Summary

• Keen AI and National Grid Electricity Transmission are developing a shared foundational model for electricity network asset management.
• FoSMo is designed to standardise how network operators collect and analyse visual data from assets such as pylons, cables, insulators, vegetation, and corrosion.
• The project addresses duplicated AI development across operators and could create a reusable base for grid inspection and maintenance tools.

Keen AI and National Grid are developing a shared AI model for electricity network asset management, aiming to reduce duplicated work across transmission and distribution operators while improving inspection and maintenance capability.

The project, known as FoSMo, or Foundational Shared Model, is backed by Ofgem’s Strategic Innovation Fund. Keen AI says the model is intended to act as a reusable base for computer vision tasks across Britain’s electricity networks, helping operators analyse visual data from assets such as pylons, cables, insulators, vegetation, and corrosion.

The problem is practical. Network operators often develop separate computer vision models to detect the same asset types and defects across similar infrastructure. That creates duplicated development, fragmented datasets, inconsistent performance, and higher costs that can ultimately flow through to consumers.

FoSMo is designed to pool learning across the sector. Keen AI says the model will provide a pre trained vision base that operators can fine tune for specific use cases, including insulator defects, vegetation management, and corrosion monitoring. It is also intended to be permissively licensed, allowing organisations to build on top of the shared model rather than starting from scratch.

Ofgem’s assessment, quoted by Keen AI, describes the project as innovative, novel, and risky, noting that it is the first attempt to develop a shared machine learning model across all UK networks. The regulator highlighted technical, regulatory, and commercial challenges around data integration, cross industry collaboration, and governance for shared model use.

Those challenges explain why the project is a useful test of infrastructure AI. Models rarely fail because they cannot identify an object in a clean test set. They struggle because asset data is inconsistent, inspections happen in different conditions, operators use different systems, and governance becomes complicated when several organisations contribute data to a shared tool.

The project also reflects a wider shift in energy network digitalisation. The UK electricity system is being asked to support renewable generation, electrified transport, heat pumps, industrial electrification, new transmission infrastructure, and growing demand from datacentres. Operators will need to inspect, maintain, and reinforce assets more efficiently while the network becomes more heavily loaded.

Computer vision can help by reducing manual review, identifying defects earlier, and prioritising maintenance. Drones, helicopters, mobile crews, and fixed sensors already generate large volumes of imagery. Value comes from turning that imagery into reliable operational intelligence.

A shared model could improve performance by learning from wider datasets than any single operator can collect alone. It could also support smaller suppliers and new entrants building inspection, analytics, and asset management tools for the electricity sector. Instead of each vendor training a narrow model against limited data, companies could build services on top of a common foundation.

Governance will decide whether that potential survives real deployment. Operators will need agreements on data quality, access rights, model updates, liability, security, and how improvements are shared. They will also need confidence that a shared model does not expose commercially sensitive information or create new dependencies without clear accountability.

The Alpha phase will involve data integration and curation, model architecture design, initial training, and governance frameworks. FoSMo is not a headline grabbing AI application, but it targets a genuine operational burden: inspecting and maintaining physical assets across a grid that has to expand and decarbonise at the same time.