Edge-Enabled Multi-Domain Condition Assessment Of Underwater Infrastructure Using Statistical Feature Engineering

Abstract

The security of subsurface infrastructure is now more crucial as subsea data centers, marine surveillance devices, and offshore facilities are now spreading. Constant exposure to hydrostatic pressure, temperature gradient, salinity changes, mechanical vibration, and electrical load changes increase degradation rates and makes them more difficult to plan maintenance. The traditional cloud-based monitoring systems are affected by latency and communication restrictions of underwater environments.

This paper proposes an edge-independent multi-domain condition estimation system combining heterogeneous sensor fusion and statistical sliding-window feature engineering with feature inference based on lightweight machine learning. Operational and environmental cues, such as temperature, pressure, conductivity, turbidity, vibration, and acoustic activity, and electrical current were modeled to give the feel of underwater infrastructure behavior. A flow-through window transformation was created to extract 49 descriptors of a window to assess both the distributional and temporal features.

Two anomaly detecting plans were tested; an unsupervised Isolation Forest baseline and a supervised Random Forest classifier. The supervised ensemble recognized the targets with an almost perfect discrimination (AUC = 1.00) in the test conditions with a controlled simulation at the expense of the baseline with a weak detection ability (accuracy ≈ 0.05). Real-time Average inference latency (mean 0.94 ms per sample) on a edge-deployed Raspberry Pi showed sub-milliseconds average, indicating that it is possible to run processes in real time.

The framework suggested in this paper creates a scalable strategy of using edge-intelligent underwater condition monitoring with a small amount of computation.