TinySysid · Workflow — Design notes¶

Reading the workflow and accuracy tables¶

The 10 / 30 / 60 / 120 second workflow and error tables remain as historical guidance. Complete inputs, device, firmware and repeated-run statistics were not recorded, so these are not guarantees for the current implementation. Sample counts assume the example's 50 Hz rate; recalculate for other rates. Core exposes six methods; check experiment-specific headers before using PW variants or method numbers.

Keep channel order, sample rate, preprocessing and windows consistent across methods, then inspect frequency, damping, mode shapes and return codes. Agreement between methods alone does not establish measurement accuracy.

Cross-Validation Workflow — Multi-Method Consensus

No single method's modal results are fully trustworthy. Always cross-validate with complementary methods and vote for consensus.

Global Parameter Reference¶

Method	Parameter	Default	Effect	ESP32 time (5ch×3000)
PP	seg_len	min(n/3,512) pwr2	Resolution vs segments	< 10 ms
	overlap	50%	PSD variance	—
	SNR threshold	4× noise_floor	Peak sensitivity	—
	peak min dist	5 bins	Deduplication	—
FDD	seg_len	1024	Frequency resolution	~50 ms
	pow-iter tol	1e-6	SVD accuracy	—
EFDD	IFFT bandwidth	±3 bins	Damping quality	~2 s/mode
	min δ	0.001	Minimum detectable damping	—
ITD	sn (AR order)	4	Candidates per pair	2-5 s
	n_lags	100	Cross-corr length	—
	cluster tol	10%	Voting strictness	—
ERA	p (block rows)	15	Hankel time span	1-3 s
	q (block cols)	10	Hankel columns	—
	SVD threshold	0.2%·S₀	Order selection	—
SSI	block_rows	12	Toeplitz size	3-8 s
	order range	2,4,...,16	Voting orders	—
	cluster tol	10%	Voting strictness	—
## Recommended Workflow

Step 1: PP Quick Scan (10s data)¶

Goal: get a quick overview of frequency content.

PP → list all prominent frequency peaks

Note: PP may miss weak modes or merge close ones. Reference only.

Step 2: FDD Confirm + Shapes (30s data)¶

Goal: confirm frequencies, get reliable mode shapes.

FDD → cleaner SV1 spectrum, SVD-derived shapes
More noise-robust than PP

Step 3: ERA Precision (60s data)¶

Goal: most accurate freq/damping/shapes.

ERA → highest overall accuracy
NExT + Hankel SVD + balanced realization

Step 4: Cross-Validate¶

Goal: independent verification of ERA results.

ITD → catches weak modes ERA might miss (short data)
SSI-PW → damping verification (sufficient data)

Step 5: Consensus Voting¶

Condition	Judgment
All methods agree	Trusted mode
Majority agree, one missed	Majority wins
Only one method detects	Flag: needs verification

Data Length vs Accuracy¶

Length	Available	Accuracy
10s (500 pts)	PP, ITD, FDD	Freq ±5%, no damping
30s (1500 pts)	+ ERA	Freq ±2%, damping ±30%
60s (3000 pts)	+ EFDD	Freq ±1%, damping ±10%
120s (6000 pts)	+ SSI	Freq ±0.5%, damping ±5%