Sensor Drift Estimation for Reactor Systems by Fusing Multiple Sensor Measurements

Rao, Nageswara S. V.; Greulich, Christopher; Ramuhalli, Pradeep; Çetiner, Sacit M.; Devineni, Pravallika

doi:10.1109/nss/mic42101.2019.9059694

Cited by 2 publications

(6 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…OLM for sensor and instrument calibration has been extensively studied [4,5,6,7,8]. Despite a positive regulatory safety evaluation [ 9 ], OLM for extending calibration intervals has not been adopted by US industry due to persistent questions related to the uncertainty bounds in OLM, as well as the need to demonstrate applicability to all anticipated sensor fault conditions and operating scenarios (steady state as well as transient).…”

Section: Online Monitoring -Prior Workmentioning

confidence: 99%

“…Transient sensor output predictions have been explored recently [14], though the approach relies heavily on physics-based models of the process and may not be readily applicable to real-time computation. More recent analyses have determined that the problem of drift detection and fault in general is learnable (i.e., solvable by machine learning techniques); generalized bounds on the model prediction have also been defined for the specific case of support-vector machines (SVM) and ensemble of trees (EOT)-based models [4].…”

Section: Online Monitoring -Prior Workmentioning

confidence: 99%

“…However, previous research has shown that such an approach results in significant cross-sensitivity between sensors, wherein drift in one sensor can influence the residual error from multiple sensors, challenging the ability to identify the specific sensor that is drifting. To robustly detect drift in a single sensor and identify the drifting sensor reliably, techniques were proposed that perform the predictions and assessments in a latent space or used models that predict sensor measurements using all other sensors as input [4,13].…”

Section: Time Series Predictionmentioning

confidence: 99%

“…Previous research also examined model prediction accuracy [4] and confidence bounds quantification [12]. Theoretical prediction accuracy bounds were developed for a number of model forms (such as SVM and EOT) using statistical learning theory [19].…”

Section: Time Series Predictionmentioning

confidence: 99%

See 3 more Smart Citations

Methodology for Online Sensor Recalibration

Ramuhalli¹,

Muhlheim²,

Yigitoglu³

et al. 2022

PHM_CONF

View full text Add to dashboard Cite

A significant contributor to nuclear power plant operations and maintenance (O&M) costs is the periodic calibration check of sensors. Periodic calibration checks provide the necessary confidence that the measurements from these sensors are correct, with the data used to monitor and verify proper operation of the reactor. The periodicity of calibrations in the nuclear industry can range from once in several weeks for some instrument channels to once every refueling outage (~18 months) for certain safety-significant pressure and level transmitters. With expected longer refueling intervals in many advanced reactor/small modular reactor concepts, fewer opportunities are expected to be available for manual calibration checks and recalibration for many instrument channels. Presently, periodic sensor calibration checks are performed manually, with manual recalibration performed if the sensor is found to be out of calibration. Although studies have shown that most (over 90%) sensors are found to stay within calibration specifications over a calibration cycle (~18 months), labor must still be spent to verify that these sensors are within calibration. Given the large number of sensors (anywhere between 100–2400 sensors) in a typical nuclear power plant, the ability to identify sensors that are failing/failed or drifting out of calibration and limit recalibration to those specific sensors has the potential to save between $0.5–1M per year per plant. The cost of calibration checks and recalibration are expected to be of greater significance to advanced reactors and small modular reactors, given the industry focus on reducing O&M costs for these reactor concepts as part of improving the economic viability of advanced nuclear power. Methods that can compensate for calibration drift by adjusting the calibration automatically and in real-time may be able to further reduce costs associated with recalibration. Such autocalibrations or online automated recalibrations reduce unavailability of instrumentation and improve online maintenance planning flexibility from both resource allocation and online risk evaluation perspectives. This paper describes an initial set of algorithms developed for the purpose of detecting drift and correcting for it through an online recalibration method. Initial results on laboratory-scale experimental data indicate the potential of these algorithms to detect calibration drift and update calibrations with prediction and drift correction accuracy exceeding 95%.

show abstract

Section: Online Monitoring -Prior Workmentioning

confidence: 99%

Section: Online Monitoring -Prior Workmentioning

confidence: 99%

Section: Time Series Predictionmentioning

confidence: 99%

Section: Time Series Predictionmentioning

confidence: 99%

See 2 more Smart Citations

Methodology for Online Sensor Recalibration

Ramuhalli¹,

Muhlheim²,

Yigitoglu³

et al. 2022

PHM_CONF

View full text Add to dashboard Cite

show abstract

“…However, good accuracy (for instance, low RMSE) on a training or test data set is not necessarily indicative of good generalization performance (prediction performance on data not previously seen by the model). While such generalization performance is difficult to quantify, generalization error bounds may be computed in some cases [37,38].…”

Section: Evaluation Metricsmentioning

confidence: 99%