Katulu Uniwear Dataset

A tidy dataset for multi-material tool wear, suitable for prognostic modeling.

As part of condition monitoring and predictive maintenance solutions, the challenge in tool wear is focused on being able to predict the current wear and forecast the remaining useful life.

A model for such solutions should be able to work for multiple tool types and workpiece materials. This could be in an online setting whereby the model is trained as data arrives or an offline setting where historical recordings are used to build a model.

In this direction, the need for a multi-material dataset, both tool, and workpiece, is required to develop robust models. To fulfill this purpose, we prepare and merge two public milling datasets for easy use in machine learning tasks, downsampled and smoothed.

Data Bundles

Under the data directory, you find the following data bundles:

• NUAA: nuaa_orthogonal_bundle_high_resolution.csv : ~15 Hz smoothed bundle of NUAA orthogonal experiments.
• PHM2010: phm2010_bundle_high_resolution.csv : ~25 Hz smoothed bundle of NUAA orthogonal experiments.
• Uniwear: uniwear.csv : ~2 Hz aligned intersection of nuaa & phm2010.

Materials Summary

Datasets are produced with different workpiece and tool materials. Here, we summarize this in a table.

Dataset	Workpiece	Tool
PHM2010	Stainless steel (HRC52)	Tungsten Carbide
NUAA	Titanium (TC4)	Solid Carbide

NUAA

The tool wear dataset comes from NUAA Ideahouse, released on IEEE's dataport.

Orthogonal experiments are part of the dataset that contains different variations: 3 factors and 3 levels (W1-W9), with the following parameters/factors fixed per experiment :

• fz feed per tooth (mm/rev) (feed_per_tooth ) [0.045, 0.05, 0.055]
• n spindle speed (rev/min) (spindle_speed ) [1750, 1800, 1850]

We only bundle these so-called orthogonal experiments. Experiments use titanium workpiece (TC4) and solid carbide cutting tool (such as Tungsten carbide).

data/nuaa_orthogonal_bundle_high_resolution.csv contains 9 experiments with raw data sampled at ~15 Hz. This is still downsampled from the original dataset but high resolution compare to extracted features in uniwear (see Uniwear dataset). We distinguish experiments with experiment_tag column.

Apart from timestamp in seconds and tool_wear in mm, the following signals appear in the dataset, column names as follows :

• axial_force
• bending_moment_x
• bending_moment_y
• torsion
• vibration1
• vibration2
• spindle_power
• spindle_current
• vibration_x
• vibration_y
• force_z

Experiment fixed variations are also given in the data set for information, column names are as follows:

• feed_per_tooth
• spindle_speed
• axial_cutting_depth

Experiment and dataset tags are also provided.

• experiment_tag
• dataset_tag

PHM2010

PHM2010 was a data challenge given by PHM society in 2010. We bundle 3 of the cutting experiments c1, c4, and c6. Stainless steel (HRC52) workpiece is used with the cutting tool being tungsten carbide. Release in IEEE dataport as well. Sensor signals.

• force_x
• force_y
• force_z
• vibration_x
• vibration_y
• vibration_z
• acoustic_emission_rms

Dataset is distinguished with 'dataset_tag', here is 'phm2010'.

Uniwear

We have merged two datasets NUAA (W1-W9 recordings) and PHM2010 (c1, c4, and c6 recordings) and align their timestamps, i.e., ~2 Hz sampling and smoothed. We use overlapping signal types on the following columns:

• force_x
• force_y
• force_z
• vibration_x
• vibration_y
• vibration_z

Label column is tool_wear in mm.

Other important columns

• timestamp : seconds since cutting.
• dataset_tag : nuaa or phm2010.
• experiment_tag : W1-W9, c1, c4, c6.

Directories

• data: Contains all bundled datasets.
• notebooks : Python notebook for plotting.

License

The bundled datasets are licensed under Creative Commons Attribution 4.0 International License . The rest is licensed under GPLv3.