Milling tool wear dataset: per-pass wear labels, dense signals, one delivery

A milling tool wear dataset records cutters dying on purpose: each archive runs one or more CNC milling tools through a sequence of material removal passes at defined speeds and feeds until the flank is worn out, measuring wear on the tool after passes while cutting force and vibration stream continuously during them. It is supervised learning for the machine shop - failure labels measured in millimetres of flank wear rather than inferred from maintenance logs.

The anchor product is the NASA Prognostics Data Set Repository - roughly twenty-one run-to-failure archives curated by NASA's Prognostics Center of Excellence at Ames Research Center, with "CNC milling machine and milling tool wear" named among its holdings. Datadory scores it 9 out of 10, the joint-highest tier in this industry's pool alongside the NBER-CES manufacturing panel, and its field definitions are marked verified rather than inferred - which matters when pass indices have to line up exactly with wear tables.

The corpus was collected between 2008 and 2019 and circulates as fixed research versions, sized from about 12 MB (C-MAPSS v1) to about 15.8 GB (C-MAPSS v2) and tens of gigabytes overall. Of all 1,744 datasets in the catalog, this is among the few whose rows are labelled failure events rather than market observations - and it is the only layer of the industrial-machinery slice that reads the component itself instead of the demand around it.

Two file families that meet at the pass index:

1. Dense signal channels - cutting force and vibration sampled at high rate throughout every pass, the raw material from which condition indicators get engineered.
2. Sparse wear labels - flank wear measured on the cutter after selected passes, dozens of labelled points per tool rather than thousands.
3. Pass metadata - speeds, feeds and pass ordering that tell you which signal window belongs to which label.

Datadory's field-granularity note reads: "Per-component run-to-failure time series - per-cycle sensor readings, per-cycle battery capacity fade, milling per-pass wear." The wear label attaches to the pass that just finished, not to individual signal samples, so alignment by pass index is the load-bearing join in any pipeline built on this data.

That structure shapes everything downstream. A turbofan archive such as C-MAPSS gives deep sequences - train_FD001.txt alone carries 20,631 labelled cycles across 100 train units - so sequence models eat well. A milling archive gives few labels and rich context per label, so the standard recipe pairs the two families: engineer RMS, kurtosis and spectral-peak features per pass from the force and vibration windows, then regress those against measured wear.

Formats follow the house style of the whole repository: ZIP archives unpacking to TXT (space-delimited), MAT and CSV files, so MATLAB users load variables directly and Python users read flat text without a converter.

Six steps, two of which decide whether the model survives contact with a real spindle:

1. Scope the archive - sibling holdings range three orders of magnitude in size, so confirm you hold the milling entries before committing disk.
2. Separate the two file families - high-rate in-cut signals versus between-pass wear tables.
3. Align by pass index - the label belongs to the pass just completed; misaligned boundaries silently inflate accuracy more than any architecture choice.
4. Engineer per-pass features - RMS, kurtosis, spectral peaks over each pass's force and vibration windows.
5. Validate leave-one-tool-out - within-tool splits memorise one cutter's trajectory and hide overfitting until deployment; cross-tool splits prove generalisation.
6. Carry the citation with the artifact - every entry ships a suggested citation (C-MAPSS's is Saxena, Goebel, Simon and Eklund, PHM 2008) and publications are asked to acknowledge both the repository and its data donors. Recording it in the experiment README at delivery time beats reconstructing provenance after review.

Steps 3 and 5 are where most projects stumble, which is why they are the two steps Datadory documents hardest in the field dictionary: the dictionary states what each column measures and at which grain, so the pass-index join is explicit rather than tribal knowledge.

Choose by decision, not popularity. If the output is a cutter-change threshold in minutes of spindle time, only per-pass wear labels give a physically meaningful target - vibration anomaly scores cannot tell an operator when a flank has reached its wear limit. If the output is a generic remaining-useful-life architecture to transfer later onto proprietary machine telemetry, C-MAPSS is the fastest place to prove the method: four difficulty tiers, true RUL files for every test unit, published leaderboards everywhere - which is why its community mirror became the standard. If the problem is spindle bearings, the same corpus carries bearing vibration including the FEMTO accelerated-degradation study - vibration-rich like the milling signals, but fault-injected and accelerated rather than naturally worn.

Teams building one condition-monitoring platform for a machine shop typically prototype on C-MAPSS, validate feature pipelines on bearing vibration, then deploy against their own spindle and cutter telemetry - the milling archive is the bridge that makes the last step credible.

Four things, and none of the plumbing is your problem:

That stability is a modelling feature, not a staleness tax. Unlike the marketplace sources in this slice - where listings reprice and expire on sellers' schedules and price history only accrues from repeated captures - a fixed corpus means a benchmark you ran in March still means something in November.

Everything here is verifiable before commitment: sample rows, the full field dictionary and coverage chips sit on the NASA Prognostics Data Set Repository product page. Read them, then request a sample scoped to the component families and fields you care about.

Data scientists and ML engineers own this corner of the slice - relevance tops out at 3 of 3 - training degradation and RUL models on the milling, bearing, battery and turbofan archives, with leave-one-tool-out validation as house discipline. The fuller workflow lives at best data for industrial machinery data scientists.

Developers and data-product builders stand ingestion and labelling tooling up against documented per-archive layouts, citations attached, so downstream consumers never inherit an undocumented schema.

Journalists, academics and students cite the repository and its per-dataset donor citations in prognostics research - Saxena et al., PHM 2008 remains the canonical reference for the engine benchmark.

Competitive intelligence teams test condition-monitoring vendors' claims against published run-to-failure benchmarks before those claims reach a purchase order.

Market researchers and consultants scope predictive-maintenance market studies against the archive catalog itself - who funds degradation science is a market signal.

One habit unites all five groups: logging citation and archive version alongside every experiment row, so a result can always be traced back to the exact bytes that produced it.

Wear models earn their keep once they plug into sourcing and demand context, and the surrounding layers supply both:

• Sourcing side: Made-in-China.com catalogs millions of live component listings carrying indicative US-dollar price ranges, minimum order quantities and audited-supplier flags - exactly what you need when a wear model says the current cutter grade wears too fast and alternatives must be priced. Roughly thirty listings arrive per search-results view, keyed by supplier storefront.
• Demand side: JMTBA Machine Tool Statistics (Japan) tracks monthly machine tool orders split domestic versus foreign back to January 2009, and Census M3 carries US manufacturers' shipments, inventories and orders monthly since January 1992 in a bulk panel of roughly 600,000 rows. When tooling spend rises in a shop, these series say whether capacity is expanding or merely replacing worn stock.
• Productivity context: the NBER-CES Manufacturing Industry Database reaches back to 1958 for the macro frame around machining intensity.

The contrast between layers is the point: the physics layer is fixed and reproducible, the demand layers refresh monthly, and the marketplace floor reprices continuously. Joining a stable wear corpus to moving commercial series is how a millimetres-of-flank-wear chart becomes a procurement decision.

Start with the industrial machinery supplies components data guide, the pillar post scoring all twenty pooled records for this industry across four working layers - including the full prognostics corpus that anchors this page. Two siblings extend the threads: turbofan engine degradation dataset walks the C-MAPSS benchmark end to end, and china industrial supplier directory covers the sourcing floor if wear analysis turns into a cutter-buying exercise. For head-to-head framing, Kaggle NASA Turbofan Jet Engine (C-MAPSS) vs JMTBA Machine Tool Statistics (Japan) weighs physics against demand, and NASA Prognostics Data Set Repository vs NBER-CES Manufacturing Industry Database weighs component physics against the productivity panel.

Product-level detail - sample rows, field dictionary, coverage chips - sits on the NASA Prognostics Data Set Repository page, and the industrial-machinery data hub indexes everything mentioned here. Open the record, read the dictionary, then request a sample cut to your components and your passes.