TinyML-UUVs: Tiny Machine Learning for Low-Power Unmanned Undersea Vehicles
Abstract
Research Problem and Objectives:The success of modern machine learning (ML) has revolutionized many fields in the past decade. Howev,er, modern ML methods require estimating millions or even billions of parameters for accurately learning a neural network model. Thi,s not only consumes much power and energy of Unmanned Undersea Vehicles (UUVs) but also takes a long time to train the ML models in,real-world underwater environments. UUVs generally have the low processing power, limited memory and data storage, low bandwidth wit,h restricted and unreliable communications, and a limited amount of battery capacity.We will investigate the UUV as an Intelligent A,utonomous System (IAS) with the objective of lowering its consumption of power, energy, and resources without sacrificing the perfor,mance of its ML tasks.Technical Approaches:Compared to regular ML without constraints of computing resources, the proposed tiny mach,ine learning (TinyML) is expected to significantly reduce the power consumption and computing resources required to run ML applicati,ons without noticeable performance loss. Although TinyML has been successfully applied on Internet-of-Things recently, it has never,been studied for UUVs. Unlike the existing applications of TinyML for saving power of general devices (e.g., smartphones), our proje,ct will specifically integrate TinyML with UUVs? onboard configurations so that the successful project will gain better insights int,o TinyML?s effects on UUVs. We will investigate lowering power consumption from three sub-components: quantization methods, model co,mpression, and unsupervised learning.Anticipated Outcomes:Given a successful completion, the outcomes of the successful project will, have five folds: 1) TinyML-UUVs will expand the survey areas for long endurance operations with less power consumption; 2) TinyML-U,UVs will quickly adapt to dynamic, unstructured, and uncertain environments with minimal bandwidth in communication; 3) TinyML-UUVs,will have ubiquitous and persistent data collection and overwhelming data analysis capability for accurate situational understanding,; 4) Deep understanding of the TinyML?s effects on sonar image analysis from UUVs will be gained; 5) A Ph.D. student will be trained, for driving the Naval research innovation with the next-generation unmanned systems.Impact on DoD Capabilities:Efficient ML/AI solu,tions of the project will realize the full potential of the Department of the Navy?s IAS strategy and Unmanned Campaign Framework be,cause efficient ML/AI methods running on the edge will be capable of online learning. In addition, the developed methods will run ef,ficiently on small-unmanned system platforms, which can support disruptive multi-vehicle teams using low-cost, energy-efficient comp,uting hardware with a small physical footprint. The success of the proposed work can ease the workload of the Unmanned Undersea Squa,dron human operators and reduce mission time to completion for Mine Counter Measure, Seabed Warfare, and High-Value Asset protection,. The proposed work will contribute to advances in UUV capabilities that will enable UUVs to be deployed in more comprehensive conte,sted and resource-constrained scenarios. This project will strengthen the overall level of research under Navy Science and Technolog,y.This project abstract is marked as ?Approved for Public Release?.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Dec 06, 2022
- Source ID
- N000142312123
Entities
People
- Yuchou Chang
Organizations
- Office of Naval Research
- United States Navy
- University of Massachusetts