A Monograph Series

Estimation and learning for vehicle health

Four short, interactive monographs on the same deep question, how to reason about what a machine will not tell you directly. Each builds its methods from scratch, verifies the math, and runs live in the browser. Together they trace a single arc from assuming a model, to learning one, to generating from one.

The map

When to reach for which

The four pieces are not alternatives so much as complementary tools, sorted by how much they assume versus how much they learn, and by what they give you back. A rough guide:

assumes a known model ←     → learns from data point estimate ←  → full distribution Kalman / EKFknown model, optimal state Particle filterknown model, any belief PINNequations + data, fitted model Diffusionlearned distribution, generative
The methods at a glance. Filters sit on the left: they assume you already know the dynamics and return the best estimate of the hidden state. The PINN moves right by learning the model while respecting physics. Diffusion sits top-right, learning an entire distribution it can sample from and score against.
The series

Four monographs

The through-line

One idea under all four

Each method, in the end, measures the same thing: the gap between what is expected and what is observed. The filter calls it the innovation, the residual between prediction and measurement. The PINN calls it the physics residual, how far a candidate solution strays from the governing equation. The diffusion model calls it the denoising error, how unlike the healthy data a signal looks. In every case that gap is both the signal used to correct the estimate and the alarm that announces a fault. Read in order, the series is a tour of four ways to compute it, and four ways to put it to work keeping a vehicle honest about its own health.