RMIS: Representation of M5 Industrial Signals

A comprehensive benchmark for evaluating foundation models on multi-modal series-like industrial signals.

Paper Platform Python Pytorch RMIS GitHub RMIS Hugging Face License

🚀 News

  • [2026.2.13] The RMIS benchmark is released.
  • [2025.7.23] FISHER-tiny, FISHER-mini and FISHER-small are released.

📊 Overview

RMIS is a comprehensive benchmark for evaluating foundation models on M5 (multi-modal, multi-sampling-rate, multi-scale, multitask, and minim fault) series-like industrial signals. The RMIS benchmark comprises 6 anomaly detection datasets and 13 fault diagnosis datasets, covering 4 distinct modalities (sound, vibration, voltage, current) with an unprecedented volume of 1.2k hours. By open-sourcing RMIS, we hope to provide a fair, convincing and off-the-shelf benchmark for the community to easily evaluate their signal models.

View the leaderboard and join us!

View Leaderboard

Anomaly Detection

Anomaly detection is to predict whether a signal is anomalous or not without any prior knowledge of anomaly. The anomaly detection part of RMIS is sourced from the annual DCASE challenge, where we follow the DCASE challenge rules to evaluate the model.

Dataset Modality Sampling Rate Volume Num Machines
DCASE 2020Sound16 kHz153 h6
DCASE 2021Sound16 kHz165 h7
DCASE 2022Sound16 kHz139 h7
DCASE 2023Sound16 kHz50 h14
DCASE 2024Sound16 kHz49 h16
DCASE 2025Sound16 kHz45 h15

Fault Diagnosis

Fault is to identify the specific fault type (health state) of a signal with labeled data provided in advance. The fault diagnosis part of RMIS is sourced from 7 public fault diagnosis datasets. To ensure proper difficulties for evaluation, we introduce sealed train-test split, where segments from the same channel of the same recording can not appear in both train and test sets.

Dataset Modality Sampling Rate Volume Task
IICASound48 kHz47 hCompressor Leakage
IIEESound44.1 kHz1 hElectric Engine Fault
WTPGVibration48 kHz14 hPlanetary Gearbox Fault
MaFaulDaSound, Vibration50 kHz19 hBearing Fault
SDUSTVibration25.6 kHz42 hBearing / Gear Fault
UMGEDSound, Vibration, Voltage, Current51.2 kHz469 hGear Eccentricity
PUVibration, Current64 kHz9 hBearing Fault

🌟 Motivation

The RMIS benchmark is launched in conjunction with our FISHER model, which is the first multi-modal Foundation model for Industrial Signal compreHEnsive Representation. Our pioneering endeavor has led to an exciting discovery: M5 industrial signals, though hugely heterogenous, can surprisingly be modeled by an universal foundation model. This model, powered by extensive training data, exhibits an unprecedented level of generalization across various modalities and tasks, and thereby opening up the scaling era for industrial signals.

Superior modeling calls for more comprehensive evaluation. Although there are dozens of open-sourced signal datasets, there has never been an open-sourced and comprehensive benchmark that ensembles all these datasets, and the experiment setups of these datasets in previous works differ more or less. To this end, RMIS is designed to be a fair, convincing and off-the-shelf benchmark with the following key features:

  • ✓

    Open-source

    RMIS is completely open-sourced. We warmly welcome fellow researchers to test their models on our benchmarks!

  • ★

    Unprecedented Scale

    RMIS comprises both anomaly detection and fault diagnosis datasets, covering 4 modalities with an unprecedented volume of 1.2k hours.

  • 🔒

    Sealed Train-Test Split

    Most previous works adopt random train-test splits, which yield over-optimistic results due to signal stationarity. In RMIS, we adopt sealed train-test split, where segments from the same channel of the same recording cannot appear in both train and test sets.

🚩 Getting Started

The RMIS pipeline can be setup in two streamlined steps:

â‘  RMIS Download

Get started by cloning our repository, downloading the preprocessed 10s wav datasets (available on Tsinghua Cloud and Hugging Face), and selecting from 6+ integrated models like BEATs and EAT.

Download Guide GitHub Repo

â‘¡ RMIS Evaluation

RMIS is designed to disentangle model representations with downstream tasks. After local configurations, you can easily add and evaluate a model by simply one command!

Test Integrated Models Test Your Model

Acknowledgements

The RMIS benchmark is sourced from 13 public datasets: DCASE20, DCASE21, DCASE22, DCASE23, DCASE24, DCASE25, IICA, IIEE, WTPG, MaFaulDa, SDUST, UMGED and PU. We sincerely thank these contributors for open-sourcing their data.

Citation

The RMIS benchmark is licensed under the MIT License. Individual datasets within RMIS may have their own licenses. If you find RMIS useful in your research, please cite our work:

BibTeX 
@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}
        }