Fault Diagnosis of Electric Submersible Progressive Cavity Pump Systems Based on Multi-Physics Forward Modeling and Regularized Inversion

The electric submersible progressive cavity pump (ESPCP) is widely used in unconventional oil production, yet its downhole unit is subject to multiple progressive faults under coupled electromagnetic, fluid, and structural physics. Conventional monitoring methods often suffer from delayed detection and high false-alarm rates, while data-driven diagnosis requires large volumes of labeled fault samples that are rarely available. This paper reformulates fault diagnosis as a Tikhonov-regularized inverse problem, eliminating the need for historical fault data. A multi-physics forward model covering the cable, motor, coupling, pump, and tubing subsystems is constructed (fixed-point coupling of the electromechanical and hydraulic steady states, verified against the coupled transient dynamics), and 22 fault parameters are defined through explicit perturbation functions. Diagnosis is a two-stage pipeline: Stage A calibrates the healthy baseline of a 12-parameter reduced quasi-steady model from one normal-condition sweep; Stage B inverts each measured sweep with the L-curve-selected regularization and scores every fault signature direction by its restricted-fit evidence (the residual reduction of the best penalised 1-D fit along that direction, healthy projection subtracted), with a per-noise-level threshold calibrated on healthy draws. The monitoring vector comprises four surface channels (flow rate, supply current, wellhead pressure, electrical power) at 12 variable-speed-drive setpoints (m = 48). A 1,720-cell experiment (22 faults × 5 severities × 4 noise levels × 3 repetitions plus 400 healthy cells), aggregated from a single master results table, yields at 3% measurement noise a detection rate of 87.3% and a per-type classification accuracy of 41.2%, rising to 69.4% at the level of the measurably confusable fault groups, with a false-alarm rate of 17.5% at the 95th-percentile healthy threshold. The identifiability boundary of the four-channel quasi-steady measurement set is quantified (condition number 6.2e7; five weak parameter directions) and reported explicitly: the fault acting through the weakest parameter direction (bearing wear, through viscous friction B) is below the noise floor at moderate severity, as are the faults whose full-model effect is itself below the noise (backlash drag, tubing friction), and phase-attributed electrical faults collapse to their dq projection. A reduced-vs-full model discrepancy study over 135 operating-condition × fault-state combinations (27 conditions × 5 fault states) bounds the surrogate error. The method requires only normal-condition data for baseline calibration, demonstrating practical potential for field deployment. Conflict of Interest Statement The authors declare no conflicts of interest.