Refinery Centrifugal Compressor Sensor Data (One Year)

It is a full operating year of one machine, captured at the control-system level. Engineers at Motor Oil (Greece) extracted twelve months of measurements - January 1 through December 31, 2022 - from their refinery's distributed control system to monitor compressor operation, calling out turbine axial displacement and oil temperature as the key monitored variables, and published the result as a single 11.1 MB Excel workbook in February 2025.

The scale is exactly what condition-monitoring work wants and rarely gets in one file: 25 sensor channels plus a Timestamp column, sampled on a 15-minute cadence, giving 35,036 readings per sensor - roughly 910,000 individual process values covering four seasons of continuous operation. Column names follow live DCS tag conventions with the .pv (process value) suffix intact: 75ZI800BA.pv rather than a sanitized axial_displacement_1. Anyone who has spent a career reading compressor screens will recognize the dialect immediately.

Get a sample of this dataset - name the channels and months you want and the rows come back cut to them.

Four observed readings spanning the year, all verbatim values:

2022-01-01 00:00   ZI800BA=0.0455   PI808=9.9148   TI821=55.686   SI865R=6116.63
2022-01-01 00:15   ZI800BA=0.0461   PI808=9.9295   TI821=55.5015   SI865R=6119.41
2022-01-01 00:30   ZI800BA=0.0467   PI808=9.9232   TI821=55.5250   SI865R=6117.89
2022-12-31 23:45   ZI800BA=0.0570   PI808=9.8439   TI821=54.9839   SI865R=6391.30

Read what those two endpoints already say. On New Year's midnight the shaft sat near zero axial displacement with suction-side pressure around 9.91 and the first bearing temperature near 55.7; by the final reading of December 31 the same channels read 0.0570, 9.84 and 54.98 - with speed up near 6391 rpm against roughly 6117 earlier, the largest move of the four. A full year of that texture sits behind each column, at 15-minute spacing, which is precisely the resolution where drift, fouling and slow mechanical change become visible before they become failures. Every number above is an observed record value, not an illustration; the remaining 24 channels ride alongside in your sample.

Seven field groups carry all 26 columns, verified against the published record. The tag grammar does most of the explaining: the numeric prefix is the loop number, the letter pair names the measurement - ZI axial displacement indicator, PI pressure indicator, PDI differential pressure, TI temperature, XI radial position probe, SI speed - and .pv marks the live process value.

Additional fields on request: per-tag engineering-unit confirmation (the record documents abbreviation prefixes but not units - magnitudes are consistent with barg, degrees Celsius, microns and rpm), tag-level location mapping for the twelve temperature channels (which sit on bearings versus discharge piping is not stated), and which compressor train within the refinery the loop numbers belong to. Anything outside the core dictionary above travels as an add-on confirmed against the workbook when your sample is prepared - never as a surprise.

Geography: deliberately narrow - one centrifugal compressor train inside a single Motor Oil (Greece) refinery. Site details beyond company affiliation are not disclosed. That narrowness is the point: this is depth-per-asset coverage, the same shape as operator-released telemetry programs, not a basin-wide survey.

Temporal: calendar 2022 exactly - January 1 00:00 through December 31 23:45 - with no gaps in the 15-minute sequence and no continuation into later years. It is a frozen one-year window, which makes it an honest benchmark: every model you build sees the same seasons, the same operating envelope, the same drift.

Granularity: per-sensor time series at fixed 15-minute resolution - 35,036 rows x 26 columns, about 910,000 readings. There is no event log, no alarm layer and no maintenance history attached; the machine's vital signs only, which is why it pairs so naturally with labeled-event corpora such as the 3W dataset.

Set against the wider Datadory catalog - 1,744 datasets averaging 7.81 on quality - this slice scores 6/10: verified field definitions and genuine industrial provenance, held back by undocumented units and a single-asset scope.

API, files, or your warehouse. Daily, weekly, or hourly.

Pick the channel your stack already speaks. The 26 columns arrive identical either way - timestamps typed, sensor channels kept wide so a 15-minute step stays one row instead of 25, and the tag grammar preserved rather than renamed away. Every delivery ships with the field dictionary above plus sample rows for validation, so your first anomaly-detection run happens against evidence, not hope.

Cadence changes are a settings conversation, not a re-integration project: name the channels, the window and the rhythm, and the sample comes back shaped to them before any commitment.

• Predictive-maintenance teams treat it as clean pre-training material: a full year of real rotating-equipment telemetry to teach anomaly-detection models what normal looks like before pointing them at live plant streams.
• Rotating-equipment engineers use the paired axial-displacement and radial-probe channels as a worked example of how thrust and proximity instrumentation behave over seasons of continuous duty.
• Data scientists and ML engineers get a compact benchmark surface - 35,036 steps, 25 channels, no missing-value archaeology - for comparing forecasting, drift-detection and autoencoder approaches on equal footing.
• Control-systems instructors teach DCS tag conventions from the real thing: .pv suffixes, ZI/PI/TI/XI/SI letter pairs and loop numbering, straight off a running refinery system.
• Condition-monitoring vendors benchmark detection sensitivity against a known-good baseline year, then quote results without touching customer data.

Deeper persona guidance lives on the data scientists, developers & builders and investors & quants pages for oil & gas equipment & services.

Within the industry's sensor-and-telemetry records, three single-machine corpora answer different questions. The Open Industrial Data Project (Valhall) streams a live North Sea gas-compressor subsystem with alarms, maintenance events, P&IDs and 3D models attached - richer context, but historical depth varies by tag. This dataset answers with completeness instead: one unbroken calendar year, every channel, fixed cadence, nothing missing. The NORA 3D multimodal dataset covers inspection imagery and 3D scans - geometry, not time series.

So the split is explicit. Need labels, events and drawings? Valhall. Need inspection imagery? NORA. Need a clean continuous year of one machine's vital signs to train or benchmark against? That is this record, and it is why it scores 6/10 against a 7.81 catalog average while still being the slice's most complete single-year time series.

Three things worth knowing upfront.

First, engineering units are not documented anywhere in the record. Magnitudes are consistent with barg, degrees Celsius, microns and rpm, but the source never states them; we flag unit inference explicitly in every delivery rather than letting a dashboard inherit a guess.

Second, the archive is one year with no successor - calendar 2022 ends at December 31 23:45 and no later vintages exist upstream. Teams needing current conditions should pair it with the live Valhall stream and let this hold the benchmark role.

Third, tag-to-location mapping stops at the abbreviation prefixes: which of the twelve temperature channels sit on bearings versus discharge piping, and which train the loop numbers describe, are not published. Name the channels driving your model in the sample request and we confirm what can be pinned down before anything ships.

Units note - no channel carries a documented unit; observed ranges (axial displacement near +/-0.1, temperatures 55-92, speed roughly 6,100-6,400) support the conventional readings but stay flagged as inferred in our deliveries.

Scope note - one compressor train, one refinery, one calendar year. Portfolio-scale or multi-site questions need different records; this one goes deep instead of wide.

Where to go next:

• Open Industrial Data Project (Valhall) - the live-stream counterpart: alarms, maintenance events and documents around a running North Sea compressor.
• NORA 3D Multimodal Oil & Gas Dataset - inspection imagery and digital-twin geometry beside this machine's time series.
• 3W Dataset 2.0.0 (Oil Well Events) - Petrobras' labeled undesirable-event corpus; where this dataset supplies raw vitals, 3W supplies the labels to train against.
• Oil & Gas Equipment & Services data hub - every cataloged dataset in the industry, this one included.
• Best oil & gas equipment & services datasets - the ranked shortlist across the vertical.