TL;DR
This paper introduces a novel fault-adaptive autonomy architecture for autonomous underwater vehicles (AUVs) that leverages Learning-Enabled Components (LECs) while ensuring safety through an Assurance Monitor (AM). The system effectively detects and mitigates actuator faults in real-time, allowing the AUV to continue its mission even under degraded conditions, significantly improving robustness and reliability.
What Problem Does This Paper Address?
Autonomous Cyber-Physical Systems (CPS), particularly those used in hazardous environments like underwater vehicles, face a critical challenge: maintaining safe and reliable operation despite potential failures. These failures can range from physical degradations to software issues. Learning-Enabled Components (LECs), often based on deep learning, offer powerful capabilities for perception and decision-making but introduce uncertainty due to their reliance on training data and potential for out-of-distribution behavior. The problem is how to integrate these powerful but potentially unreliable LECs into safety-critical systems and ensure that faults are detected, understood, and managed effectively to prevent mission failure or hazardous situations. Specifically, this paper focuses on actuator faults, which can significantly impact vehicle dynamics and safety.
Methods
The core of the proposed solution is a fault-adaptive autonomy architecture built around a Behavior Tree (BT) for high-level mission control. Key components include:
- Fault Detection and Isolation (FDI) LEC: A Deep Neural Network (DNN) trained to detect and classify thruster faults and their degradation levels. This LEC takes thruster commands and actual RPMs as input.
- Assurance Monitor (AM): This component, based on Inductive Conformal Prediction (ICP), quantifies the trustworthiness of the FDI LEC's output. It provides "credibility" and "confidence" metrics for each classification.
- Assurance Evaluator (AE): This module uses the credibility and confidence metrics from the AM to make a final decision on whether the FDI LEC's classification is trustworthy. It employs a selective classification approach, deciding whether to accept the classification or flag it for further investigation.
- Behavior Tree (BT): The BT orchestrates the overall mission execution, integrating the FDI LEC, AM, and AE. It uses the outputs of these components to trigger contingency actions, such as control reallocation, when a fault is detected and deemed trustworthy.
- Thruster Allocation Matrix (TAM): Used for control reallocation, this matrix helps re-distribute thrust among the remaining healthy thrusters to compensate for the loss of a faulty thruster, maintaining stability and allowing mission continuation.
Datasets
The paper describes a data generation process using a Gazebo-based simulation of a BlueROV2 underwater vehicle.
- Training Data: Generated from multiple simulated waypoint-following missions with various thruster degradations. The initial 20 seconds of each simulation were excluded to account for system boot-up. The data was split into an 80% training set and a 20% calibration/validation set.
- Evaluation Data: Further simulations were conducted with different random seeds and scenarios, including hazardous environments with obstacles, to evaluate the system's performance.
Key Insights and Ideas
- Bridging the Gap Between LECs and Assured Autonomy: The paper effectively demonstrates how to integrate powerful but inherently uncertain LECs into safety-critical systems by augmenting them with an assurance mechanism. This is crucial for deploying AI in real-world applications where reliability is paramount.
- The Power of "Trustworthiness" Metrics: The use of credibility and confidence metrics from ICP provides a quantifiable way to assess the reliability of an LEC's output, allowing for informed decision-making about whether to act on a detected fault.
- Real-time Fault Mitigation: The architecture enables real-time fault detection and control reallocation, allowing the AUV to adapt to degraded conditions without necessarily aborting the mission. This significantly enhances operational resilience.
- Behavior Trees as an Orchestration Framework: BTs prove to be a flexible and effective tool for managing complex mission logic and integrating fault-adaptive behaviors, making the system more modular and maintainable.
Key Results
- High Recall and Accuracy: The proposed system, with the FDI LEC, AM, and AE, achieved a recall of 98.37% and an accuracy of 93.85%. This is a significant improvement over a raw LEC without assurance (recall 84.05%, accuracy 84.05%) and even a raw LEC with a credibility threshold (recall 91.64%, accuracy 92.37%).
- Reduced Rejection Rate: The system was able to accept a higher percentage of correct LEC outputs (12.54% rejected) compared to a raw LEC with a SoftMax threshold (81.24% rejected), indicating better decision-making.
- Successful Mission Completion: In waypoint scenarios, the AUV with fault-adaptive autonomy successfully completed missions even with severe degradations, whereas without the FDI system, missions failed in 27 out of 60 cases.
- Efficient Reallocation: The average reallocation time was around 2.42 seconds, indicating that the system can detect and react to faults quickly.
Limitations and Future Work
- Simulation-Based: The current evaluation is based on simulations. Transitioning to real-world hardware and validating the system's performance in actual underwater environments is a crucial next step.
- Focus on Actuator Faults: The paper primarily addresses actuator faults. Future work could extend this to other types of faults, such as sensor failures or software anomalies.
- Limited Fault Types: While the system handles various degradation levels, the number of distinct fault classes is finite. Exploring more complex and emergent fault behaviors could be beneficial.
- Future Extensions: The authors plan to explore further contingency possibilities and expand the safety of the autonomous systems for the BlueROV2 AUV.
Summary updated