Concepts to code

Paper ideas → the function that implements them. Pair with DATA_DICTIONARY.md. Paths are under src/fdia_graph/.

Modules

The paper's math lives in src/fdia_graph/engine/; at the top level, generation.py/profiles.py drive the engine and the rest load and serve data.

SDK file	job
registry.py	dataset versions, aliases, cache
download.py	fetch + cache a shard
generation.py	assemble the classification shard (the recipe)
streams.py	assemble a continuous timeline
dataset.py	loader → tensors / PyG (what fg.load returns)
profiles.py	real load series → operating points

engine/ file	formula it implements
core.py	FdiaGenerator: grid + noise setup (__init__), attack targeting; composes the three mixins
measurement.py	emit_from_state (the measurement function h(x)), emit, state_from_net
physics.py	solve / resolve_states — AC re-solve under new loads
attacks.py	corrupt (Ad/As/Ar) and lra_delta (Al redistribution)

State estimation

• WLS x̂ = argmin (z−h(x))ᵀW(z−h(x)): h(x) = engine/measurement.emit_from_state; full solvers ship in fdia_graph.se (WLS and the robust/prior classes) — see ../guides/state_estimation.md.
• Bad-data r_i=(z_i−h_i)/σ_i, J=Σr_i²: σ_i from the engine FdiaGenerator.SD; pass/fail stored as the stealthy flag.
• Noise: engine FdiaGenerator.SD (accuracy-class), each reading = true + per-meter bias + per-scan jitter.

Attack families

family	paper	build	code
Aq	A_o	scale load, re-solve	generation.make (1) + engine/physics.solve
At	A_t	slow ramp, re-solve	generation ramp loop (5)
Al	A_l	load-conserving redistribution	engine/attacks.lra_delta (6)
Ad	A_d	z ← z(1±u)	engine/attacks.corrupt
As	A_s	z ← βz	engine/attacks.corrupt
Ar	A_r	replay z(t−k)	engine/attacks.corrupt

Aq/At/Al re-solve power flow so readings stay a consistent AC state (invisible to the residual test by construction). Ad/As/Ar tamper readings directly (detectable). All changes kept in a 2–20% plausibility band (generation.NOISE_FLOOR, attack_intensity).

Temporal feature

temporal_delta = scan-to-scan [ΔP,ΔQ]; swing = that as a z-score of recent volatility. Built in generation._fin and streams._store. Any above-noise attack spikes the swing, so localization is per-bus and needs no graph; the slow ramp At stays inside the swing, so it is the open case.

Metrics

Localization macro-F1 (per-bus F1 over attackable buses); DR/FA (always report FA with DR); strict localization accuracy. Runnable examples that compute these are in the top-level README.md.