VISTA Sensor Naming
M.Sc. Thesis · ISO 19848 · Applied ML
Overview
Ships have thousands of sensors. Each sensor has a name. None of them are named the same way. ISO 19848 exists to change that, and DNV built VISTA on top of it. I spent over a year working on VISTA as a Data Engineer at DNV, and somewhere in the middle of that I wrote my M.Sc. thesis on the classification problem underneath it: a systematic comparison of ML methods for predicting correct sensor tag names across the VIS hierarchy, identifying which approaches hold up at scale and which ones fall apart.
The challenge
ISO 19848 does not tell you which naming convention to use. It says here is a structure, pick one. Two main conventions exist in practice: JSMEA, which is service-oriented and quick to roll out, and DNV's VIS, which is semantically rich but comes with 2,883 functions, 773 product types, and 19 engineering disciplines. For a classifier that is a very large label space, very little training data, a vocabulary that shifts between shipbuilders, and zero tolerance for being wrong because the data downstream is safety-critical.
The approach
- 01 Spent serious time with both JSMEA and DNV VIS: how they are structured, where they differ, what that means for an ML system trying to map legacy tag names to a standard.
- 02 Ran a structured comparison across applicable ML methods: supervised classification, NLP sequence labelling, semantic embedding similarity, and anomaly detection for outlier tags. None are a clean fit — part of the value is documenting exactly why.
- 03 Evaluated each approach against two independent sub-tasks: Primary and Secondary Item prediction, and Tag Element prediction. They behave differently enough that no single method wins across the board.
- 04 Can not share the data or results. NDA. But the methodology and the comparison are fully in the thesis.
Outcome
Completed thesis. Not many people have written research that sits exactly here: ISO 19848, applied NLP, maritime domain, and real deployment constraints all at once. It is a niche corner but it is a real one, and the problems in it are genuinely hard.
Working on something in the maritime software or applied AI space? Let's talk →
Next project