Skip to content

v0.3.0 — 406 multimodal public dataset

Choose a tag to compare

View all tags
@dctthree dctthree released this 25 Aug 06:07
v0.3.0
e118ba6

v0.3.0 — 406 mm multimodal public case

This release adds the real 406 mm four-point-bending / in-line pull-test campaign while preserving the modality and truth boundaries of the existing P110 EXP2 release.

Added

  • A fully documented case_studies/406_multimodal package with MATLAB source, derived CSV/JSON tables, 13 real blind-repeat figures plus one frozen calibration figure, and an independent fail-fast validator.
  • The 406 calibration partition: zero load S0 plus six loaded states S1–S6, shared MEM/remanence magnetic CSVs and strain-gauge records.
  • The 406 blind-repeat partition: C1 and C2 complete, C3 partial S0–S2, for 17 paired magnetic/ETP packets.
  • Public field photographs with provenance, colour normalization, metadata stripping and an explicit background-person privacy note.
  • Real calibration and repeatability figures on the repository front page.
  • A versioned data Release index, per-file and per-asset SHA-256 manifests, channel contracts and a bilingual dataset card.

Sensor contracts

  • The 406 magnetic CSV contract has 1307 named numeric fields. The first 1280 fields are 320 × (X,Y,Z,T).
  • There are 32 physical columns with 10 sensing positions each. One-based odd physical columns are remanence and even physical columns are MEM; each group therefore contains 160 sensing channels.
  • MEM and remanence share the same magnetic CSV stream. They must not be counted as two raw files.
  • ETP is an independent 67-field stream with 20 (Amplitude, Phase) channel pairs plus acquisition fields.

Frozen evidence boundary

  • MAG-F1-DW-Q90-v1: primary same-pipe relative-order/change evidence. Strict C1↔C2 repeatability after the pre-declared C2/S2 rejection: Spearman ρ 1.000, Lin CCC 0.996, normalized RMSE 3.35%.
  • MEM-F4-ZSD-v1: unsigned, same-cycle-S0 auxiliary evidence. Strict C1↔C2 CCC 0.980, normalized RMSE 7.51%.
  • ETP candidates: research/QC only because target specificity and negative-control gates do not pass.
  • Fusion: fail-closed modality/QC eligibility gating only; no numeric fusion value. Explicit stagewise cross-modality conflict scoring is not implemented in v0.3.0.
  • Blind absolute MPa: disabled. The blind-repeat partition has no contemporaneous strain/MPa truth.

The calibration MPa column is derived as 206 GPa × median bending strain; it is not a load-cell measurement and must not be copied into the blind-repeat labels.

Deliberately excluded

  • blind test/data/未处理数据, which duplicates or precedes the curated clipped magnetic source tree.
  • Acquisition screenshots, old single-cycle conclusions, deprecated reports, large MAT/FIG runtime products and local absolute paths.
  • The old continuous ETP recording as a stage-by-stage calibration source; it represents one strain state and cannot be aligned to S0–S6.

Compatibility

The generic Python data platform remains backward-compatible with the P110 v0.2.0 package. The 406 raw-data reader is supplied as a dedicated MATLAB case because its 1307-field fragmented magnetic contract and independent complex ETP contract differ materially from the P110 three-axis CSV layout.

Download and integrity

This Release contains 13 attachments: 8 raw-data ZIP parts plus DATASET_CARD.md, channel_schema.csv,
aw_file_manifest.csv,
elease_index.json, and SHA256SUMS.txt. Verify every downloaded file against SHA256SUMS.txt before use. Extract ZIPs while preserving their canonical
aw/... paths.

Assets 15
Loading