Prognosis Framework

Particle Filters are a novel class of non-linear filters that combine Bayesian learning techniques with importance sampling to provide good state tracking performance while keeping the computational load tractable. The idea is to represent the system state (in this case the battery SOC or voltage or capacity) as a probability density function (pdf) that is approximated by a set of particles (points) representing sampled values from the unknown state space, and a set of associated weights denoting discrete probability masses. The particles are generated from an a priori estimate of the state pdf, propagated through time using a nonlinear process model, and recursively updated from measurements through a measurement model. The main advantage of PFs here is that model parameters can be included as a part of the state vector to be tracked, thus performing model identification in conjunction with state estimation. After the model has been tuned to reflect the dynamics of the specific system being tracked, it can then be used to propagate the particles till the failure (e.g. EOD or EOL) threshold to give the RUL pdf.

Sample EOD prediction:

Sample EOL prediction: