README.md

---
pretty_name: CAD-TCM
license: cc-by-4.0
configs:
  - config_name: default
    data_files: data.csv
language:
  - en
task_categories:
  - tabular-classification
task_ids:
  - tabular-multi-class-classification
size_categories:
  - 100K<n<1M
tags:
  - anomaly-detection
  - continual-learning
  - continual-anomaly-detection
  - predictive-maintenance
  - steel-manufacturing
  - tandem-cold-mill
  - tabular
  - synthetic
---

# CAD-TCM

## Dataset Summary

**CAD-TCM** is a single-source continual anomaly detection benchmark scenario for predictive maintenance in steel manufacturing. It is derived from **TCM: Benchmark Datasets for Predictive Maintenance in Steel Manufacturing**, a synthetic dataset generated from a physics-based mathematical model of a 5-stand tandem cold mill (TCM), and converts the original tabular sensor data into a sequence of concept-grouped tasks.

The dataset contains **108,872 samples**, **8 tasks**, and has a reported **18.66% anomaly ratio in the test set**.


## Intended Use

This dataset is intended for research on:

- continual anomaly detection;
- continual learning for tabular data;
- predictive maintenance in industrial/manufacturing processes;
- robustness under distribution shift;
- task ordering in continual-learning benchmarks;
- forgetting and knowledge transfer across related process-monitoring concepts;
- benchmarking anomaly detectors under sequential task exposure.

The intended use is **defensive/industrial machine learning research**, i.e. detecting equipment faults and process anomalies to support predictive maintenance.

## Dataset Source

- TCM: Benchmark Datasets for Predictive Maintenance in Steel Manufacturing: `https://zenodo.org/records/11469702`

The source data is entirely **synthetic**, generated from a physics-based mathematical model of a 5-stand tandem cold mill simulating four fault types: reduction-scheme irregularities, work-roll friction increase, bearing failure (elevated motor torque), and motor efficiency loss.


## Dataset Files

The repository contains the following files:

| File | Description |
|---|---|
| `data.csv` | Main tabular dataset file. |
| `orderings.json` | Predefined task orderings for continual-learning evaluation. |
| `croissant.json` | Croissant metadata describing the dataset. |

## Dataset Structure

The main file is:

```text
data.csv
```

The dataset contains task metadata, binary labels, and numerical rolling-mill process/sensor features.

### Core Columns

| Column | Type | Description |
|---|---:|---|
| `task_id` | integer | Numeric identifier of the continual-learning task. |
| `task_name` | string | Name of the task, e.g. `tcm_0`. |
| `task_split` | string | Split assignment for the row. |
| `label` | integer | Binary anomaly label. `0` denotes normal operation and `1` denotes an anomalous/faulty condition. |

### Task Identifiers

The dataset contains the following task identifiers:

`tcm_0`, `tcm_1`, `tcm_2`, `tcm_3`, `tcm_4`, `tcm_5`, `tcm_6`, `tcm_7`

### Feature Columns

The remaining columns are numerical process/sensor features from the 5-stand tandem cold mill, including per-stand measurements. Representative examples include:

- `thickness_entry`, `thickness_exit`
- `width`
- `ys_entry`, `ys_exit` (yield strength)
- `work_roll_diam_1` … `work_roll_diam_5`
- `work_roll_mileage_1` … `work_roll_mileage_5`
- `reduction_1` … `reduction_5`
- `tension_0` … `tension_5`
- `roll_speed_1` … `roll_speed_5`
- `force_1` … `force_5`
- `torque_1` … `torque_5`
- `gap_1` … `gap_5`
- `motor_power_1` … `motor_power_5`

For the complete schema, see `croissant.json`.

## Task Orderings

The dataset provides six predefined orderings in `orderings.json`. These orderings define different continual-learning evaluation regimes over the same task set.

| Ordering | Task sequence |
|---|---|
| `curriculum_asc` | `tcm_2` → `tcm_5` → `tcm_7` → `tcm_6` → `tcm_4` → `tcm_0` → `tcm_1` → `tcm_3` |
| `curriculum_desc` | `tcm_3` → `tcm_1` → `tcm_0` → `tcm_4` → `tcm_6` → `tcm_7` → `tcm_5` → `tcm_2` |
| `generalization_desc` | `tcm_0` → `tcm_5` → `tcm_2` → `tcm_7` → `tcm_4` → `tcm_6` → `tcm_1` → `tcm_3` |
| `generalization_asc` | `tcm_3` → `tcm_1` → `tcm_6` → `tcm_4` → `tcm_7` → `tcm_2` → `tcm_5` → `tcm_0` |
| `smooth_drift` | `tcm_5` → `tcm_7` → `tcm_0` → `tcm_2` → `tcm_1` → `tcm_6` → `tcm_4` → `tcm_3` |
| `abrupt_drift` | `tcm_0` → `tcm_3` → `tcm_5` → `tcm_1` → `tcm_2` → `tcm_6` → `tcm_7` → `tcm_4` |

These orderings are intended to expose complementary continual-learning dynamics, including curriculum-like adaptation, generalization-oriented ordering, smooth drift, and abrupt drift.

## Dataset Creation
The details of dataset creation can be found in our paper: [link](https://arxiv.org/abs/2607.18289)

## Citation
When using the dataset, please cite:
```
@article{faber2026towards,
  title={Towards Principled Continual Anomaly Detection: A Systematic Framework and Benchmark Scenarios},
  author={Faber, Kamil and Smendowski, Mateusz and Corizzo, Roberto},
  journal={arXiv preprint arXiv:2607.18289},
  year={2026}
}
```