Nonastreda: Multimodal Dataset for Identifying Tool Wear Condition

# 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