README.md

metadata

license: cc-by-nc-4.0
language:
  - en
  - pt
tags:
  - automotive
  - ecu
  - telemetry
  - synthetic-data
  - engine-control
  - automotive-ml
  - powertrain
  - knock-detection
  - diagnostics
task_categories:
  - tabular-classification
  - tabular-regression
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

ThermoDrive-ECU-50M

A large-scale synthetic dataset of automotive ECU telemetry for thermodynamic modeling, engine diagnostics, and powertrain ML research.

Table of Contents

1. Overview
2. Dataset at a Glance
3. Motivation and Use Cases
4. Data Generation Methodology
5. Schema Reference
6. Engine Taxonomy
7. Statistical Characterization
8. Data Access
9. Benchmark Tasks
10. Known Limitations
11. License and Citation

---

Overview

ThermoDrive-ECU-50M is a synthetic, tabular dataset comprising 50 million ECU (Engine Control Unit) telemetry records spanning multiple engine families, displacement classes, and forced induction configurations. Each row represents a discrete engine operating state captured across a wide operational envelope — from cold idle through full-load, high-RPM conditions.

The dataset is purpose-built to support machine learning workloads in the automotive domain: detonation detection, torque/power prediction, thermal management modeling, and anomaly detection in combustion systems. Records are generated under a physically consistent simulation framework, with feature distributions anchored to real-world ECU calibration maps for naturally aspirated, turbocharged, and twin-turbocharged powertrains.

---

Dataset at a Glance

Property	Value
Total records	50,000,000
Split	Train only
Storage format	Apache Parquet
Compressed size	~911 MB
Features	17 columns (2 categorical, 15 numerical)
Engine families	3 (aspirado, turbinado, biturbo)
Engine variants	11 unique engine_id codes
Target variables	detonacao, torque_nm, potencia_cv
License	CC BY-NC 4.0
Host	Hugging Face Datasets Hub
Identifier	Kodjaoglanian/ThermoDrive-ECU-50M

---

Motivation and Use Cases

Automotive ECU datasets with sufficient scale for deep learning are rarely available in the public domain. Real-world OBD logs are constrained by proprietary calibration IP, instrumentation cost, and the narrow operational coverage of naturalistic driving. ThermoDrive-ECU-50M addresses these gaps by providing a high-volume, physically grounded synthetic corpus.

Primary research applications:

Knock detection and detonation prediction — The binary detonacao column provides a direct supervised signal for classifier development. Operating at 50M samples enables training deep architectures without data augmentation while preserving class imbalance characteristics present in real systems.

Engine performance regression — torque_nm and potencia_cv targets span realistic ranges across all engine variants, supporting development of neural surrogate models for powertrain calibration.

Thermal management modeling — temp_agua_c and temp_ar_admissao_c serve as input features or prediction targets for thermal load models used in cooling system control and warm-up strategy optimization.

Anomaly detection — Boundary cases in oil pressure (pressao_oleo_bar), lambda deviation (sonda_lambda_afr), and EGT (egt_c) provide realistic seeds for out-of-distribution detection benchmarks.

Cross-engine transfer learning — The structured engine taxonomy allows controlled domain shift experiments: train on naturally aspirated engines, evaluate generalization on turbo variants.

---

Data Generation Methodology

Records are generated via a parametric simulation pipeline conditioned on engine type and displacement. The generative process follows these principles:

Operating point sampling — RPM values are sampled from engine-type-specific distributions that reflect realistic speed-load profiles. Naturally aspirated engines are overrepresented at mid-range RPM; turbocharged variants include higher proportions of low-RPM, high-load operating points corresponding to turbo spool behavior.

Thermodynamic consistency — Key relationships between variables are enforced: ignition advance (ponto_ignicao_graus) is inversely correlated with load (carga_tps_map_pct) following knock-limited spark advance logic; injection duration (tempo_injecao_ms) scales with fuel demand derived from volumetric efficiency and air mass flow approximations; EGT rises monotonically with load and RPM.

Forced induction modeling — Turbo pressure (pressao_turbo_bar) is absent (zero or near-zero) for aspirado entries and follows a boost-RPM relationship for turbinado and biturbo variants. biturbo engines exhibit elevated boost ceilings and higher torque floors.

Detonation labeling — The detonacao binary label is assigned via a rule-based engine that penalizes high-load, low-advance, high-intake-temperature, and low-lambda conditions. Class imbalance is intentionally preserved to reflect the rarity of knock events in a well-calibrated ECU map.

Batch generation — Records are produced in large batches with per-engine stochastic seeds, then shuffled globally. No temporal ordering is implied — each row is an independent operating state snapshot, not a time-series sample.

---

Schema Reference

All columns and their semantics are described below. Categorical columns are stored as string; numerical columns use float32 or int32 as indicated.

Categorical Features

tipo_motor

Engine aspiration type. Controls the overall thermodynamic regime.

Value	Description
aspirado	Naturally aspirated — no forced induction; intake pressure at or below atmospheric
turbinado	Single turbocharger — positive boost with moderate ceiling
biturbo	Twin-turbo configuration — higher boost ceiling and wider torque band

engine_id

Compound identifier encoding aspiration prefix and displacement in liters.

Family prefix	Example IDs	Displacement range
NA_	NA_1.0, NA_1.6, NA_2.0, NA_3.5	1.0 L to 3.5 L
T_	T_1.0, T_1.6, T_2.0, T_2.5	1.0 L to 2.5 L
BT_	BT_2.0, BT_3.0, BT_4.0, BT_5.0	2.0 L to 5.0 L

---

Numerical Features

ponto_ignicao_graus — float32

Spark ignition advance in degrees before top dead center (BTDC). Positive values indicate advance. Ranges from approximately 1° to 31° depending on load, RPM, and fuel quality assumptions.

Higher values indicate more aggressive advance timing, present under low-load, low-temperature conditions. Advance is retarded (lower values) as knock risk increases.

tempo_injecao_ms — float32

Fuel injector pulse width in milliseconds. Proportional to the quantity of fuel delivered per cycle. Naturally aspirated small-displacement engines register values around 1–7 ms at part load; large-displacement biturbo engines at full load may exceed 20 ms.

pressao_turbo_bar — float32

Intake manifold boost pressure in bar (gauge). For aspirado entries this value is zero or marginally negative (due to throttle restriction). For turbinado and biturbo entries, values range from approximately 0.01 to 1.82 bar, with biturbo configurations reaching the upper end of that range.

rpm_atual — int32

Current engine speed in revolutions per minute. Range spans from approximately 700 RPM (warm idle) to 7,000 RPM (redline). Distribution differs across engine families: naturally aspirated engines include more high-RPM samples; turbocharged engines feature more low-to-mid RPM content.

temp_agua_c — float32

Engine coolant temperature in degrees Celsius. Ranges from near-ambient values (~0–20 °C) representing cold-start conditions up to approximately 110 °C. Operating temperature is typically 80–100 °C. High values may indicate thermostat failure or coolant loss scenarios.

temp_ar_admissao_c — float32

Intake air temperature in degrees Celsius. Reflects both ambient conditions and heat soak from the engine bay. Range is approximately 0–55 °C. Elevated intake temperatures reduce effective octane rating and can induce detonation; this variable has high predictive weight for the detonacao label.

carga_tps_map_pct — float32

Engine load percentage derived from throttle position sensor and manifold absolute pressure. Ranges from near 0% (closed throttle, overrun) to approximately 100% (wide-open throttle). Normalized to a 0–100 scale.

tensao_bateria_v — float32

Battery / alternator voltage in volts. Represents the electrical system voltage as seen by the ECU. Typical operating range is 12.0–15.0 V. Values outside this range can indicate charging system faults and affect injector timing due to dead-time compensation.

marcha — int32

Engaged transmission gear. Integer values from 1 to 6. Combined with rpm_atual and carga_tps_map_pct, this variable encodes the vehicle speed and driveline load state, which conditions torque delivery.

torque_nm — float32

Crankshaft torque output in Newton-meters. This is a primary prediction target. Small naturally aspirated engines (NA_1.0) may output as little as 3 Nm at idle/overrun; large biturbo units (BT_5.0) can reach approximately 800 Nm at peak load.

potencia_cv — float32

Engine power in CV (cheval vapeur, equivalent to metric horsepower, 1 CV ≈ 0.7355 kW). A derived secondary target. Peak values for BT_5.0 configurations exceed 550 CV. Power is close to zero under overrun conditions regardless of engine size.

sonda_lambda_afr — float32

Wideband oxygen sensor output expressed as Air-Fuel Ratio (AFR). Stoichiometric combustion for gasoline is approximately 14.7. Values above ~15.0 indicate lean operation (excess air); values below ~14.0 indicate rich operation (fuel enrichment). Deviations from stoichiometry appear under wide-open throttle enrichment (rich) and overrun fuel cut (lean/open-loop).

egt_c — float32

Exhaust gas temperature in degrees Celsius, measured pre-turbine (on forced induction variants) or at the exhaust manifold. Ranges from approximately 700 °C at light load up to 1,080 °C near peak power. EGT is a critical thermal protection variable for turbine and piston crown durability.

pressao_oleo_bar — float32

Engine oil pressure in bar. Ranges from approximately 2.0 bar at warm idle up to approximately 6.0 bar at high RPM and cold oil. Low oil pressure at a given RPM is an anomaly flag for bearing wear or oil starvation.

detonacao — int32

Binary detonation (knock) event flag. 1 indicates a knock event was detected in that operating state; 0 indicates clean combustion. This is the primary classification target. The dataset preserves realistic class imbalance — knock events represent a minority of records, consistent with a properly calibrated production ECU map.

---

Engine Taxonomy

The following matrix summarizes the engine variants available and their characteristic operating envelopes.

engine_id	tipo_motor	Displacement	Boost ceiling (bar)	Typical torque ceiling (Nm)	Typical power ceiling (CV)
NA_1.0	aspirado	1.0 L	—	~80	~55
NA_1.6	aspirado	1.6 L	—	~130	~75
NA_2.0	aspirado	2.0 L	—	~180	~130
NA_3.5	aspirado	3.5 L	—	~350	~250
T_1.0	turbinado	1.0 L	~0.9	~200	~115
T_1.6	turbinado	1.6 L	~0.8	~240	~150
T_2.0	turbinado	2.0 L	~0.9	~300	~200
T_2.5	turbinado	2.5 L	~1.3	~370	~260
BT_2.0	biturbo	2.0 L	~1.5	~470	~310
BT_3.0	biturbo	3.0 L	~1.0	~580	~380
BT_4.0	biturbo	4.0 L	~1.8	~720	~560
BT_5.0	biturbo	5.0 L	~1.7	~820	~470

Values above are approximate upper bounds observed in the dataset and not hard-coded constraints.

---

Statistical Characterization

The following summaries are derived from the visible data samples and reflect expected distributions across the full 50M-record corpus.

RPM distribution: Right-skewed within each engine family. Modal operating range is 1,000–4,000 RPM across the dataset. High-RPM samples (>6,000) are present but under-represented, consistent with naturalistic driving distributions.

Torque and power: Strongly correlated with engine_id, rpm_atual, carga_tps_map_pct, and pressao_turbo_bar. Biturbo engines exhibit the widest variance in both targets. Under overrun (low TPS, high gear), torque may fall to single-digit Nm values regardless of displacement.

Lambda (AFR): Tightly clustered around 14.2–15.2 for closed-loop stoichiometric operation. Outliers appear at full load (enrichment, AFR ~10–13) and at closed throttle (open-loop lean, AFR ~16–17). Lambda is a strong predictor of detonacao at extreme values.

EGT: Monotonically increases with load and RPM. Biturbo engines show consistently higher EGT due to higher thermal load on pre-turbine gas. Cold-start and idle samples exhibit EGT values below 800 °C.

Detonation prevalence: The detonacao column is imbalanced. Knock events are correlated with high load, elevated intake air temperature, aggressive ignition advance, and lean AFR. Dataset users should apply appropriate class-weighting or resampling strategies during model training.

Oil pressure: Follows a predictable RPM-dependent curve with cold-oil spikes (high pressure at low RPM and low temperature) and warm-idle troughs. Outliers represent simulated sensor drift or fault injection.

---

Data Access

Hugging Face Datasets (Python)

from datasets import load_dataset

ds = load_dataset("Kodjaoglanian/ThermoDrive-ECU-50M")
df = ds["train"].to_pandas()

Direct Parquet Access

import pandas as pd

url = "hf://datasets/Kodjaoglanian/ThermoDrive-ECU-50M/data/train-*.parquet"
df = pd.read_parquet(url)

Dask (recommended for full 50M corpus)

import dask.dataframe as dd

df = dd.read_parquet("hf://datasets/Kodjaoglanian/ThermoDrive-ECU-50M/data/train-*.parquet")

Polars

import polars as pl

df = pl.scan_parquet("hf://datasets/Kodjaoglanian/ThermoDrive-ECU-50M/data/train-*.parquet")

Filtering by Engine Type

# Load only biturbo records
from datasets import load_dataset

ds = load_dataset(
    "Kodjaoglanian/ThermoDrive-ECU-50M",
    split="train"
).filter(lambda x: x["tipo_motor"] == "biturbo")

---

Benchmark Tasks

The following tasks define concrete evaluation protocols for this dataset.

Task 1 — Knock Detection (Binary Classification)

Target: detonacao Input features: All 15 numerical and categorical columns excluding the target. Metric: Area Under the ROC Curve (AUC-ROC), Average Precision (AP). Baseline: Gradient-boosted tree (XGBoost / LightGBM) with default hyperparameters. Note: Class imbalance must be handled explicitly. Report metrics on a stratified held-out split.

Task 2 — Torque Regression

Target: torque_nm Input features: All columns excluding torque_nm and potencia_cv. Metric: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R². Evaluation protocol: Stratified split by engine_id to ensure all engine types appear in both train and evaluation sets.

Task 3 — Cross-Engine Generalization

Objective: Train on aspirado records, evaluate on turbinado and biturbo records. Target: torque_nm or detonacao. Metric: Relative degradation in MAE or AUC-ROC versus in-domain evaluation. Research value: Quantifies domain shift between naturally aspirated and forced induction thermodynamic regimes.

Task 4 — Thermal Anomaly Detection

Objective: Detect records where temp_agua_c > 105 °C in combination with pressao_oleo_bar < 2.5 bar as compound fault states. Task type: Unsupervised / semi-supervised anomaly detection. Approach: Isolation Forest, Autoencoder, or one-class SVM on a subset excluding known fault records.

---

Known Limitations

Synthetic origin. All records are generated via a simulation pipeline, not captured from physical vehicles or engine test benches. Real ECU logs exhibit noise patterns, sensor quantization artifacts, and temporal autocorrelation that are not fully reproduced here.

No temporal structure. Records are independent snapshots. This dataset is not suitable for time-series architectures (LSTM, Temporal Convolutional Networks) without additional data engineering to reconstruct plausible operational sequences.

No fuel type variation. The dataset assumes a single gasoline fuel type with fixed octane rating assumptions. Ethanol blends, flex-fuel operation, and diesel combustion are not represented.

Idealized sensor model. Sensor noise follows simplified distributions. Physical sensors exhibit non-linear drift, temperature-dependent offset, and quantization at specific bit depths — none of which are modeled in the current version.

Geographic / regulatory invariance. No altitude, humidity, or ambient pressure variation is encoded. Real ECU maps account for barometric pressure compensation, which affects both boost targets and ignition advance.

Single operating mode. The dataset does not differentiate between ECU operating modes (e.g., sport, economy, launch control). All records reflect a single unified calibration map per engine type.

---

License and Citation

License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Use of this dataset for commercial purposes is not permitted without explicit authorization from the dataset author. Attribution is required for all research and derivative works.

Full license text: https://creativecommons.org/licenses/by-nc/4.0/

Dataset page: https://huggingface.co/datasets/Kodjaoglanian/ThermoDrive-ECU-50M

Citation (BibTeX):

@dataset{kodjaoglanian2026thermodrive,
  author    = {Kodjaoglanian},
  title     = {ThermoDrive-ECU-50M: A Large-Scale Synthetic ECU Telemetry Dataset for Automotive ML Research},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Kodjaoglanian/ThermoDrive-ECU-50M},
  license   = {CC BY-NC 4.0}
}

---

Documentation version 1.0 — June 2026.