What question did this study set out to answer?

The aim is to improve the quality of data and reliability of models used in autonomous systems for failure prediction and object detection.

May 17, 2026Open Access

Enhancing data quality and model reliability for the design of reliable autonomous systems in prognostics and vision-based object detection

Key Points

The aim is to improve the quality of data and reliability of models used in autonomous systems for failure prediction and object detection.
Proposed methodologies for uncertainty modeling and reduction in autonomous systems
Evaluated through case studies in Remaining Useful Life estimation for batteries and engines
Developed risk-aware models for visual detection enhancements using new bounding box techniques.
PHT-based models achieved a 46% and 93% reduction in RMSE compared to baseline models in battery and engine datasets respectively.
Improvements of up to 7% in detection accuracy and 12% in object coverage were observed with the new Prob-Bbox and ABC Bbox techniques.

Abstract

Autonomous systems used for failure prediction in Prognostics and Health Management (PHM) and vision-based object detection in self-driving vehicles face numerous uncertainties. These uncertainties stem from sensor measurement errors influenced by environmental factors, operational variability related to the health of physical components, and modeling challenges in machine learning (ML) algorithms. These factors pose significant obstacles to system reliability, complicating the decision-making processes in autonomous systems. This dissertation proposes advanced methodologies for uncertainty modeling and reduction, targeting two key dimensions: data quality and model reliability. By systematically addressing uncertainty at both levels, the research aims to improve the decision-making capabilities of autonomous systems, thereby enhancing their safety, robustness, and reliability. The proposed methodologies are evaluated through extensive case studies in (i) Remaining Useful Life (RUL) estimation for batteries and engines and (ii) vision-based object detection for autonomous navigation. Performance assessments, including both quantitative and qualitative analyses, demonstrate significant improvements in uncertainty reduction and predictive accuracy.For Prognostics and Health Management (PHM), we introduce Physical Health Timesteps (PHTs), a temporal re-indexing scheme that aligns multivariate battery- and engine-degradation signals collected under diverse duty cycles. In NASA C-MAPSS and Li-ion datasets, PHT-based models cut root mean squared error (RMSE) by 46 % and 93% relative to cycle-count baselines. For vision-based object detection, we develop a Probability-based Bounding Box (Prob-Bbox) and Asymmetrically Bias-Corrected (ABC) Bbox, which jointly quantify and shrink localization uncertainty. Evaluated on BDD100K, these transform raw detection outputs into precise, risk-aware spatial models, enhancing detection accuracy (IoU) by up to 7% and increasing object coverage (region-wise Recall) by up to 12%. This innovation significantly reduces the risk of accidents involving vision-based autonomous vehicles for safer transportation.Future work will aim to further mitigate model uncertainties in challenging conditions, such as sensor degradation, distribution shifts, and rare-event scenarios, with the goal of achieving highly reliable autonomous systems in real-world deployments.

Perguntar à IA

Bookmark

View Full Paper