--- license: other library_name: scikit-learn tags: - xgboost - anomaly-detection - time-series-forecasting - research --- # DrivePulseAI KIT forecast novelty artifact This repository file is a fitted research artifact for the KIT Automotive OBD-II next-window novelty experiment. ## Source and attribution The artifact was trained on prepared windows from the KIT Automotive OBD-II Dataset by Marc Weber (2023), Karlsruhe Institute of Technology, [DOI 10.35097/1130](https://doi.org/10.35097/1130). The source dataset is licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Please cite the dataset and retain its attribution when using or redistributing this model artifact. ## Intended use and limits The artifact forecasts selected sensor medians in the immediately following 10-second window from the current window's ten sensor medians, standard deviations, and missing fractions. It contains fitted per-signal XGBoost or persistence forecasters, residual scales, feature/target order, and a 99th-percentile training calibration threshold. KIT has no verified mechanical fault labels. A high novelty score means a forecast residual was unusual relative to calibration data; it is not a fault probability, diagnosis, or validated fault-detection result. This artifact is for research and recorded-data replay demonstrations only. The artifact was serialized with joblib and must only be loaded by the matching trusted DrivePulseAI serving code. Do not load untrusted pickle/joblib files. ## Oxford RUL estimator configuration `oxford_rul_theilsen.joblib` contains the configured `TheilSenCapacityTrendRUL` estimator (1,000-cycle lookback and a four-observation minimum). It is an algorithm configuration, **not a fitted model learned from pooled Oxford measurements**. At inference time, the serving API fits a robust trend to the capacity history supplied in that request and extrapolates to the caller-supplied capacity threshold. The estimator is used with Oxford Battery Degradation Dataset 1 for experimental RUL research: DOI [10.5287/bodleian:KO2kdmYGg](https://doi.org/10.5287/bodleian:KO2kdmYGg). The source dataset's ODbL/DbCL terms and required citations are documented in [`data/DATASET_ATTRIBUTION.md`](https://github.com/direkkakkar319-ops/DrivePulseAI/blob/main/data/DATASET_ATTRIBUTION.md). The dataset itself is not included in this model repository. The 75% first-crossing target is unconfirmed, only two cells cross it, and cell-disjoint RUL performance is not established. Outputs are experimental proxy estimates, not dependable battery-life forecasts.