Virtual Sensing

Cutter Tooth Wear Assessment in Turbid Water

Estimating cutter-tooth abrasion where visual inspection is impossible using PINNs to solve inverse problems from motor signals.

Read Time: 8 min read | Published: 2025-01-15

The Challenge

In dredging, visual inspection of cutter heads is impossible during operation. Wear progression directly correlates with a 12-18% increase in energy consumption per 100 operating hours (Mechanical Sciences, 2025). As teeth wear, the cutting force to normal force ratio drops dramatically from 33:1 to 5:1, yet current practice still relies on inefficient manual inspection stops.

Our Approach

We utilize a dual-network inverse PINN architecture (arXiv:2512.12074) where both the solution (cutting behavior) and the unknown parameter (tooth sharpness) are learned simultaneously from the same loss function. This physics-based approach uses the established relationship between tooth geometry and mechanical resistance to infer wear directly from motor torque and pump pressure.

Validation

While direct validation in turbid water is challenging, physics-guided deep learning models in analogous high-speed milling applications have achieved ~3.9% MAPE in wear estimation. By incorporating the specific 33:1 to 5:1 force ratio change as a physical constraint, our model can mechanistically distinguish between harder soil strata and worn teeth.

Recommendation

Proceed with a gated pilot program. Do not rely on this as the sole indicator for maintenance until a wider library of soil-type training data is established.

Related Articles

Slurry Transport Instability & Cavitation Modeling

Dual-Model Architecture combining PINNs for density sensing with Surrogate Models for real-time CFD proxies to handle instability....

Pipelay Tensioner Slip Prediction

Physics-Informed Friction Models using existing pressure/speed data to predict slip risk with a 30-minute horizon....

Data Scarcity in Maritime Predictive Maintenance

Critical offshore assets do not fail often enough to feed classical machine learning, so most pure data driven predictive maintenance initiatives stal...