Organizing unfamiliar signal families
A multi-domain feature and clustering workflow that groups similar discharge candidates, preserves outliers and returns every cluster to representative waveforms for expert review.
TL;DR
01 · A
Engineers cannot inspect thousands of candidate events one by one, and new signal families may not fit a predefined classifier. The analysis needs to compress the review space without pretending that a cluster ID is a physical diagnosis.
01 · B
At an anonymized industrial measurement technology company, clustering followed preselection and classification so domain engineers could compare distributions, representative waveforms and outliers. Customer and device details remain private.
Different discharge mechanisms can form distinct waveform families, while noise and unusual operating conditions create their own structures. Clustering reduces the number of examples an expert must inspect and can reveal unfamiliar groups, but only if the result stays connected to the original waveform and measurement context.
Raw waveform distance alone is sensitive to phase, amplitude and duration.
Feature choice and scaling can change the apparent cluster structure.
Density-based noise points are not automatically non-discharge events.
Two-dimensional projections help inspection but distort high-dimensional distance.
System path
Extract time-, frequency-, spectral- and wavelet-domain feature families
Scale robustly and compare original feature space with reduced views
Apply density-based and hierarchical clustering while preserving outliers
Inspect medoids, waveforms, spectra and wavelet signatures for each group
Feature-to-review pipeline
Complementary feature families are scaled, grouped and then traced back to representative source waveforms for expert interpretation.
Candidate events with source links
Time, spectrum and wavelet descriptors
Prevent one magnitude from dominating
Keep dense groups and outliers
Return each group to real waveforms
Key trade-offs
Time, spectral and wavelet views capture complementary aspects and remain individually inspectable.
Forcing every event into a cluster hides uncertainty and can erase rare patterns.
Embedding plots organize attention; representative raw signals remain the object of engineering review.
04 · A
04 · B
Three views, three jobs
The original feature space supports grouping, the 2D projection supports navigation, and the raw waveform supports engineering judgment.
Distances used by the clustering workflow
A lossy map that helps humans navigate
Signal shape, spectrum and context
01A clean two-dimensional cluster picture can be less truthful than a messy but traceable review set.
02Unsupervised analysis becomes useful when it reduces expert work while preserving uncertainty.
What I can share
I cannot publish customer recordings, throughput or cluster-quality figures from this confidential work. I share the feature, clustering and representative-waveform review pipeline through a fixed-seed illustration.
Public sources