TL;DR
This paper addresses the critical need to accurately model battery behavior in formal verification of autonomous robots. By integrating realistic battery features like capacity fade and non-linear discharge rates into probabilistic model checking, they demonstrate how overlooking these aspects can lead to significantly misleading safety and reliability assessments for missions involving recharging.
What Problem Does This Paper Address?
Autonomous robots, especially in challenging environments like offshore wind farms, rely heavily on batteries. However, current formal verification methods often simplify battery models, treating them as generic energy sources. This simplification overlooks crucial battery characteristics such as capacity fading over time and the non-linear way discharge rates change as the battery's state of charge (SOC) decreases. These simplifications can lead to overly optimistic or pessimistic predictions about mission success, safety, and required mission time, potentially causing significant financial losses and safety risks. The paper aims to bridge this gap by showing how incorporating realistic battery behavior, informed by Battery Prognostics and Health Management (PHM) principles, can lead to more accurate and reliable formal verification of autonomous robot missions.
Methods
The authors employ probabilistic model checking using the PRISM tool. They formalize a specific use case: an unmanned aerial vehicle (UAV) inspecting an offshore wind farm. This involves creating a model composed of four interconnected modules:
- Drone Module: Represents the UAV's finite state machine, detailing its actions like take-off, flying, inspection, and returning to base.
- Grid Module: Models the wind farm layout as a grid, defining the inspection route and the number of turbines to be checked in each cell.
- Environment Module: Captures environmental dynamics, specifically modeling two levels of wind speed and their probabilistic transitions, which influence power demand.
- Battery Module: This is the core of their contribution. It explicitly models two key battery features:
- Capacity Fading: The battery's maximum capacity decreases with each recharge cycle.
- Non-linear Discharge Rate: The discharge rate increases as the battery's SOC decreases for a constant power demand.
They use Probabilistic Computation Tree Logic (PCTL) to specify properties of interest, such as mission success probability and expected mission time.
Datasets
Not discussed in the paper. The paper uses a hypothetical/generic battery PHM method to provide parameters for their model.
Key Insights and Ideas
- Battery Features Matter: Overlooking battery capacity fade and non-linear discharge rates can lead to significantly inaccurate formal verification results, potentially by orders of magnitude in mission time.
- Interplay of Factors: Battery features, dynamic environments (wind speed), and battery safety strategies (e.g., minimum SOC thresholds) are not independent. They jointly influence the overall mission outcome and verification results.
- Risk vs. Reward in Battery Safety: A more conservative battery safety strategy (higher safe SOC threshold) leads to higher mission success probability but also increases mission time and recharges. Conversely, a riskier strategy can reduce mission time but may decrease success probability.
- Simplified Models Can Be Misleading: Even "basic" models that account for some battery aspects (like voltage levels) but not others (like capacity fade) can still produce misleading results, especially in scenarios with multiple recharges.
Key Results
The paper presents verification results for a typical UAV inspection mission. They compare four scenarios:
- Case #1 (Baseline): Safe battery strategy (safe_t=0.3), moderate environment (p_wsp_c=0.1). Results: 100% mission success, ~4700 min mission time, ~42 recharges.
- Case #2 (Risky Strategy): Lower safe SOC threshold (safe_t=0.25), moderate environment. Results: 91% mission success, ~3886 min mission time, ~34 recharges. (Demonstrates risk-reward trade-off).
- Case #3 (Dynamic Environment): Safe battery strategy, more dynamic environment (p_wsp_c=0.3). Results: 100% mission success, ~7483 min mission time, ~72 recharges. (Shows dynamic environment significantly increases mission time).
- Case #4 (Risky Strategy + Dynamic Environment): Lower safe SOC threshold, more dynamic environment. Results: 89% mission success, ~5622 min mission time, ~53 recharges. (Highlights potential for reduced success with combined risks).
They also show that simplified battery models can overestimate or underestimate mission time by 1-3 working days compared to their "advanced" model, which can lead to significant economic losses.
Limitations and Future Work
Limitations:
- Simplified Battery Model: While more advanced than previous work, the battery model still uses simplified assumptions, such as two levels of wind speed and rounded estimations of battery consumption.
- Hypothetical PHM: The paper uses a hypothetical/generic battery PHM method for parameter generation, rather than integrating a real-time PHM system.
- Focus on Out-of-Battery Failure: The verification primarily focuses on the "out-of-battery" failure mode, simplifying other potential failure probabilities to zero.
- PRISM Limitations: PRISM's ability to solve "reachability reward" properties is limited to target states reached with probability 1, requiring careful selection of target states for numerical examples.
Future Work:
- Integrate Real-time PHM: The authors propose integrating battery PHM techniques with formal verification at runtime, or at least during recharging at the base, using newly collected log data.
- Refine Models: Further refine the PRISM model based on more detailed observations and data.
- Scalability: Address the scalability challenges of running online PHM and formal verification algorithms concurrently.
- Extend to Other Robots/Environments: Apply the developed framework to other types of autonomous robots and more complex environments.
Summary updated