Expert review at industrial scale

Long Time-Series Annotation Workbench

A desktop workbench for navigating, annotating and resuming work on very long measurement records while keeping model suggestions separate from expert-approved labels.

Period: Oct 2024 — Mar 2026
Focus: Industrial Time Series · Human-in-the-loop Tooling · Software Engineering
Core stack: Python · PyQt · PyQtGraph · NumPy

TL;DR

Problem: Long paired recordings cannot be treated like ordinary desktop charts. Engineers still need to locate suspicious intervals precisely, assign a signal type and return to unfinished reviews without losing work.
What I did: Translated expert review needs into desktop interaction and positioning rules, including combined keyboard-and-wheel navigation.
Result / limit: Created the human-reviewed data entry point used by downstream classification, segmentation and clustering work.

A technician carrying out a high-voltage field inspection near electrical equipment. — Public contextual photograph—not the project site. The workbench was built for expert review of industrial measurement records.

01 · A

Problem

Long paired recordings cannot be treated like ordinary desktop charts. Engineers still need to locate suspicious intervals precisely, assign a signal type and return to unfinished reviews without losing work.

01 · B

Context

At an anonymized industrial measurement technology company, I iterated the interaction model and positioning rules with future users and domain engineers. To protect confidential work, I focus this case study on the workflow rather than customer data or identifiable device details.

Workflow illustration

A workbench designed around expert decisions

I use this workflow illustration to show the interaction model: machine candidates remain amber, expert-confirmed intervals become teal, and every edit stays tied to an original signal position.

Workflow illustration for a long time-series annotation workbench with synchronized channels and expert review. — The illustration focuses on synchronized review, editable intervals and traceability to the original signal position.

Why it matters

Partial discharge is not one uniform waveform: internal, surface and corona activity can appear alongside switching and environmental noise. Reliable interval and type labels are the foundation for separating these mechanisms and for studying insulation degradation before it develops into a more serious fault.

Partial-discharge measurement circuit with coupling capacitor, impedance and measuring instrument. — A measurement chain turns high-frequency electrical events into signals that engineers can inspect and label.

Explore technical detailsWhat made it hard · System path · Key trade-offs

02 · B

What made it hard

01
Memory and rendering cost grow faster than a desktop interface can tolerate.
02
Screen coordinates, downsampled views and source indices must remain reversibly aligned.
03
Keyboard and mouse workflows must support fast review without hiding state changes.
04
Automatic preselection may prioritize intervals, but it cannot replace expert approval.

Coordinate contract

The view may be compressed; the annotation may not drift

Long recordings are rendered through windows and reduced views. Every screen selection must still map reversibly to the source sample range.

01full resolution

Source record

Immutable sample positions

02bounded memory

Chunk window

Read only the active region

03screen resolution

Display view

Downsample for responsive drawing

04reversible

Annotation

Store original start and end

Correctness is defined in source coordinates, not pixels.

My approach

System path

Chunked access

Read only the required record windows and preserve source positions

Fast navigation

Shift + wheel pans, Ctrl + wheel zooms, and overview/detail views stay synchronized

Expert annotation

Select, cancel, delete and type intervals while review state remains explicit

Recoverable output

Persist progress, export labels and retain model suggestions separately

Human review loop

A candidate becomes a label only through an explicit decision

Preselection reduces where experts look. It does not collapse proposal, correction and approved truth into one state.

01proposed

Machine candidate

Suggested interval and score

02reviewing

Expert inspection

Zoom, pan and compare channels

03decision

Edit or reject

Move, resize, cancel or delete

04effective

Confirmed label

Type and boundaries approved

05recoverable

Persist

Save progress and export provenance

Key trade-offs

Display resolution is not analysis data

The interface can downsample what it draws while annotations continue to reference the original signal positions.

Suggestions are not labels

Signal-processing and model output is shown as a candidate layer until an engineer accepts or changes it.

Persistence is part of the workflow

Resume and export behavior were treated as core domain behavior rather than a final save button.

04 · A

My contribution

Translated expert review needs into desktop interaction and positioning rules, including combined keyboard-and-wheel navigation.
Implemented interval selection, cancellation, deletion, persistence and export paths for long records.
Connected Hilbert-envelope and quantized TCN autoencoder experiments to a reviewable preselection layer.

04 · B

What was implemented

Implemented a modular GUI and a recoverable annotation workflow.
Created the human-reviewed data entry point used by downstream classification, segmentation and clustering work.

04 · C

Lessons learned

01
For long signals, interaction latency and coordinate correctness are one product problem.

02
Human-in-the-loop systems need provenance for both machine suggestions and final labels.

What I can share

Because this work used confidential industrial recordings, I cannot publish customer data or internal performance benchmarks. I can show the workflow I implemented: modular signal processing, resumable annotation and export of original sample positions.

Public sources

Field-inspection context photo · CC BY-SA 3.0 https://commons.wikimedia.org/wiki/File:Portable_TEV_Detector_in_use.jpgPartial-discharge measurement diagram · CC BY-SA 3.0 https://commons.wikimedia.org/wiki/File:Partial_discharge_1a_svg.svgOMICRON guide to partial-discharge measurements https://www.omicronenergy.com/uploads/tx_downloader/51c20/A_Guide_for_PD_Measurements_on_MV_and_HV_apparatus_-_Part1.pdf