, the risk for issues such as degradation of components, sensor faults, and erroneous controls also increase. These anomalies pose significant risks for vehicles conducting critical missions, those operating in hard to reach areas, or for aerial vehicles flying over densely populated areas. It is therefore crucial to detect and mitigate these anomalies.

There exist several approaches for anomaly detection, such as traditional rule or threshold-based methods, model-based approaches, supervised machine learning-based methods, and even unsupervised methods to detect different types of abnormal behaviors. These methods have inherent drawbacks including lack of sensitivity, inability to detect previously unknown faults, not being robust to compromised in-network information, or requiring sophisticated system models.

ARADISS (Adaptive Real-time Anomaly Detection and Identification in Sensored Systems) is a new method of fault detection for cyber-physical systems, such as satellites, UAVs, and automobiles. It uses machine learning models to learn physics-based dependencies, and an unsupervised algorithm to detect and identify anomalies. It is inspired by the physical dependencies between a vehicle’s operation and the associated power consumption, allowing the use of the battery as a trustworthy sensor to detect anomalies in a vehicle’s operation (a hardware-based root-of-trust).

ARADISS utilizes features extracted from run-time battery voltage and current information to construct learning models (norm maps) that map dependencies independently between battery metrics and each system operational variable. Anomalies are detected when deviations in this trend are observed, which are quantified using five key parameters of the unsupervised anomaly detection algorithm.

This approach is applicable to most systems with electric batteries and can be rapidly optimized and adapted for efficient and cost-effective onboard fault management.

The algorithms are self-adapting to gradual system changes, allowing for highly sensitive and accurate diagnostics.  ARADISS is a cost effective fault management solution for customized platforms, and allows for rapid implementation of mitigation strategies.

Approach

  • ABSTRACTION: A system’s normal behavior is abstracted as a digital twin using machine learning models connecting each pair of system variables
  • DETECTION: Model predictions from a ground truth variable are compared to observed behavior to flag anomalies/faults
  • IDENTIFICATION: Flagged anomalies are verified using additional models and fault isolation is carried out using series of edge checks following graph techniques

Benefits

  • ARADISS is cheaper and faster to implement than traditional fault management while being transparent.
  • It provides comprehensive and robust fault coverage
  • It can be rapidly and automatedly adapted to new platforms
  • Low computational requirements
  • On-board or off-board (ex: ground control, telemetry-based) deployments
  • Improves autonomy
  • Can act as a “second line of defense” on top of traditional fault detection for highly risk-averse missions

Value Propositions

Onboard Deployment
    ⦿ For deep space and remote terrestrial missions 🡒 Increase autonomy

Ground-based monitoring
    ⦿ For near-Earth missions 🡒 Reduce human effort

AIT Support
     ⦿ During mission development 🡒 Faster testing and validation

Post-Event Analysis
     ⦿ For analysis of safety event data 🡒 Rapid insights/root-cause identification

 


Applications

  • Small Satellites
  • Unmanned Aerial vehicles
  • Unmanned Ground Rovers
  • Underwater underwater/surface vehicles
  • Automobiles