Publication:
Censoring-aware neural survival modeling for catalyst lifetime prediction in industrial campaigns
| cris.virtual.department | #PLACEHOLDER_PARENT_METADATA_VALUE# | |
| cris.virtual.department | #PLACEHOLDER_PARENT_METADATA_VALUE# | |
| cris.virtual.orcid | 0000-0002-9524-4205 | |
| cris.virtual.orcid | 0000-0001-6600-1792 | |
| cris.virtualsource.department | 7bac28ac-f3c2-462d-aea4-cc71c4892295 | |
| cris.virtualsource.department | 3d0467d5-8f2f-463b-9a89-cd2e89911f08 | |
| cris.virtualsource.orcid | 7bac28ac-f3c2-462d-aea4-cc71c4892295 | |
| cris.virtualsource.orcid | 3d0467d5-8f2f-463b-9a89-cd2e89911f08 | |
| dc.contributor.author | Tysebaert, Stefaan | |
| dc.contributor.author | Deschrijver, Dirk | |
| dc.contributor.author | Couckuyt, Ivo | |
| dc.date.accessioned | 2026-11-03T08:57:20Z | |
| dc.date.available | 2026-11-03T08:57:20Z | |
| dc.date.createdwos | 2026 | |
| dc.date.issued | 2026 | |
| dc.description.abstract | Catalyst-driven shutdowns are a major source of uncertainty in utility and production planning at chemical sites. Unfortunately, for data-driven solutions, catalyst datasets are often incomplete: only a small number of campaigns reach observed end-of-life, while many are terminated earlier by operational events. The challenge involves incorporating incomplete campaigns into training. This work proposes a survival-based neural sequence modeling framework for catalyst lifetime prediction. Inspired by survival analysis and remaining useful life methods, lifetime is represented as an operational budget, where steam generation serves as a proxy for cumulative reaction exposure. A population-level model provides campaign-level budget estimates and upper bounds for censored campaigns. A parametric Accelerated Failure Time (AFT) likelihood is used to train neural sequence models on exact and censored observations. The AFT formulation is combined with Monte Carlo dropout to capture uncertainty, while conformal calibration is applied to improve coverage. The framework is evaluated on a real-life industrial dataset comparing LSTM, attention-augmented LSTM, and patch-based transformer models. Results show these models outperform a traditional population-level baseline model. Moreover, the incorporation of censored campaigns provides an additional reduction in prediction error. Overall, the proposed approach reduces prediction error by more than 50% over the baseline. These results demonstrate that sequence modeling and survival-based techniques achieve good performance on censored data, while the AFT formulation provides a unified basis for censoring-aware learning, multitask regularization, and uncertainty quantification. These promising results serve as a practical basis for data-efficient catalyst-lifetime forecasting in plant operations. | |
| dc.description.wosFundingText | This work was funded by the Flemish Agency for Innovation and Entrepreneurship (VLAIO) through the Baekeland Research Project HBC.2025.0227, entitled "Physics-informed spatio-temporal modeling and robust optimization for energy & utility management decision-making". | |
| dc.identifier.doi | 10.1016/j.compchemeng.2026.109923 | |
| dc.identifier.issn | 0098-1354 | |
| dc.identifier.uri | https://imec-publications.be/handle/20.500.12860/60704 | |
| dc.language.iso | eng | |
| dc.provenance.editstepuser | greet.vanhoof@imec.be | |
| dc.publisher | PERGAMON-ELSEVIER SCIENCE LTD | |
| dc.source.beginpage | 109923 | |
| dc.source.endpage | 12 | |
| dc.source.journal | COMPUTERS & CHEMICAL ENGINEERING | |
| dc.source.numberofpages | 12 | |
| dc.source.volume | 216 | |
| dc.subject.keywords | REGRESSION | |
| dc.title | Censoring-aware neural survival modeling for catalyst lifetime prediction in industrial campaigns | |
| dc.type | Journal article | |
| dspace.entity.type | Publication | |
| imec.internal.crawledAt | 2026-11-01 | |
| imec.internal.source | crawler | |
| imec.internal.wosCreatedAt | 2026-11-01 | |
| Files | Original bundle
| |
| Publication available in collections: |