Technology

Physics-informed condition intelligence for marine hulls and propulsors

We combine a first-principles clean-speed model, a constrained learning layer, and uncertainty-aware tracking to convert vessel telemetry into actionable maintenance decisions.

live schematic

From measurements to maintenance evidence

physics + AI

01Estimate

Predict the clean-hull speed from independent operational telemetry.

02Compare

Convert the difference against observed performance into speed and fuel evidence.

03Decide

Trigger recommendations only when the confidence band is above the threshold.

Clean-speed model architecture

telemetry - clean-hull physics - constrained AI residual - diagnostics - action

module

Telemetry inputs

Power, shaft speed, pitch, draft, trim, wind, heading, and motion proxies.

module

Physics core

Computes the clean-hull speed prediction.

module

Constrained residual

Learns only the structured correction that physics misses.

module

Residual diagnostics

Builds speed-loss and fuel-penalty signals from the difference to observed performance.

module

Decision engine

Acts only when the conservative confidence bound exceeds a maintenance threshold.

model equation

\hat{v}_{\mathrm{clean}}(\mathbf{x}) = \hat{v}_{\mathrm{phys}}(\mathbf{x}) + \delta_{\theta}(\mathbf{x})

The clean-hull estimate is built from a physics baseline plus a residual learning term. The residual term is kept small and anchored to zero on clean nominal data, so the model remains interpretable rather than becoming a black box.

input integrity firewall

\mathbf{x} = \{\,P_{\mathrm{shaft}},\, n_{\mathrm{s}},\, \beta_{\mathrm{CPP}},\, \text{draft},\, \text{trim},\, \text{wind},\, \text{motion}\,\},\quad v_{\mathrm{GPS}} \notin \mathbf{x}

The model uses independent operational telemetry only. GPS speed is reserved for validation and residual diagnostics, so the clean-speed prediction is not built from the quantity it is trying to estimate.

PredictStep 1 of 3

Three coupled physical subsystems

The clean-speed prediction is not a single empirical rule. It comes from three coupled subsystems that explain the dominant loads and propulsive balance. Open the cards below to see the structure.

Hull resistance model

HYDRODYNAMICS

How strongly the water pushes back in calm water.

R_{\mathrm{H}} = R_{\mathrm{F}} + R_{\mathrm{W}} + R_{\mathrm{APP}}

The clean-hull baseline starts from the loads that dominate vessel speed. Instead of hiding everything inside a single empirical coefficient, the method separates resistance mechanisms so the model remains interpretable when loading, trim, and operating point change.

Inputs: delivered power, draft or displacement proxy, trim, heading, and operating state.
Output: a clean-hull resistance envelope consistent with marine propulsion physics.
Benefit: the residual is less likely to confuse normal hydrodynamic variation with genuine degradation.

engineering interpretation

Why it matters

This block contributes to the clean-hull prediction. By making the physical structure explicit, the downstream residual is more likely to represent real degradation rather than unmodelled operating variability.

Propulsor operating-point balance

PROPULSOR-HULL

How shaft power, propeller pitch, and resistance meet at one feasible speed.

T(n_{\mathrm{s}},\, \beta_{\mathrm{CPP}}) = R_{\mathrm{H}} + R_{\mathrm{E}}

The propeller and hull are treated as a coupled equilibrium. The predicted speed is the operating point where available thrust balances total resistance while remaining compatible with the measured shaft state and controllable-pitch setting.

Inputs: shaft speed, delivered shaft power, and controllable-pitch propeller setting.
Output: a feasible thrust level and a clean-hull speed estimate.
Benefit: the vessel cannot be predicted to move faster than the measured propulsive state can physically support.

engineering interpretation

Why it matters

Environmental load correction

WIND + MOTION

What weather and seaway motion add to the calm-water baseline.

R_{\mathrm{E}} = R_{\mathrm{AA}} + R_{\mathrm{AW}}

Wind and wave effects are handled explicitly so they are not falsely interpreted as fouling. Relative wind enters directly, while the effect of waves can be inferred from vessel-motion proxies when external sea-state measurements are unavailable.

Inputs: relative wind, heading, and motion-response proxies.
Output: an environment-induced resistance correction.
Benefit: the monitoring signal stays stable during poor weather and changing seaway conditions.

engineering interpretation

Why it matters

DetectStep 2 of 3

Target-independent diagnostics and condition tracking

Once the clean-hull prediction is available, the unexplained performance loss becomes the core diagnostic signal. We convert that loss into physically meaningful indicators and track them through time with uncertainty awareness.

interpretation

The objective is not merely to fit speed well. The objective is to make the remaining residual explainable: after correcting for propulsion state, loading, and weather, the remaining speed loss should be attributable to degradation of the hull or propulsor system.

Clean-speed deficit

signal

\Delta v_t = \hat{v}_{\mathrm{clean},t} - v_{\mathrm{obs},t}

A positive deficit means the vessel should have gone faster under the same power, loading, and weather. This is the direct evidence of performance loss.

Excess fuel ratio

signal

\Delta q_{f,t} = 100\,\dfrac{\dot{m}_{f,t} - \dot{m}_{f,\mathrm{ref},t}}{\dot{m}_{f,\mathrm{ref},t}}\ [\%]

The speed deficit is translated into additional fuel consumption relative to the clean-hull reference. This turns degradation into cost and emissions language.

Posterior confidence band

signal

\left[\,\underline{c}_t,\ \overline{c}_t\,\right]

The estimator tracks a latent condition state together with uncertainty. Maintenance actions are triggered from the conservative lower bound, not from a single optimistic point estimate.

ActStep 3 of 3

Uncertainty-aware maintenance logic

We act on the confidence band, not on a single point estimate. That makes the system conservative by design and prevents noisy voyages from triggering unnecessary maintenance events.

decision rule

\underline{c}_t > \tau \;\Longrightarrow\; \text{recommend maintenance}

The lower confidence bound must exceed the action threshold. In other words, the recommendation is triggered only when the evidence is strong even under uncertainty.

Action	Threshold τ	Indicative fuel impact
Watch	Alert score ≥ 1	< 1%
Alert	Speed loss ≥ 0.07 kn	~1.5%
Recommend Cleaning	Speed loss ≥ 0.28 kn or alert score = 3	~4.3%
Schedule Drydock	Speed loss ≥ 0.70 kn	~10.7%

decision pipeline

threshold τ

low evidence

strong evidence

1. Estimate clean speed

Compute the clean-hull speed from target-independent telemetry inputs.

2. Form residual signals

Derive speed-loss and fuel-penalty signals as technical and economic evidence.

3. Track condition with uncertainty

Fuse repeated evidence over time to estimate condition and confidence bounds.

4. Recommend maintenance conservatively

Act only when the conservative confidence bound clears the threshold.

Validated performance

Real-vessel results should be read together with methodology. Accuracy matters, but the value lies in combining accuracy, interpretability, and uncertainty-aware decisions.

>99%

Prediction accuracy

Validated against full-coverage operational data

<4%

Error with sparse sensors

The physics baseline holds when signals are missing

<1%

False alarms

Independent checks must agree before an alert fires

Vessels validated

Independent real-world operational datasets

Interested in the technical details?

We are happy to discuss methodology, deployment constraints, telemetry integration, and collaborative research on marine condition monitoring.