Introduction
Cyber-physical systems (CPS) are integration of computational and physical processes. Typical examples of CPS are process control applications, where physical plants are controlled with signals computed in embedded systems through the feedback mechanism. In these applications it is important to detect and correct process perturbations in a timely manner in order to prevent defects in the final products. This is a complex task that involves precise knowledge of the process, in the form of a model and its current state.
In industry, considerable effort is dedicated to the development of physics-based models of processes. Such models are commonly used in offline process optimization. In addition to that, these models can also be used to reconstruct the probability density function of the state of the process by combining the stochastic model with in-process measurements following a Bayesian framework. One way to perform these calculations in nonlinear systems is with Sequential Monte Carlo (SMC) methods. Since their introduction over 20 years ago, the particle filter—a well-known SMC method—has been used in state and parameter estimation in a broad spectrum of applications, such as computer vision and financial econometrics. A major drawback of particle methods is the computational cost they incur. When using SMC methods for online estimation in CPS, it is important that computation and data communication are performed much faster than the sampling rate of the process. The speed of particle filters and similar methods is greatly enhanced when implemented on parallel computer architectures, such as general-purpose graphics processing units (GPGPUs). GPGPUs allow massive parallelization by distribution of independent tasks over several computing units. However, there is a long way to go from accurate CPS model development to particle filter estimator implementation on GPUs. The automated particle filter design system (APFDS) project aims to bridge such a gap by building a real-time automated tool on CPU+GPU architecture realizing systmatic design of particle filter estimator for CPS, efficient implementation on GPUs, and real-time analysis of the application. This project is on going.Publication
Conference:
-
"Implementation of a particle filter on a GPU for nonlinear estimation in a manufacturing remelting process",
Felipe Lopez, Lixun Zhang, Joseph Beaman, and Aloysius K. Mok,
2014 IEEE/ASME Conference on Advanced Intelligent Mechatronics (AIM), 2014, 340-345.
Journal:
-
"Particle filtering on GPU architectures for manufacturing application",
Felipe Lopez, Lixun Zhang, Joseph Beaman, and Aloysius K. Mok,
Computers in Industry, 2015(71): 116-127.