Chemical Process Monitoring Time-Series Dataset

Kaggle — Chemical Process Monitoring Time-Series Dataset is a fully synthetic, minute-level multivariate record of six simulated commodity-chemical reactors - 777,600 rows and 21 columns covering ninety unbroken days from midnight on 1 January 2024 to the last minute of 30 March. Its author describes it as a continuous chemical reactor system built with realistic industrial sensor behavior, control dynamics, seasonal effects, multiple operating regimes and gradual fault development, and states plainly that no real plant or proprietary data is involved.

What makes it interesting for Commodity Chemicals work is that it is not a flat table of random numbers. Three reactors run regime A baseline operation and three run regime B stress operation - hotter, pressurized, with cooling that underperforms and noise that misbehaves. A slow sinusoidal ambient-temperature term bends cooling efficiency across the whole quarter, so the same reactor looks different in March than it did in January. Faults arrive gradually, not as instant step changes, and every row carries the labels that make supervised work possible: fault_type, efficiency_loss_pct and time_to_fault_min. Get a sample of this dataset scoped to the reactors and windows you want to test against.

Three opening minutes from reactor A_R1, straight from the file:

# CHEMICAL_PROCESS_TIMESERIES.CSV -- REACTOR A_R1, REGIME A, THE OPENING MINUTES
timestamp            reg  reactor  temp      press    feed      coolant  agit_rpm
2024-01-01 00:00:00  A    A_R1     181.1356  15.7910  101.1089  79.1546  305.7799
2024-01-01 00:01:00  A    A_R1     182.2498  15.7070  98.9324   -        302.2836
2024-01-01 00:02:00  A    A_R1     183.1297  15.5934  99.7922   80.5937  -

# THE SAME FIRST MINUTE, WIDER VIEW
ambient_temp_effect : 0.0000000        agitator_speed_rpm : 305.7799
reaction_rate       : 0.7245           conversion_rate    : 99.1518
selectivity         : 91.9274          yield, vibration, motor current,
                                       power, both setpoints : further channels alongside

# LABEL DISTRIBUTION, ALL 777,600 ROWS
fault_type 0  normal         : 749,699 rows
fault_type 1  gradual fault  :   9,143 rows
fault_type 2  gradual fault  :   6,842 rows
fault_type 3  gradual fault  :   4,474 rows
fault_type 4  gradual fault  :   7,442 rows
faulted share                :   ~3.6 percent

# SHAPE
reactors         : 6  (A_R1-A_R3, B_R1-B_R3)     regimes : 2
rows per reactor : 129,600                        columns : 21
cadence          : 1 minute x 90 days             total   : 777,600 rows

The texture tells you how to work with it. Sensor values arrive at full float precision, so a temperature reads 181.1355580507772 rather than 181.14 - keep the digits, they carry the drift signals. Some cells arrive empty: the simulator stages roughly 6 percent sensor dropouts, concentrated in the coolant-flow and reactor-pressure channels, which is exactly the missing-data problem a real historian hands you. And the labels are rare by design - about one row in twenty-eight carries a fault, spread across four modes, which is the class imbalance any honest detector has to survive.

Twenty-one columns group into five jobs. Identity and context: timestamp, operating_regime, reactor_id and ambient_temp_effect tell you which unit, which minute, which world-state. Control-room sensors: reactor_temp, reactor_pressure, feed_flow_rate, coolant_flow_rate and agitator_speed_rpm are the channels an operator would watch on a trend screen. Quality variables: reaction_rate, conversion_rate, selectivity and yield_pct say how well the reaction is going. Equipment health: vibration_rms, motor_current and power_consumption_kw watch the machinery instead of the chemistry.

Then the part most sensor dumps lack. Control intent: temp_setpoint and pressure_setpoint record what the loops were asked to hold, so setpoint-versus-measurement gaps become a diagnostic rather than a guess. Labels: fault_type classes the minute, efficiency_loss_pct grades its damage and time_to_fault_min counts down to the event. Those three columns are why this dataset supports predictive-maintenance modeling instead of just plotting.

• Geography: none, deliberately. These are simulated reactors, so there is no country, region or site dimension to join against - the physics is the coverage.
• Temporal: ninety continuous days, 2024-01-01 00:00 through 2024-03-30 23:59, sampled every minute without a gap. Long enough for the ambient sinusoid to bend the baseline, short enough to profile in an afternoon.
• Granularity: one row per reactor per minute - six units times 129,600 minutes equals 777,600 rows, each carrying 21 columns.

A caution worth more than most metadata: the catalog blurb undersells the depth, describing roughly a quarter-million rows over about half a year, while the file itself holds 777,600 rows across 90 days. We profile the artifact, not the brochure - and the difference matters when you budget rows for training versus validation.

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

You pick the channel and the cadence; the twenty-one-column dictionary above travels unchanged across all three. Files suit benchmark runners loading the full 90-day window once and iterating locally, API payloads suit products that pull a reactor-day at a time inside a dashboard or a demo, and warehouse delivery suits teams joining minute-level telemetry with their own maintenance logs in SQL. Changing cadence is a settings conversation, not a re-integration project.

• Data scientists and ML engineers benchmark anomaly detection, fault diagnosis and forecasting models on labeled ground truth - four gradual fault modes, a 3.6 percent positive class, and leave-one-reactor-out splits ready-made by reactor_id.
• Reliability and process-engineering teams prototype predictive-maintenance logic on time_to_fault_min before pointing anything at production historians, and use efficiency_loss_pct to tie alarms to money instead of just to alerts.
• Developers and data-product builders stand up digital-twin demos, monitoring dashboards and alerting workflows against realistic dynamics - regimes, seasonality and dropouts included - without waiting on a plant.
• Journalists, academics and students teach and write about industrial sensing with data that carries no confidentiality baggage, because nothing in it came from a real facility.
• Consultants and analytics vendors score competing monitoring tools against one shared reference instead of everyone grading their own homework.

Data scientists and ML engineers get the rarest thing in applied forecasting work: labels. fault_type classes the minute, time_to_fault_min converts the problem into remaining-useful-life regression, and the two-regime design punishes models that quietly memorized one operating envelope. Developers and data-product builders get realistic motion for demos - a reactor that drifts with the seasons and drops a coolant reading now and then beats a sine wave every time. Competitive-intel and product teams get a neutral yardstick for scoring monitoring vendors. Investors and quants get a reference environment for stress-testing industrial-IoT theses without buying hardware. Journalists, academics and students get citable, confidential-free material for teaching what a control room actually sees.

Because the raw artifact arrives as a quarter-gigabyte CSV with opinions. We ship it typed against the twenty-one-field dictionary above, with a documented policy for the ~6 percent dropout cells instead of silent nulls, label columns joined and typed so a classifier trains on arrival, and regime-aware splits built in so nobody accidentally validates on the operating condition they trained on. Need engineered features - rolling drift statistics, per-reactor baseline residuals, one-row-per-episode fault tables? Those come mapped onto the same identifiers.

Get a sample of this dataset cut to the reactors, windows and label modes you care about, then put the full window on the delivery cadence your sprint plan actually runs on.

Notes that pair well with this page:

• NIST Chemistry WebBook - evaluated thermochemistry and physical constants to anchor the simulated sensor ranges in real chemistry.
• EPA TSCA Chemical Substance Inventory - the regulatory universe of US commercial chemistry behind the substance classes these reactors stand in for.
• OECD eChemPortal - federated hazard and classification records across 34 government databases for the same chemistry.

Background reading: Predictive Maintenance and Multivariate Time Series. Browse the vertical on the Commodity Chemicals industry hub or the ranked best commodity chemicals datasets.