Nonastreda: Multimodal Dataset for Identifying Tool Wear Condition

Published: 9 January 2025| Version 1 | DOI: 10.17632/m892d2wtzh.1
Contributors:
Hubert Truchan,

Description

# Nonastreda: 9 Multimodal Dataset Featuring Time Series and Image Data for Flank Tool Wear Classification and Regression * Detailed description: 'Data in Brief' Journal (available soon) * Repository: https://github.com/hubtru/Impala * Repository: https://github.com/hubtru/Girape * Notebooks for converting forces_xyz_raw.mat into spectrograms, scalograms or wavelets: https://github.com/hubtru/Girape/tree/main/scripts # Overview: Nonastreda (Nona) * 'Nona' from Latin "ninth" * Dataset Size: 512 samples (instances, observations) * Modalities: 9 modalities * Tasks: * Classification: 3 classes (sharp, used, dulled) * Regression: 3 targets (flank wear [µm], gaps [µm], overhang [µm]) * Additional subtasks: * Uni/Multi-Modal Classification * Multilabel Regression * Anomaly Detection * Remaining Useful Life (RUL) Estimation * Signal Drift Measurement * Zero-Shot Flank Tool Wear Classification * Diagnostic Feature Engineering * Domain: industrial flank tool wear of the milling machine * Input (per sample): * Images: 1 tool image, 1 chip image, 1 workpiece image * Mel-Spectrograms: x, y, z axes (3 images) * Complex Morlet Scalograms: x, y, z axes (3 images) * Extra Modalities: raw (time-series) force signals in x, y, z axes * Output: * Machine state classes: sharp, used, dulled * Regression targets: flank wear [µm], gaps [µm], overhang [µm] * Evaluation metrics: * Classification: accuracies, precision, recall, F1-Score, ROC curve * Regression: MAE, MSE, RMSE * Data splitting: * Protocol: 10-Fold Cross Validation * Training and Validation: data from 9 tools * Testing: data from the 10th tool * Results: accuracy averaged over ten splits * The dataset includes measurements from ten tools Extra Time-Series Modality * Raw forces signal in x, y, z axes is provided in `forces_xyz_raw.mat` file. * The `*.mat` file can be used with scripts from the Girape repository to generate spectrograms, scalograms, and wavelets. * Source force signals (Fx, Fy, Fz) allow experimentation with new types of feature engineering and embeddings, such as Shannon, Daubechies, or Morlet wavelets. * Sampling rate for force signals: 1 kHz. * forces_xyz.mat + Girape/scripts -> spectrograms or scalograms or wavelets Future Work * Improvements of (zero-shot flank) tool wear classification and regression. * Incorporating raw force signals (Fx, Fy, Fz) into multimodal studies. * Calculating new modalities using the raw force signals (Fx, Fy, Fz). * Conducting experiments on: * Anomaly Detection * Remaining Useful Life (RUL) estimation * Signal Drift measurement * Designing Diagnostic Feature Engineering. * Modalities Correlation Analysis. # Data Structure Nonastreda/ │ ├── chip/ ├── scal/ │ ├── x/ │ ├── y/ │ └── z/ ├── spec/ │ ├── x/ │ ├── y/ │ └── z/ ├── tool/ │ ├── work/ │ ├── labels.csv ├── labels_reg.csv └── forces_xyz_raw.mat

Files

Steps to reproduce

* Detailed description: See the upcoming publication in the 'Data in Brief' journal. The Nonastreda dataset is obtained from a real industrial milling device processing material with a shaft milling tool. During the milling process, time-series and image data were collected to model industrial tool wear. * **Time-Series Data**: - Three force signal sequences (Fx, Fy, Fz) were collected using: - Industrial dynamometer - Amplifier - Bus coupler - Industrial PC * **Image Data**: - Images were captured using an industrial unit microscope.

Institutions

Leibniz Universitat Hannover

Categories

Feature Extraction, Regression Analysis, Image Classification, Time Series, Drift Analysis, Multimodality, Remaining Useful Life, Device Efficiency, Multimodal Learning, Multimodal Deep Learning

Licence