Introduction
Unmanned Underwater Vehicles (UUVs) are increasingly adopting bio-inspired flapping-fin propulsion systems to achieve superior agility in complex environments, such as near-shore navigation and station-keeping. However, controlling these systems presents significant challenges. While predicting the thrust generated by a specific fin movement (the "forward" problem) is straightforward, determining the optimal movement to achieve a desired thrust (the "inverse" problem) is computationally intensive. Furthermore, these vehicles operate on limited battery power and low-resource onboard hardware, necessitating control strategies that are both power-efficient and computationally lean.
System Architecture
To address these challenges, this research proposes a "power-aware" inverse-search machine learning framework designed for real-time execution on edge hardware like the Raspberry Pi 4B. The architecture consists of two primary components: a forward surrogate model and an inverse search optimizer.
The forward model utilizes a Long Short-Term Memory (LSTM) neural network trained to predict thrust and power consumption based on gait parameters (stroke, pitch, and frequency). The inverse search component then uses this LSTM to evaluate potential gaits against a multi-objective cost function:
$$Cost = w_T(Thrust) + w_P(Power) + w_K(Kinematic Smoothness)$$
By adjusting the weights ($w_T, w_P, w_K$), the system can dynamically prioritize different performance metrics, such as maximizing thrust for rapid maneuvers or minimizing power for extended endurance.
Experimental Setup and Results
The system was evaluated using a dataset of 864 unique gaits across three fin designs: a standard rigid fin and two flexible polymer (PDMS) fins with varying ratios (1:10 and 1:20). Benchmarking on a Raspberry Pi 4B demonstrated that the inverse search could be completed in under 0.5 seconds, meeting the real-time requirements for a 2 Hz flapping cycle.
A key finding was the non-linear relationship between thrust and power. The results indicated that a minor sacrifice of 0.1 N in thrust could yield a substantial power reduction of 3 Watts, highlighting the inefficiency of "max thrust" gaits. Additionally, flexible fins (PDMS) were found to be more conducive to multi-objective optimization, showing lower thrust losses when power and smoothness were prioritized compared to the rigid design.
Conclusion
This work demonstrates a feasible approach for implementing sophisticated, multi-objective machine learning control on low-resource UUV hardware. By integrating power awareness into the control loop, significant energy savings can be achieved with minimal performance impact. Future work will focus on physical validation in water tank environments to confirm that these predicted efficiency gains translate to extended operational battery life in real-world conditions.
Summary updated