Reproducibility package for "Imbalanced failure prediction at the industrial edge: a cross-dataset, cross-device benchmark of cost-sensitive batch and online tree models"
Year: 2026
Type: software
Abstract: Benchmark code and experiment results supporting the paper "Imbalanced failure prediction at the industrial edge: a cross-dataset, cross-device benchmark of cost-sensitive batch and online tree models". The package contains the runners, ablation scripts and complete result logs for a matched, leakage-controlled evaluation of nine batch and three online tree configurations under chronological evaluation, validation-only cost-sensitive thresholding and prequential learning. Experiments were run directly on single-board computers: a Raspberry Pi 3 Model B Rev 1.2 and, for the SECOM experiments, additionally a Raspberry Pi 4 Model B Rev 1.5. Contents: - ai4i2020/ : benchmark runner, cost-sensitive and online-learning robustness scripts, and results for the AI4I 2020 dataset. Includes the canonical five-seed run behind the paper's main AI4I table, two independent reproduction runs (one per Raspberry Pi board), and the cost-ratio sweep and memory-replay robustness checks. Each run folder carries per-seed performance and resource logs, figures, and a metadata.json recording the board, operating system, Python version and exact library versions used. - secom/ : data-preparation script, benchmark runner, three ablation scripts (feature-count ablation, split-type ablation, drift diagnosis), the processed SECOM dataset produced by the preparation step, the aggregate tables behind the paper's SECOM results, and per-seed run logs from both Raspberry Pi boards. The raw AI4I 2020 and SECOM datasets are not redistributed here. Both are third-party public datasets at the UCI Machine Learning Repository and should be downloaded from there: AI4I 2020 Predictive Maintenance Dataset (Matzka, 2020), https://doi.org/10.24432/C5HS5C ; SECOM Dataset (McCann and Johnston, 2008), https://doi.org/10.24432/C54305 . Trained model files are also excluded. The code and the fixed random seeds recorded in each run's metadata regenerate identical models, which is itself one of the cross-device findings the paper reports. See README.md at the top of the archive, and the per-folder READMEs, for setup and run instructions.
Author: Abd. Hallim, Hallim
Institution: Binus University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203571014