Skip to content

data-inventory.md

Latest commit

 

History

History
56 lines (43 loc) · 2.8 KB

data-inventory.md

File metadata and controls

56 lines (43 loc) · 2.8 KB
Raw

Data inventory

Azure PdM (baseline)

  • Kaggle slug: arnabbiswas1/microsoft-azure-predictive-maintenance
  • License string reported by Kaggle CLI: unknown (UNVERIFIED, flagged; scout notes downstream mirrors treat this as Microsoft sample data + Kaggle ToS).
  • Fetched: 2026-08-28
  • Location: data/raw/azure/

Files

File SHA256 (first 16) Rows (ex. header) Notes
PdM_errors.csv 9c2a2a010ad77227 3919 datetime,machineID,errorID. Error1..error5 vocabulary.
PdM_failures.csv 0c6c31a4fd52ef2d 761 datetime,machineID,failure. comp1..comp4 vocabulary.
PdM_machines.csv 5e8e1571c4999bf8 100 machineID,model,age. Exactly 100 distinct machines as expected.
PdM_maint.csv 481ed4e155f609e6 3286 datetime,machineID,comp. Starts 2014-06-01, precedes telemetry.
PdM_telemetry.csv d957f3c45bb83416 876100 datetime,machineID,volt,rotate,pressure,vibration. Hourly.

Full SHA256 in data/raw/azure/CHECKSUMS.sha256.

Sanity vs scout report

  • Distinct machines: 100 (matches scout expectation).
  • Telemetry timestamps: 876,100 rows (matches scout's 876,101 stated timestamp count; off by 1 header-line accounting).
  • Telemetry window: 2015-01-01 06:00 -> 2016-01-01 (spot-checked below).
  • PdM_maint extending before telemetry (2014-06-01) is expected: the dataset preserves maintenance history for machines whose telemetry starts in 2015. Loader must handle this so early maintenance events don't become NaN or drop.

Seed-failure batch (discovered 2026-08-28)

The PdM_failures table contains 18 rows at exactly 2015-01-02 03:00:00 (one or more per machine, various comp codes) that have NO matching entry in PdM_maint. Every failure AFTER that timestamp does join to a maint row.

Read as a seed batch planted by the synthetic-data generator to bootstrap failure history at the start of the observation window. Kept as legitimate terminal_failure events. The Azure loader defines SEED_FAILURE_TS = 2015-01-02 03:00:00 for downstream code that needs to filter or annotate them.

Impact: any pattern-mining or predictive-window sampling that requires a preceding PdM_maint row for every failure MUST exclude or special-case this timestamp. The seed failures have essentially no pre-failure event history (telemetry starts 2015-01-01 06:00) and should probably be excluded from Phase 2 window sampling.

Alibaba cluster-trace-v2018 (primary)

Not yet fetched. See BACKLOG.md "Phase 1 — Alibaba primary".