Hertz

Open-source conducted-EMI analysis and EMI-filter design toolkit for power electronics. Sibling of Kirchhoff (in-process SPICE) and Kelvin (parts librarian) in the OpenConverters ecosystem.

Hertz turns the mathematics of the conducted-emissions world — LISNs, CISPR receiver detectors, limit lines, CM/DM separation, line-filter synthesis — into tested, reusable code. It is vendor-neutral by design: synthesis works over any manufacturer's component catalog supplied as candidate lists.

Pre-compliance estimates only. Nothing here replaces accredited chamber testing, and no output of this library is a compliance statement.

Modules (v0.1)

Module	What it does
hertz.limits	Limit lines (CISPR 32 Class A/B mains, CISPR 25 classes 1–5 conducted) with log-frequency interpolation. Frequencies outside a line's coverage raise — CISPR 25 defines limits only inside protected bands, and "no limit here" is information.
hertz.lisn	50 µH (CISPR 16) and 5 µH (CISPR 25) LISN models: EUT-side impedance vs frequency and SPICE subcircuit export.
hertz.detector	CISPR 16-1-1 measuring-receiver emulation on sampled data: Gaussian-window STFT envelope + quasi-peak charge/discharge detector + critically damped meter (Krug & Russer, IEEE TEMC 2005). Bands A–D constants included.
hertz.separation	Exact CM/DM separation from the two LISN line signals. Deliberately refuses magnitude-only spectra (phase is required — that is physics, not a missing feature).
hertz.filter_design	Line-filter synthesis per Würth Elektronik's public application note ANP015, generalized over the line count: single-phase L/N, DC supply pair, 3-phase 3-wire (delta X, each pair sees 1.5·C) and 3-phase + neutral (star X). 1- and 2-stage CM+DM sizing, component rounding onto explicit candidate lists, Y-cap leakage current at phase-to-earth voltage, X-cap discharge resistor. Validated against the app note's worked example.
hertz.traces	Spectrum-analyzer CSV ingestion with unit detection; ambiguous units raise instead of guessing.

Architecture

Two implementations of the same engine, sharing one set of golden test vectors:

• cpp/ — the product core (C++20, header-only): compiles natively, to WASM (browser tools run entirely client-side — no simulation servers), and to Python via pybind11 (PyHertz). Reuses the OpenMagnetics MKF library (its radix-2 FFT today; its wideband choke impedance model next), so build MKF first and point HERTZ_MKF_ROOT at the checkout.
• src/hertz/ — the Python reference (numpy): the readable spec and cross-validation implementation; every C++ behavior is asserted against the same numbers here.
• src/hertz_cpp/ — the same API on the C++ engine: a drop-in replacement for hertz that calls PyHertz. It adapts shapes only (numpy in/out, tuple returns, path-taking CSV readers) and reimplements no formula, so the reference suite can be pointed straight at the C++ engine.

CSV parsing is string-based in the C++ core by design: the host (browser File API, Python open()) reads the file and passes its content — the WASM-native pattern.

Install & test

# Python reference
pip install -e ".[dev]"
pytest

# C++ core (Catch2 binary, run it directly — no ctest)
cmake -S cpp -B cpp/build -G Ninja -DCMAKE_BUILD_TYPE=Release
ninja -C cpp/build -j4
./cpp/build/test_hertz

# The SAME 92 tests against the C++ engine through PyHertz (built by the
# command above; -DHERTZ_BUILD_PYBIND=OFF skips it). The tests are the golden
# vectors shared by both implementations, so a failure here is a real
# disagreement between the two engines — never something to silence.
pytest --backend=cpp

PyHertz mirrors src/hertz/*.py name for name, plus the surface the Python reference does not carry: SPICE deck emission (filter_spice_deck, lisn_reference_deck, deck_abcd_il), the radiated CM-attenuation target, the cable-ferrite picker, and limit_polyline_runs.

Web GUI (web/)

The full engine is exposed as a browser instrument — Vue 3 + Vite + the WASM engine, zero server compute — deployed at hertz.openconverters.com. Two destinations, because the workflow is measure-then-design: Spectrum (scan vs limit verdict, worst offenders, required attenuation, handoff to the designer) and Filter — the ANP015 designer with any manufacturer's candidate lists + SPICE export, laid out as a workbench: an input rail on the left, two independently switchable output panes on the right.

Everything else is a pane of that bench rather than a screen of its own, because each one is an input to a filter design or a reading of one, and answers little on its own: MEASURE · SCOPE CAPTURE (CISPR 16-1-1 peak/quasi-peak/average emulation of an uploaded waveform, which hands per-mode CM/DM targets straight to the rail), MEASURE · CM PROBE (radiated pre-scan from a CM-current capture, plus the cable-ferrite mitigation picker), TEST SETUP (LISN) (impedance explorer + SPICE model), and the design outputs — schematic, catalog parts, BOM, insertion loss, sizing & safety, SPICE netlist, predicted result.

scripts/build_wasm.sh        # engine → web/public/ (needs emsdk + WebLibMKF)
cd web && npm install && npm run dev    # develop
npm run build && npm test               # dist/ + Playwright e2e (headless)

Roadmap

• Harmonic-comb detection / switching-frequency estimation from measured spectra
• Filter synthesis driven by Kelvin candidate lists and verified in-circuit with Kirchhoff (real source/LISN impedances instead of 50 Ω, CISPR 17 worst-case 0.1 Ω/100 Ω terminations)
• Component model pipeline: measured/FEM impedance → vector-fitted SPICE subcircuits (scikit-rf) → nonlinear saturation cores (Verilog-A/OSDI)
• Middlebrook input-filter stability check and damping design
• CISPR band A LISN branch; MIL-STD-461 / DO-160 limit packs (after verification against the standards)

License

MIT — see LICENSE.