The reliable prediction of future system behavior using sensor data is often hindered by inherent uncertainties, especially in cases where the data undergo gradual changes over time. These uncertainties typically arise from environmental factors or system degradation, posing significant challenges to accurate prognosis and decision-making. In this study, we propose a solution to address this issue by employing confidence intervals to quantify uncertainty in prognosis based on progressively drifted sensor data. Our approach aims to establish a robust framework for evaluating the uncertainty associated with predictions derived from sensor data affected by gradual changes. To illustrate the importance of our proposed method, we mathematically model an exponentially growing sinusoidal pattern with additive noise and outliers, a pattern commonly observed in vibration signals from rotating machinery. Through various deep learning models, well-trained and optimized under hyperparameter optimizations and validation, our empirical validation and analysis demonstrate the effectiveness of our approach in enhancing the reliability and accuracy of prognosis models in dynamic sensor data environments. Thus, we draw important conclusions about the trustworthiness of predictions. This research contributes to advancing the understanding and application of statistical techniques in managing uncertainty within sensor-based prognostic systems, thereby improving their effectiveness across diverse real-world applications.
Previous Article in event
A Wearable Reflectance PPG Optical Sensor Enabling Contact Pressure and Skin Temperature Measurement
Previous Article in session
Next Article in event
Confidence Intervals for Uncertainty Quantification in Sensor Data-Driven Prognosis
Published:
26 November 2024
by MDPI
in 11th International Electronic Conference on Sensors and Applications
session Sensors and Artificial Intelligence
https://doi.org/10.3390/ecsa-11-20501
(registering DOI)
Abstract:
Keywords: Confidence intervals; Decision-making; Deep learning; Prognosis; Sensor data; Uncertainty; Vibration signals