systems theory//engineering patterns

Engineering patterns are recurring shapes of reasoning that appear under different names across sensing, estimation, learning, diagnosis, decision and control, and they are used to orient quickly in a problem no textbook covers: a pattern recognised is worth more than ten algorithms memorised, because it says which question to ask before any method is chosen. The Kalman filter is the standard example of why they matter. It weighs a prediction against a measurement, it is the dual of the LQR, it lives on its innovation, it relinearizes in its extended form, it is recursive least squares and it carries its own covariance: six patterns in one method, and a method that fits so many tends to last.

Engineering patterns are recurring shapes of reasoning that appear under different names across sensing, estimation, learning, diagnosis, decision and control, and they are used to orient quickly in a problem no textbook covers: a pattern recognised is worth more than ten algorithms memorised, because it says which question to ask before any method is chosen. The Kalman filter is the standard example of why they matter. It weighs a prediction against a measurement, it is the dual of the LQR, it lives on its innovation, it relinearizes in its extended form, it is recursive least squares and it carries its own covariance: six patterns in one method, and a method that fits so many tends to last.

The patterns also give an order of questions for any new system. Is what you want to know observable from what you measure? What fits on the platform, within its cycles, memory, energy, latency and certification? Only then, which algorithm; start with the heuristic, and centralize while you can. A pattern is a question and never a hammer: everything is optimization does not mean every problem should be posed as one, and estimating is averaging does not make every average a good estimator.

Estimation and its evidence. Almost every estimator is a reliability-weighted mean, and what differs is where the weights come from (weighted averaging). Measured minus expected drives both correction and detection (residual as surprise). A number without an error bar cannot support a decision (uncertainty as first-class). And if what you seek leaves no trace in the measurements, no filter will find it; the fix is another sensor or another experiment (observability before the algorithm).

The mathematics underneath. Estimation and control are the same equations with the matrices transposed, so learning one side gives the other (estimation-control duality). Least squares, Kalman, LQR, MPC, training and assignment all minimize a cost and differ in which cost is reasonable to write (optimization). Nonlinear problems are solved as a sequence of linear ones (linearize-solve-repeat). A question about a system becomes a matrix whose eigenvalues answer it, and the skill is choosing the matrix (spectral reading); one matrix, D−AD-AD−A, turns up in heat diffusion, fleet consensus, clustering, graph networks and random walks (graph Laplacian).

Time. Every millisecond between measuring and acting costs phase (delay as the enemy). Systems that work stack loops by time scale, each treating the one inside as perfect and the one outside as constant (hierarchical control). What sinks a system lives in the tail of the distribution, so design with percentiles and worst cases (tails over means).

Design choices. Error is never removed, only moved, and a budget says where it fits (error relocation). The simplest tool that solves the problem with margin is the baseline and plan B (flyswatter rule). The best algorithm is the best one that fits the platform (embedded system). Physics and data compete and combine, and the industrial frontier is the hybrid (model versus data).

Architecture. Everything that senses and acts is a loop, and a weak stage cannot be rescued by the best algorithm elsewhere (closed-loop system). A centre that sees everything optimizes, debugs and certifies better, so distribute only what must survive without a link or react faster than the round trip (fleet architecture).

Two frames place any of these in practice: the six context levels from physics to business at which a method lives, and its technology maturity in the domain at hand. The book adds two observations that live in their own notes: a sampled system is a different system from the continuous one (discretization), and feedback forgives a crude model, but only for what is slower than the loop and arrives in time.

The book closes its list with a warning it calls same name, different animal: words are reused, assumptions are not. Control consensus averages a continuous value with neighbours, approximately, tolerating loss and delay, while Raft votes a discrete value exactly among servers. Trajectory tracking makes a vehicle follow a path; multi-target tracking follows other objects with sensors. The state of a state-space model is physical; the state of a Markov decision process is whatever the decision needs. A digital filter shapes a signal's spectrum; an estimation filter infers a hidden state. SIL is both software-in-the-loop and safety integrity level. When two things share a name, ask what each one guarantees.