Oxy-fuel Cutting Task State Image Dataset

Associated Paper: CNN-based Task State Estimation for Safer Automation of Oxy-fuel Metal CuttingPaper Status: Published (IEEE CASE 2023, doi: 10.1109/CASE56687.2023.10260647)

PAPER ABSTRACT:

The industrial operation of oxy-fuel metal cutting via gas torches involves tasks such as ignition, preheating, and combustion along the target surface. Automated oxy-fuel cutting systems are exposed to risks and anomalies that can lead to incorrect actions and safety hazards. In this paper, we develop a classifier for online task state estimation to assess the cutting robot’s actions, detect anomalies, and reduce the risk of hazards. Using representative footage from our robotic cutting experiments, we curate an image dataset labeled with four types of cutting task states. Using deep learning methods, we design and train a convolutional neural network model for classifying the cutting task state from input images. The classifier architecture is optimized for rapid inferences during online estimation. After evaluation, our classifier achieves an overall accuracy of 93.8% with high inference speeds on two types of representative hardware. Our ‘Oxy-fuel Cutting Task State’ (OCTS) dataset is available at  doi.org/10.5281/zenodo.7734951.

DATASET DESCRIPTION:

The Oxy-fuel Cutting Task State (OCTS) dataset contains image data from footage recorded during a series of robotic oxy-fuel metal cutting experiments labeled with one of four cutting task states, identified using their prominent feature:

Torch flame (TF): Associated with the vision system calibration task.

Preheating pool (PP): Associated with the surface conditioning task.

Combustion pool (CP): Associated with the combustion control task.

Not applicable (NA): Associated with halting operations since none of the previous elements are identified; this is an anomaly.

The dataset files consist of:

Data: Available as a ZIP archive split into 5 volumes (~2.8 GB each).

Labels: Available in JSON format.

Metadata: Available in CSV and PDF formats, contains the individual experiment set IDs, their dates and times, their total frame counts, and their label-wise frame counts.

DATASET LICENSE:

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

DATA INSTRUCTIONS:

To extract the ZIP archive, download all five ZIP volumes into a common directory and extract the file dataset.zip.001. After extraction, 50 directories are obtained ('S01', 'S02', …, 'S50'). These contains the raw data (image frames) of each of the 50 individual cutting experiments. The image filenames are their frame number ('000000.jpeg', '000001.jpeg', …). All images are in JPEG format. All image filenames are their 6-digit frame number (includes leading zeroes such as in '004021.jpeg') for a particular experiment. Essentially, this is the sequential image data from the footage of each experiment.

JSON INSTRUCTIONS:

The JSON file partitions the images of each experiment set into the four labels. The first JSON level contains the experiment set ID as a string ('S01', 'S02', …, 'S50'). The second JSON level contains the four labels ('TF', 'PP', 'CP', 'NA') for each set ID. The third JSON level contains arrays of strings containing the frame number (filename without extension) of each image (e.g., ['000000', '000001', , …]). Usage of the JSON file is illustrated in the following Python code snippet:

import json

with open ("labels.json") as json_file:
labels = json.load(json_file) # `labels` is a dictionary.

# Get all image filenames (frame numbers) from experiment `S01` in the `NA` label.
labels['S01']['NA'] # returns list of image filenames (strings)
#output: ['003843', '003844', …, '003978', '003979']

Thus, the labels are retrieved for each of the images in each of the set IDs.

METADATA INSTRUCTIONS:

The metadata associates the set ID of each experiment to its recording date and time. In addition, it lists the total frames of each experiment and the frame count in each of the four labels ('TF', 'PP', 'CP', 'NA'). This is available in PDF format for convenient viewing but also in CSV format.