Kalman Filter-based Digital Twin Framework for Anomaly Detection and System Identification of a Pipe Conveying Fluid
Bibliographic record
Abstract
RÉSUMÉ: Au Québec, 94 % de l’énergie provient de l’hydroélectricité. Or, ce mode de production est confronté à des enjeux croissants : imprévisibilité climatique, acceptabilité sociale limitée pour de nouveaux barrages et vieillissement des turbines. L’intégration de sources énergétiques intermittentes pour varier les modes de production énergétique induit un changement d’opération des turbines hydrauliques. Initialement conçues pour un régime permanent, ces dernières sont soumises à des cycles de charge partielle et des séquences de démarrage/arrêt plus fréquentes. Ces conditions additionnées au vieillissement des machines accentuent les risques de bris causant l’arrêt de production énergétique des turbines. Les stratégies de maintenance périodique actuelles ne permettent pas d’identifier en amont les sources de défaillance. Ainsi, prédire ces défaillances dans les turbines permettrait d’optimiser leur maintenance et de réduire les charges monétaires associées à la perte de la machine et l’arrêt de production énergétique. L’objectif général de ce projet est de développer un cadre d’applications et d’algorithmes pour la construction d’un jumeau numérique des turbines hydrauliques. À partir des données des capteurs, et des formulations mathématiques décrivant la dynamique du système, le jumeau numérique vise à reconstruire une copie digitale de la turbine. Conjointement avec les algorithmes développés, le but est de prédire les sources de défaillances dans le système via l’analyse vibratoire de la machine, l’identification des paramètres mécaniques inconnus et réaliser le suivi tout au long de la vie utile de la machine en temps réel. Construire un jumeau numérique d’une turbine s’avère fastidieux et coûteux d’où l’utilisation du systèmes académiques du tuyau parcouru par un écoulement. Outre sa simplicité, ce tuyau imite les interactions entre l’eau et la structure rencontrées dans une turbine comme l’amortissement ajouté de l’eau, le flottement de la structure, et l’excitation induite par l’écoulement. Aussi, ce montage expérimental permet d’étudier une large gamme dynamique, soit du linéaire au chaos, en variant la vitesse de l’écoulement. Ce modèle largement étudié dans les 70 dernières années a un cadre mathématique décrivant cette dynamique vérifié et validé. Ainsi, l’hypothèse du projet stipule qu’il est possible de transférer le savoir acquis aux turbines hydrauliques en testant notre cadre de jumeau numérique sur le tuyau. Le montage consiste en un tuyau en caoutchouc siliconé et deux caméras évènementielles enregistrant la variation d’intensité lumineuse des pixels du capteur. Ces capteurs produisent un nuage continu de coordonnées des pixels excités dans le temps provenant du mouvement du tuyau. ABSTRACT: Ageing of hydro turbines and the addition of intermittent energy sources to the grid imply partial load operation of the turbines under stringent conditions. The digital twin (DT) as a predictive maintenance tool could be a solution to optimise maintenance, reduce power downtime, and related economic losses. However, building a functional DT of a hydro turbine from scratch is tedious and expensive. Hence, to study the fluid-structure interaction phenomena in hydro turbines, the main objective of this project is to build a digital twin framework of a pipe conveying fluid system instead. The nonlinear pipe oscillation is captured with two event-based cameras and reconstruct in 3D through a parallax and polynomial fitting algorithm. Then, an unscented Kalman Filter (UKF) with the nonlinear formulation of the pipe is used to solve the inverse problem and estimate the unknown flow rate inside the pipe. To accelerate the monitoring procedure, we run a linear Kalman filter (LKF) combined with dynamic mode decomposition (DMD) to track the modal changes in the event-based data reconstruction. With the latter code, anomaly detection is performed by comparing the LKF estimates with the known modal patterns of the DMD using the Euclidean distance inside a steepest ascent hill-climbing algorithm and an entropy-based algorithm on the Euclidean distance evolution. The whole is merged inside a workable framework that performs in our lab the flow rate identification and anomaly detection. In this work, with event-based outputs, we reconstructed the pipe oscillation with an average error of 2.3 % in which a maximal error of 6 % was observed once the tip of the pipe reached a large amplitude. To ensure that the polynomial fitting has sufficient event coordinates, we find that the optimal time window should be set to Δt = 0.025. The parameter identification with UKF is dependent on the accuracy of the prior knowledge formulation. For simulated data, the UKF is able to estimate the deflection with an average error of 0.47 % and the unknown flow rate with an error of 0.2 %. The validation of the system identification with UKF is done by using the event-based data and a flow metre. The UKF estimates increase to 0.61 % for the deflection and 6 % for the flow rate. LKF-DMD manage to accelerate the process from 12 s per iteration to 10−3 s per iteration. We find that the entropy method catches well anomalies representing slow modal changes such as machine efficiency decrease. The Euclidean method alone detected instantaneously sudden perturbation peaks such as sensor defects. Overall, this work shows that the proposed DT framework is still in its infancy. Hence, building a mature DT framework will require, among others, the interoperability of multiple algorithms, hybrid modelling, high computational power, and efficient machine learning technique to generalise and automate the process.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".