Seven engineering packages that take an automotive Ethernet cable all the way through: measure it, prove the measurement can be trusted, understand what the factory did to it, compare laboratories across three countries, and predict whether a harness will pass the standard months before the first sample exists.
7 / 7 demonstrations verified end-to-end
Every package stands alone and is installable on its own, but each one consumes what the previous ones produce. The arrows are real dependencies, not a diagram drawn after the fact.
Every box is a page — click any package for the full engineering case study.
Each is a working package with its own test suite, command-line interface and technical report stating the mathematics it implements and the limits of what it can claim. Every card opens a full case study — and one page follows a single sample through all seven.
Removes the fixture, converts raw four-port data into the quantities engineers specify, and issues a verdict against the standard with the margin and the frequency where it was worst.
Knows what measurement is being performed, what a valid calibration is, which checks must pass, and whether a result is trustworthy enough to enter the database at all.
Reproduces a legacy-style shielding tool exactly so old results stay reproducible, implements the modern method beside it, and quantifies the cost of migrating.
Turns a frequency-domain reflection into characteristic impedance metre by metre, so a defect becomes a position rather than a bump on a graph.
Correlates extrusion-line parameters against measured high-frequency behaviour, and predicts loss and impedance from a production record alone.
Gives every result a full uncertainty budget, so a 0.8 dB disagreement against 1.2 dB of uncertainty is reported as not distinguishable rather than as a fault.
Assembles a virtual link from measured, archived or predicted cable data and answers whether it passes, how far it reaches and what the receiver sees.
From a single noisy 10 m measurement it predicts 3 to 25 m at 23 to 125 °C to within 0.1 dB, extrapolates an octave past the band to within 0.04 dB, and its worst-case eye is a tight lower bound on a bit-level simulation — 420 against 421 mV.
When several sites measure the same product, the question is not what each one measured but whether they agree — and how you would know. Every result carries a full uncertainty budget, so a difference is only called a problem when it is larger than the uncertainty of that difference.
Insertion loss at 600 MHz on the circulating artefact, aggregated over ten fortnightly rounds. The reference is a Cox largest-consistent-subset consensus of 7.9149 dB. En is the difference from that reference divided by the expanded uncertainty of the difference: any value below 1 means the sites are not distinguishable. Site symbols are displaced where two sites coincide at map scale.
Every image below is generated by running the code in this repository — no mock-ups, no illustrations. Each figure is reproducible from a single command.
Loss-aware layer peeling turns a frequency-domain reflection measurement into characteristic
impedance along the cable. The plain transform is biased by 4 to 9 Ω over 15 m;
this reconstruction is within a few tenths of an ohm. zprofile
Calibration, repeatability, fixture, instrument, environment and noise, combined the
GUM way and stored with the result. labplatform
Trend, change point and EWMA on each site's artefact series, with the root cause attributed
against calibration and environment records. labplatform
One-at-a-time sensitivity says the link hinges on connector balance — 0.96 dB — while
every other parameter moves it by under 0.01 dB. linktwin
Loss and impedance predicted from manufacturing parameters alone, with honest 90 %
prediction intervals, cross-validated against held-out samples. cableanalytics
Cable, connectors and PCB traces cascaded on all four ports, judged against the same limit
files a measured harness would face. linktwin + cablecheck
The old tool is reconstructed byte-for-byte so historical results stay reproducible, and the
modern method runs beside it with the difference quantified. shieldeval
Nothing here could have been built in a different order. Each phase needed the one before it to exist first — the digital twin is only possible because the measurement layer underneath it is trustworthy.
The measurement layer: network algebra, de-embedding, mixed-mode decomposition and limit evaluation, plus the two standalone algorithm projects.
Linking what happens on the extrusion line to what the laboratory measures, and turning that link into a predictive model with stated intervals.
Making measurements reproducible six months later from the archive alone, then making several laboratories comparable to each other.
Using every layer below to answer design questions — pass or fail, reach, connector budget, eye height — before any hardware is built.
Each package ships a technical report that gives the mathematics with notation defined where it is used, cites the standards and papers it rests on, reports the validation against ground truth, and ends with a section declaring what the method does not model. The reports are built from source in continuous integration, so they cannot drift from the code.
I am Sreeram Anil, an MSc Electromobility student at FAU Erlangen-Nürnberg, writing my thesis on model-predictive control, based in Bavaria. This toolchain is a portfolio project: it was built to find out whether an entire measurement-to-prediction chain for high-frequency cables could be made to hold together, and to build it the way it would have to be built in industry.
Every package has a test suite that runs on three Python versions on every push, a technical report built from source, an uncertainty statement where one is warranted, and a declared list of its own limitations. The engineering judgement — what to model, what to measure, what to validate against, and what to refuse to claim — is the substance of the work.
No proprietary or customer measurements appear anywhere in this toolchain. Cables come from a physics-based synthesiser, instruments from a simulator with a realistic error model, and production records from a generator. That is a deliberate choice, and it is what makes the validation meaningful: the ground truth is known exactly, so every claim about accuracy is checked against a number rather than asserted.
The commit history records it openly. The architecture, the validation strategy, the decisions about what each method may and may not claim, and the limitations stated in every report are the engineering content — and they are what I can walk you through line by line.