RMIS Benchmark Datasets

Introduction

RMIS is a benchmark dataset collection for evaluating representation learning on multi-modal industrial signals. It brings together the datasets used in the RMIS benchmark, covering anomaly detection and fault diagnosis across four modalities: sound, vibration, voltage, and current.

This Hugging Face repository is meant to provide the benchmark datasets themselves in an easy-to-load format. If you are looking for the full benchmark codebase, evaluation pipeline, or leaderboard, please refer to the RMIS GitHub repository.

RMIS is closely related to FISHER:

• FISHER is the foundation model proposed for industrial signal representation.
• RMIS is the benchmark used to evaluate FISHER and other signal models.
• This Hugging Face repository hosts the dataset side of RMIS, while the GitHub repository hosts the benchmark code and evaluation pipeline.

In the current release, the dataset includes 19 configurations:

• Anomaly detection: dcase20, dcase21, dcase22, dcase23, dcase24, dcase25
• Fault diagnosis: iica, iiee, mafaulda_sound, mafaulda_vib, pu_cur, pu_vib, sdust_bearing, sdust_gear, umged_cur, umged_sound, umged_vib, umged_vol, wtpg

What is included

Each configuration corresponds to one benchmark subset. The exact schema varies by subset, but all configurations provide an audio column together with file-level metadata such as file_name, and task-specific annotations such as status, scene, mt, ori, domain, or related attributes.

The split names follow the benchmark tasks:

• ad: anomaly detection subsets
• fd: fault diagnosis subsets

Please note that this repository focuses on dataset hosting and loading. Detailed benchmark construction, preprocessing rationale, evaluation protocol, and model integration are documented in the RMIS GitHub repository.

Usage

Load one subset with datasets:

from datasets import load_dataset

ds = load_dataset("jiangab/RMIS", "dcase20", split="ad")
print(ds)
print(ds[0])

If you want decoded waveforms instead of deferred audio objects:

from datasets import load_dataset, Audio

ds = load_dataset("jiangab/RMIS", "dcase20", split="ad")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

sample = ds[0]
audio = sample["audio"]["array"]
sr = sample["audio"]["sampling_rate"]
print(audio.shape, sr)

Load another configuration in the same way:

from datasets import load_dataset

# fault diagnosis example
fds = load_dataset("jiangab/RMIS", "umged_vib", split="fd")
print(fds.column_names)

If you want to download the entire dataset repository at once instead of loading one configuration at a time, you can download the whole Hugging Face dataset repo locally:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="jiangab/RMIS",
    repo_type="dataset",
)
print(local_dir)

You can also use the Hugging Face CLI:

hf download jiangab/RMIS --repo-type dataset

Please note that load_dataset is typically used one configuration at a time, while whole-repository download is useful when you want all released subsets stored locally.

Recommended usage in the RMIS project

If your goal is to benchmark a model on RMIS, this dataset repository is only one part of the workflow. A typical setup is:

1. Load a subset from this Hugging Face dataset repository.
2. Extract signal representations with your model.
3. Evaluate the representations with the RMIS benchmark pipeline from the GitHub repository.

In other words, this repository provides the benchmark data, while the RMIS codebase provides the standardized evaluation.

Acknowledgements

RMIS is built from multiple public industrial signal datasets. We thank the original dataset creators for making these resources available.

If you believe that any content in this repository infringes your rights, please contact us and we will address the issue promptly.

Citation

If you find RMIS useful, please cite the following paper.

@article{fan2025fisher,
  title={FISHER: A Foundation Model for Multi-Modal Industrial Signal Comprehensive Representation},
  author={Fan, Pingyi and Jiang, Anbai and Zhang, Shuwei and Lv, Zhiqiang and Han, Bing and Zheng, Xinhu and Liang, Wenrui and Li, Junjie and Zhang, Wei-Qiang and Qian, Yanmin and Chen, Xie and Lu, Cheng and Liu, Jia},
  journal={arXiv preprint arXiv:2507.16696},
  year={2025}
}