Early Failure Detection in Industrial Systems Using Physics-Informed Neural Networks (PINNs): Focused on Short-Term Prediction

Abstract

Industrial digital twins for oil and gas production systems require forecasting models that are both accurate and physically consistent, yet purely data-driven neural networks often violate underly ing dynamical constraints over extended forecast horizons. A SINDy-regularised neural forecasting framework is proposed for downhole pressure prediction in gas-lifted wells. The approach integrates sparse system identification with physics-informed deep learning. A convolutional encoder compresses multi-variate telemetry windows into compact latent representations, from which a multi-layer per ceptron decoder predicts future pressure at lead times of 30 to 120 seconds. Separately, the Sparse Identification of Nonlinear Dynamics (SINDy) discovers a parsimonious 22-term ordinary differential equation from smoothed operational data, selected via ℓ1-regularised regression followed by Ridge refitting from a 34-term observable-space candidate library incorporating polynomial, boundary-flow, and thermo-hydraulic surrogate terms. The SINDy-discovered ODE is embedded in the training procedure as a physics-based regularisa tion penalty. Random collocation points are evaluated across the forecast horizon, where the mismatch between the network-predicted temporal derivative and the SINDy ODE is penalised. Evaluation is performed on Well 00019 of the 3W dataset under normal operating conditions. The baseline config uration achieves a validation R2 of 0.9286 and a test R2 of 0.8675 at the 120-second forecast horizon, with RMSE remaining below 0.098 MPa across all four horizons. At λ = 10−7, the validation data loss decreases by 24.28% and physical consistency improves by 17.41% relative to the unregularised baseline. Varying λ reveals a narrow effective regime: at λ = 10−6, the 30 s test R2 rises to 0.9731 at the expense of long-horizon accuracy; at λ = 10−5, the test-set R2 collapses to negative values. SINDy-based physics regularisation, at an appropriately small weight, yields pressure forecasts with improved temporal consistency and long-horizon stability, without requiring a priori specification of the governing equations.