Real-Time Mechanistic Anomaly Detection in Drilling: Bayesian Fusion of Reduced-Order Multiphase Flow and Surface Data

Abstract

Drilling operations routinely rely on surface instrumentation to infer evolving downhole conditions, yet the mapping from surface signals to annular multiphase states remains underdetermined when flow transitions and sensor artifacts co-occur. In managed-pressure and deepwater contexts, early recognition of anomalous influx signatures is particularly challenging because small deviations in pressure, flow, and pump behavior can be consistent with multiple latent mechanisms. This paper develops a physics-constrained digital-twin framework that fuses reduced-order multiphase flow dynamics with uncertainty-aware representation learning to detect, localize, and quantify anomalies in real time using primarily surface measurements. The central contribution is a differentiable observer that embeds a quasi-one-dimensional annular flow model inside a probabilistic state-space estimator while learning only the closure discrepancies needed to reconcile model and data. Unlike purely discriminative detectors, the proposed approach outputs calibrated posterior distributions over latent annular states, regime-consistent transport parameters, and an anomaly score tied to mechanistic residuals. The method supports both abrupt events and slowly drifting conditions through Bayesian change detection on innovation statistics, and it provides decision support by mapping posterior risk to constraint-aware operational advisories. The paper details identifiability conditions, stabilization via dissipativity constraints, and robustification to sensor delays and rate limits. A comprehensive evaluation protocol is proposed to quantify detection latency, false-alarm control, and out-of-distribution generalization across geometries and fluids without assuming access to downhole labels.}