Condition, Don't Remove: Speed-Conditioned Normal Modeling for Unsupervised Fault Detection in Rotating Machinery

Unsupervised (normal-only) anomaly detection for rotating machinery under operating-speed shift. Instead of removing speed (order-tracking / invariance) or treating it as a discrete domain label, we condition the normal-only model on the continuous tachometer speed covariate. On MaFaulDa (accelerometers + microphone + tachometer, five fault types, 12–60 Hz) conditioning raises pooled ROC-AUC from 0.955 to 0.989 at matched speed and from 0.897 to 0.937 on held-out speeds, wins on every fault type, is backbone-agnostic (PCA and autoencoder), holds across accelerometer and microphone channels, and generalizes to the MIMII-DG acoustic benchmark.

Read the paper

• Web (GitHub Pages): https://apartsinprojects.github.io/EngineKnock/
• Word: docs/EngineKnock_speed_conditional.docx
• Source HTML: paper/index.html

Repository

• paper/, docs/ — the paper (HTML source + figures + DOCX; docs/ serves the Pages site).
• scripts/ — all experiments (exp001–exp010): ingest, baselines, the speed-conditional method, controls, multi-fault / backbone / multi-channel validation, and the MIMII-DG breadth test.
• experiments/ — per-experiment registry (hypothesis, setup, numbers, conclusion) and INDEX.md.
• results/ — result JSONs and figures.
• PROJECT_LOG.md, BACKLOG.md, PAPER_OUTLINE.md — project log and plan.

Data

The datasets are public but not redistributed here. MaFaulDa: https://www02.smt.ufrj.br/~offshore/mfs/. MIMII-DG (DCASE 2022 Task 2): Zenodo 10.5281/zenodo.6529888. Scripts read the dataset zips from data/ (git-ignored).

Status

Preprint draft. Author block anonymized.