Collaborative Research:SHF:Medium:Machine Learning on the Edge for Real-Time Microsecond State Estimation of High-Rate Dynamic Events

点击翻译按钮获取中文摘要

Computer control of dynamic systems from the manufacturing, robotics, and aviation fields traditionally operate on timescales of 10s or 100s of milliseconds. For example, an avionics system traveling at 1000 kilometers per hour and operating at 10 milliseconds per control decision will move three meters in the time allocated to each control decision. However, emerging hypersonic, space, and military systems require active control while operating at extreme velocities or while being subjected to accelerations or decelerations caused by explosions or high-speed collisions. These applications require control at timescales on the order of microseconds. Making control decisions for such systems often requires that the controller estimate the state of the system from indirect measurements such as vibration. Traditional methods for state prediction are based on first principles using finite element analysis (FEA), whose execution time scales as a square of the number of elements. This makes it impractical to evaluate FEA models at microsecond timescales. Models derived from machine learning can estimate the state of the system based on pre-curated datasets and require less workload as compared to an equivalent FEA model. Such models, when combined with domain-specific processors, could provide equivalent accuracy with higher throughput than FEA models, making microsecond-scale state modeling possible. However, there are currently no suitable development methodologies for systematic generation of machine-learning models at such extreme performance constraints. The objective of this research is to develop a structural model compiler that meets a given accuracy constraint, as well as a corresponding overlay generator on which the generated model meets a given microsecond-scale latency constraint. This research will advance the fundamental knowledge and skills required for the real-time decision-making and control of active structures that experience high-rate dynamic events.This project addresses two distinct but synergistic problems: (1) technologies to enable real-time decision-making and control of active structures that experience dynamic events at the microsecond timescale and (2) development of tools for optimization and synthesis of domain-specific processors for trained models. Recent academic and industrial work focusing on development of specialized architectures for evaluating Long Short Term Memory (LSTM) models generally yield “one-off” designs tuned to a specific Field Programmable Gate Array (FPGA)--often a server class FPGA--and have rigid, “baked in” design decisions. This makes it difficult to compare alternative or competing optimization techniques for a desired target FPGA platform. To solve this, this project is developing a generalized programmable processor architecture that incorporates a repertoire of optional features designed to accelerate specific aspects of LSTMs and support associated model optimizations. The architecture is both programmable and customizable, allowing it to serve as a common platform for evaluating different approaches for accelerating LSTM models. Concurrently, the investigators are developing a set of benchmark datasets for structural state estimation with accuracy and performance requirements. The project is also developing useful artifacts for subsequent research in edge-based machine learning, including a method for comparing different LSTM model-pruning and compression approaches and comparing different microarchitecture designs. Code and hardware designs developed from this project are open-source.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.