⚡ electrical engineering

A corrosion-aware digital-twin framework for condition monitoring of saline-water centrifugal pumps

This paper presents a reproducible, corrosion-aware digital-twin framework for monitoring saline-water centrifugal pumps, which integrates experimental data from a 200-hour degradation study with multidomain state representations and CFD modeling to support fleet management without relying on fully validated predictive algorithms.

Original authors: Nsini Udo

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Nsini Udo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, salt-heavy world of the ocean, metal pumps work tirelessly to move water for cooling systems, fire protection, and ballast. These machines are the unsung heroes of marine engineering, yet they face a constant, invisible enemy: corrosion. When saltwater eats away at the smooth metal surfaces inside a pump, it creates tiny pits and rough patches. This degradation is not just a cosmetic issue; it changes how the water flows, making the pump work harder, vibrate more, and lose its ability to push water efficiently. For decades, engineers have tried to predict when these pumps will fail, often relying on broad guesses about their health. However, a new approach is emerging that seeks to understand the machine's condition with much greater precision, treating the pump not as a single black box, but as a complex system where surface texture, water flow, and mechanical vibration tell a connected story.

This story is told through a concept known as a "digital twin." Imagine a virtual copy of a physical pump that lives inside a computer. This copy is fed real-time data from sensors on the actual machine, allowing engineers to see how the pump is performing without having to take it apart. The challenge has always been how to translate the messy reality of a corroding metal surface into numbers that a computer model can understand. A recent study by independent researcher Nsini Udo tackles this specific problem. Rather than claiming to have built a perfect, working virtual pump that predicts failure dates, the study offers a clear, step-by-step framework for how to build one. It focuses on a specific type of pump made of stainless steel, running in synthetic seawater, and traces its slow decline over two hundred hours.

The researchers began by looking at data from a controlled experiment where a stainless steel pump ran in salty water for two hundred hours. They checked the pump's performance every fifty hours. Over this time, the pump did not break suddenly; instead, it slowly got worse. The amount of water it could push per second dropped from 1.50 liters to 1.38 liters. The height the water could be pushed to, known as the "head," fell from 12.0 meters to 10.7 meters. The electrical efficiency, which measures how well the pump converts electricity into moving water, slipped from 70.6 percent down to 58.0 percent. At the same time, the machine began to shake more, with vibration levels rising from 0.20 millimeters per second to 0.75 millimeters per second. Crucially, the researchers also measured the texture of the metal inside the pump. The surface, which started out relatively smooth with an average roughness of 0.8 micrometers, became significantly rougher, reaching 3.4 micrometers by the end of the test.

The core of this work is how these different changes are organized into a single, understandable picture. Instead of combining all these numbers into one simple "health score," which might hide what is actually wrong, the study proposes keeping them separate. It creates a detailed list of the pump's state, noting the flow, the pressure, the vibration, and the surface texture as distinct pieces of information. This is vital because a drop in water pressure could be caused by a clogged pipe, a worn seal, or a rough surface, and each cause requires a different fix. By keeping the data separate, the system can tell the difference between a mechanical shake and a surface problem, even if they happen at the same time.

A significant part of the study addresses a tricky gap between what can be measured and what a computer needs to know. Engineers can easily measure the roughness of a metal surface with a tool, but computer models that simulate water flow need a different kind of number to represent that roughness. The study does not pretend to have found the perfect formula to convert one number into the other. Instead, it openly admits the uncertainty. It suggests a range of possibilities, showing that if the measured roughness is 3.4 micrometers, the number used in the computer model could reasonably be anywhere between 1.7 and 6.8 micrometers, depending on the specific shape of the corrosion pits. This transparency is a key finding: it shows that we cannot simply assume a single conversion rule works for every pump. The study explicitly rules out the idea that a single, universal number can describe all corrosion, and it refuses to claim that a computer model has been fully validated when it has not yet been tested against real-world data.

The researchers also emphasize that this work is a blueprint, not a finished product. They have not built the final computer model that predicts when the pump will break, nor have they trained an artificial intelligence to guess the remaining life of the machine. The dataset used, consisting of only five measurement points over two hundred hours, is too small to train such advanced systems. Instead, the study provides the necessary structure to make those future steps possible. It defines how to take a physical measurement, how to handle the uncertainty when turning that measurement into a computer input, and how to compare the computer's prediction with the real-world sensor data. This structure includes a way to track the "residual," which is simply the difference between what the computer thinks is happening and what the sensors say is actually happening. If this difference grows, it signals that something is wrong, but the framework allows engineers to see exactly which part of the system is causing the mismatch.

This approach offers a new way to think about maintenance for ships and offshore platforms. Currently, maintenance often relies on fixed schedules or reacting to obvious failures. This framework suggests a more intelligent path where sensors watch the pump's vibration and flow, while periodic inspections check the surface texture. If the sensors show a change, the system can suggest an inspection rather than immediately ordering a part replacement. It allows for a staged approach: first, understanding the physical changes; second, building a reliable virtual model; and finally, using that model to predict future problems. The study makes it clear that until the virtual model is tested against independent data and proven to work in real marine environments, it remains a theoretical tool.

The findings serve as a reminder that in complex engineering, the path to a solution is often about clarity rather than complexity. By separating the different types of data and being honest about what is known and what is assumed, the study creates a solid foundation for the future. It shows that while we cannot yet predict the exact moment a pump will fail, we can build a system that understands the slow, steady march of corrosion with much greater precision. The work does not claim to have solved the problem of pump maintenance, but it has provided the essential map for how to solve it, ensuring that every step from the physical pump to the digital model is traceable, understandable, and grounded in real evidence.

Technical Summary: A Corrosion-Aware Digital-Twin Framework for Saline-Water Centrifugal Pumps

Problem Statement
Marine centrifugal pumps operate in harsh environments where corrosion-induced surface degradation and mechanical condition evolution occur simultaneously. Current condition-monitoring paradigms often fail to capture the coexistence of these distinct degradation modes. Specifically, there is a lack of evidence-based frameworks that clearly separate measured physical evidence from derived quantities, model-form assumptions, and virtual-model outputs. Furthermore, the relationship between measured arithmetic surface roughness ( $R_a$ ) and the equivalent sand-grain roughness ( $k_s$ ) required for Computational Fluid Dynamics (CFD) modeling is often treated as a universally calibrated constant, despite the uncertainty inherent in corroded 316L stainless-steel surfaces. The paper addresses the need for a transparent, reproducible architecture that links corrosion measurements to a digital twin without making unsupported claims regarding failure probabilities or Remaining Useful Life (RUL).

Methodology
This study is a secondary analysis of a previously published 200-hour controlled saline-water experiment involving a 316L stainless-steel centrifugal pump operating in 30 parts per thousand (ppt) synthetic seawater at approximately 2900 rpm. The methodology does not involve new physical experiments or the execution of CFD simulations. Instead, it focuses on reorganizing existing data into a multidomain state representation and defining the interfaces for a future digital twin.

Key methodological steps include:

Data Provenance and State Vector Formulation: The study reorganizes experimental data (flow, head, efficiency, vibration, and surface roughness) into a heterogeneous state vector $z(t)$ . This vector explicitly distinguishes between measured variables (e.g., $R_a$ , vibration), derived quantities (e.g., efficiency), and assumed parameters (e.g., the transfer multiplier $\alpha$ for roughness).
Roughness Transfer Sensitivity: To address the gap between measured $R_a$ and CFD-required $k_s$ , the paper proposes a deterministic sensitivity envelope defined by the relation $k_s = \alpha R_a$ . Three transfer multipliers are tested: $\alpha = 0.5, 1.0,$ and $2.0$. This creates a screening envelope for $k_s$ rather than a single calibrated value.
Residual-Based Interface: A conceptual interface is defined where the residual $r$ is the difference between observed outputs ( $y_{obs}$ ) and future virtual model outputs ( $\hat{y}$ ). The framework emphasizes that residuals must be auditable, containing metadata on operating context, model versions, and uncertainty tags, rather than collapsing into a single health score.
Exclusion of Validation: The author explicitly states that no CFD solver, live data feed, or validated model-measurement residual was generated in this work. The study stops at defining the architecture and evidence chain.

Key Results
The secondary analysis of the 200-hour dataset reveals the following degradation trajectories:

Hydraulic Performance: Pump head decreased from 12.0 m to 10.7 m (−10.83%), and flow rate decreased from 1.50 L/s to 1.38 L/s (−8.00%).
Efficiency: Electrical-input efficiency dropped from 70.6% to 58.0% (−17.85%).
Mechanical Response: Vibration velocity increased from 0.20 mm/s to 0.75 mm/s (+275%).
Surface Degradation: Arithmetic surface roughness ( $R_a$ ) increased from 0.8 µm to 3.4 µm (+325%). Pit density increased from 0 to 68 pits/mm², with a maximum pit depth of 210 µm.

Based on the final $R_a$ of 3.4 µm, the proposed sensitivity envelope yields equivalent sand-grain roughness values ( $k_s$ ) of 1.7, 3.4, and 6.8 µm for $\alpha = 0.5, 1.0,$ and $2.0$, respectively. This contrasts with literature benchmarks (e.g., Adams et al., 2012) which suggest $k_s \approx 19.9$ µm for the same $R_a$ , highlighting the significant uncertainty in roughness transfer for corroded surfaces.

Key Contributions
The paper contributes a methodical framework rather than a validated operational tool. Its primary contributions are:

Corrosion-Aware State Vector: A multidomain state representation that separates hydraulic, mechanical, and surface degradation channels to support explainable monitoring.
Explicit Evidence Architecture: A clear distinction between measured evidence, derived quantities, transfer assumptions (the $\alpha$ multiplier), and virtual-model outputs.
Deterministic Sensitivity Envelope: A model-form sensitivity analysis for the $R_a$ -to- $k_s$ transfer, demonstrating that CFD inputs should not rely on a single literature coefficient without validation.
Residual Interface Specification: A definition for how residuals should be structured to include operating context and uncertainty, enabling maintenance staff to distinguish between consistent anomalies and environmental variations.
Staged Validation Pathway: A proposed roadmap moving from surface characterization to matched-hydraulic validation and eventually to data-driven prognostics, acknowledging that current five-point data is insufficient for machine learning models.

Significance and Claims
The author positions this work as a "proof-of-concept" for a corrosion-aware digital-twin architecture. The significance lies in the transparency of the evidence chain: the framework prevents the confusion of assumed CFD inputs with measured properties. The paper explicitly claims not to have produced a validated digital twin, a failure probability model, an RUL estimator, or a machine-learning classifier. Instead, it provides the necessary "pre-implementation" evidence, state representation, and interface specifications required to enable and validate fleet management in the future. The work serves as a methodological bridge, ensuring that future implementations of intelligent marine systems can trace their decisions back to specific, uncertainty-aware measurements rather than opaque black-box models.

Technical Summary: A Corrosion-Aware Digital-Twin Framework for Saline-Water Centrifugal Pumps

More like this