NigerStress-PdM: An Open, Reproducible Stress-Test Benchmark for Resilient Predictive Maintenance of Energy Infrastructure Where Field Data Are Scarce

Public predictive-maintenance (PdM) benchmarks are recorded under controlled conditions that differ sharply from those of industrial assets in West Africa, where high humidity, saline exposure, unstable grid power, telemetry loss, mixed legacy and modern instrumentation, and pipeline vandalism routinely change what sensors record. This lab-to-field gap makes it difficult to evaluate transfer learning and domain adaptation for local deployment, because the data that would be needed are rarely shared. We present NigerStress-PdM, a documented and reproducible protocol that converts four open benchmarks (NASA CMAPSS, UCI AI4I 2020, the Petrobras 3W offshore-well dataset and ESPset for electrical submersible pumps) into stressed target domains using six physics-informed perturbations: humidity-induced sensor drift, intermittent telemetry loss, sabotage-type impulses, progressive corrosion drift, sampling-regime mismatch and wave-induced cyclic loading. Each perturbation is parameterised from documented operating conditions and applied only to the asset classes it can physically affect. The protocol yields a 43,000-instance corpus (19,000 clean source and 24,000 stressed target instances). We validate the benchmark by showing that the perturbations create measurable, channel-specific distribution shift (composite maximum mean discrepancy 0.82; per-perturbation range 0.59 to 0.84), that a domain classifier separates source from target with 91.4% accuracy and a latent-space silhouette of 0.74, and that the added stressors carry genuine predictive signal (humidity-drift residual and telemetry-gap count rank 8th and 11th among SHAP features). The protocol, parameters and code are provided to support reproducible evaluation of domain adaptation for developing-economy industrial settings.