Project Title:
Data-Driven Localization of Magneto-Mechanical Resonators from Time Signals

Principal Investigators / Authors:
* Artyom Tsanda, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
* Konrad Scheffler, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
* Sarah Reiss, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
* Jonas Faltinath, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
* Janik Bach, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
* Pascal Stagge, Fraunhofer Research Institution for Individualized Medical Technology and Engineering IMTE, Luebeck, Germany
* Fabian Mohn, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
* Tobias Knopp, Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany

Funding Acknowledgement:
This data set was generated as part of the DFG-funded project CRC 1615: SMART Reactors for Future Process Engineering (DFG Project Number: 503850735).

# General Information
## Dataset title
Raw data and trained model checkpoint for the paper "Data-Driven Localization of Magneto-Mechanical Resonators from Time Signals."
## Short Description
This dataset supports the training and experimental evaluation of a deep learning method for localizing a magneto-mechanical resonator (MMR) from multichannel voltage time signals. It contains simulated signals and metadata, recorded system noise, experimental signals, robot-recorded ground-truth positions, and the trained model checkpoint used in the study.
## Date of Data Collection
2026
## Geographical Coverage
Hamburg, Germany.
## Keywords
* Magneto-mechanical resonator
* Magnetic localization
* Wireless sensing
* Deep learning
* Time-series regression

# Methodological Information
## Data Collection and Processing
The experimental data were acquired with a 16-coil inductive array arranged as a 4 x 4 grid. The MMR consists of two antiparallel cylindrical neodymium-iron-boron magnets in a 3D-printed housing. The resonator was positioned by a three-axis robot on a 5 x 5 x 3 grid, resulting in 75 ground-truth positions. Its orientation remained fixed during the experiment.

The signals were sampled at 61,035 Hz. The original signals contain 53,700 samples per channel and were downsampled by a factor of 20 for the released model inputs, resulting in 2,685 samples per channel. The experimental data contain 75 positions and 40 frames per position, with 16 channels per frame.

The simulated data were generated from the MMR signal model with randomized position, orientation, natural frequency, quality factor, initial phase, and maximum deflection angle. The training and validation signals contain 100,000 and 10,000 samples, respectively. Recorded system noise is included for data augmentation during training.
## Experimental Design / Study Context
The study evaluates end-to-end localization of an MMR from time signals. Experimental measurements were acquired at 75 positions on a robot-controlled 5 x 5 x 3 grid. Each position has repeated static measurements, and the released experimental array contains 40 frames per position. The trained model estimates the three-dimensional position from the 16-channel signal.
## Data Validation and Quality Assurance
The experimental ground-truth positions were recorded by the positioning robot. The released arrays, metadata files, and checkpoint were checked for consistency with the data organization and processing described in the paper.

# Data and File Overview
## Files
The uploaded files are flattened from the original directory structure. The complete release contains 21 files:

* README.md - dataset description and metadata
* flatten.sh - utility used to flatten and restore the original directory paths
* experiment__data.npy - experimental signals
* experiment__gt.csv - robot-recorded ground-truth positions
* model__checkpoint.pt - trained model checkpoint
* training__noise.npy - recorded system-noise samples
* training__train_chunk01.npy through training__train_chunk10.npy - simulated training signals
* training__train_meta.csv - training metadata and targets
* training__val_chunk1.npy - simulated validation signals
* training__val_meta.csv - validation metadata and targets
* training__test_chunk1.npy - simulated test signals
* training__test_meta.csv - test metadata and targets
## Naming Convention
The files were flattened for upload. The double underscore (`__`) separates the original directory name from the filename. For example, `training__train_meta.csv` corresponds to `training/train_meta.csv`.

The original directory structure was:

```text
.
├── experiment
│   ├── data.npy
│   └── gt.csv
├── model
│   └── checkpoint.pt
└── training
	├── noise.npy
	├── test_chunk1.npy
	├── test_meta.csv
	├── train_chunk01.npy
	├── train_chunk02.npy
	├── train_chunk03.npy
	├── train_chunk04.npy
	├── train_chunk05.npy
	├── train_chunk06.npy
	├── train_chunk07.npy
	├── train_chunk08.npy
	├── train_chunk09.npy
	├── train_chunk10.npy
	├── train_meta.csv
	├── val_chunk1.npy
	└── val_meta.csv
```

## Number of Records / Observations
The release contains 100,000 simulated training signals, 10,000 simulated validation signals, and 10,000 simulated test signals. The experimental array contains 75 positions with 40 frames per position. The simulated signal chunks have shape `(10000, 16, 2685)`, and the experimental array has shape `(75, 40, 2685, 16)`, with the channel dimension last.

# Access and Licensing Information
Repository and Persistent Identifier:
Published via TORE, DOI: 10.15480/882.18147
License for Use:
CC0 1.0 (Public Domain Dedication)
Access Restrictions:
Open
Text for Citation:
A. Tsanda, K. Scheffler, S. Reiss, J. Faltinath, J. Bach, P. Stagge, F. Mohn, and T. Knopp, Supplementary Data for the Paper: “Data-Driven Localization of Magneto-Mechanical Resonators from Time Signals” (2026). doi:10.15480/882.18147

# Reproducibility and Software Dependencies
To reproduce the original study follow the instructions from the following code repository: https://github.com/IBIResearch/mmr-learned-localization-from-signals.git

# Ethical and Legal Aspects
Data Protection:
No personal or sensitive data is contained in this dataset. The study involved only non-human materials (PMMA containers, chemical mixtures, and polypropylene inserts) and did not include human subjects or animals. Therefore, no anonymisation or pseudonymisation measures are applicable.
Consent Statement:
Not applicable. No human participants or animals were involved in the data collection for this study.

# Versioning and Updates
Version Number:
v1.0
Date of Release:
2026-09-02

# Contact Information
Artyom Tsanda
Institute for Biomedical Imaging, Hamburg University of Technology, Hamburg, Germany
artyom.tsanda@tuhh.de