Codifying Failure Knowledge for Industrial Decision Support: An Evaluation of Failure-Type-Aware Generative Augmentation
Interdisciplinary Journal of Information, Knowledge, and Management · 2026 · 21 · 23
DOI: 10.28945/5904 ↗Kurz gefasst
Fachartikel in der Zeitschrift Interdisciplinary Journal of Information, Knowledge, and Management, 2026. Verfasst von Alexis Lazanas und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Aim/Purpose: This study investigates whether partitioning generative data augmentation by recorded failure mechanism improves industrial equipment failure detection, both overall and for each failure mechanism, compared with single-generator augmentation, conventional resampling, and a cost-sensitive baseline.
Background
Data-driven maintenance systems learn from operational histories in which failure is rare. Established remedies for this scarcity treat the minority class as one population, yet failures arise from distinct physical mechanisms with distinct sensor signatures. The homogeneity assumption is seldom examined, and its organizational consequence, decision support that systematically underrepresents infrequent failure modes, is seldom measured. Methodology: A leakage-safe comparison of eight strategies on the AI4I 2020 benchmark of 10,000 records, with a failure prevalence of 3.4% and four labeled failure mechanisms. The strategies were a cost-sensitive baseline, random undersampling, SMOTE, SMOTENC, Borderline-SMOTE, ADASYN, single-generator WGAN-GP augmentation, and a failure-type-aware design with one generator per failure mechanism. A Random Forest classifier received the same tuning under every strategy across 30 repeated stratified data splits. PR-AUC is the primary criterion; per-mechanism recall and comparisons at matched operating points are secondary criteria. Contribution: The study treats failure-mechanism labels as codified organizational knowledge and tests whether analytics can use that knowledge to improve failure detection. It demonstrates, through a fair, leakage-safe comparison, that partitioning augmentation by mechanism is no better than a tuned cost-sensitive baseline, and that the value of codified mechanism knowledge lies in revealing the failure modes that decision support misses.
Findings
Across 30 repeated splits, the cost-sensitive baseline achieved the highest mean PR-AUC (0.818). The failure-type-aware strategy (0.803) was not significantly different from the baseline or the interpolation-based methods; its advantage over single-generator augmentation (0.770) was consistent but not significant after accounting for dependence between splits. At matched operating points, augmentation offered no benefit over changing the baseline decision threshold, and neither generative strategy improved detection of the rarest mechanism, tool wear. With generator settings fixed in advance instead of tuned, both generative strategies performed worse. Once its settings were fixed, the failure-type-aware strategy required approximately 200 times the computation of the baseline per split, in addition to a one-time search of about four hours. Recommendations for Practitioners: Before investing in generative augmentation, tune a cost-sensitive model and set its decision threshold according to the local cost of a missed failure relative to an unnecessary inspection; in this study, that combination matched every augmentation strategy at a small fraction of the computational cost (around 3 seconds per model build versus around 9 minutes for the failure-type-aware strategy, after a one-time search of about four hours). If failure mechanisms are documented, use them to evaluate detection for each mechanism, as overall scores may mask poorly detected failure modes. This study did not estimate the economic return of recording mechanisms. Recommendation for Researchers: Comparisons of imbalance strategies should include a tuned cost-sensitive baseline with an adjusted decision threshold, and all strategies should receive the same tuning. Studies should compare strategies at matched operating points and across repeated data splits, with tests that account for dependence between splits. They should report the composition of the minority class and evaluate results for each subtype. Impact on Society: Unplanned industrial downtime carries costs in energy, materials, and worker safety. Detection systems that miss infrequent failure mechanisms concentrate those costs on events organizations are least equipped to handle; evaluating detection for each failure mechanism makes these blind spots apparent. Future Research: Evaluation on operational maintenance records with partial or noisy mechanism labels; comparison of partitioned generation with a label-conditioned generator of matched capacity and with tabular diffusion models; assessment of the fidelity and memorization of synthetic records; and evaluation against explicit maintenance cost models.
Zitieren
APA
Lazanas, A., & Kampouropoulos, G. (2026). Codifying Failure Knowledge for Industrial Decision Support: An Evaluation of Failure-Type-Aware Generative Augmentation. Interdisciplinary Journal of Information, Knowledge, and Management, 21, 23. https://doi.org/10.28945/5904
Vancouver
Lazanas A, Kampouropoulos G. Codifying Failure Knowledge for Industrial Decision Support: An Evaluation of Failure-Type-Aware Generative Augmentation. Interdisciplinary Journal of Information, Knowledge, and Management. 2026;21:23. doi:10.28945/5904
BibTeX
@article{Lazanas2026,
author = {Lazanas, Alexis and Kampouropoulos, Georgios},
title = {Codifying Failure Knowledge for Industrial Decision Support: An Evaluation of Failure-Type-Aware Generative Augmentation},
journal = {Interdisciplinary Journal of Information, Knowledge, and Management},
year = {2026},
volume = {21},
pages = {23},
doi = {10.28945/5904}
}
EUVIMED