07 April 2025

Innovation through collaboration: the Smart Image Database project

Elia Group and its partners are creating a comprehensive image database to train AI models for fault recognition. While individual collections lacked sufficient images of assets (like pylons), pooling these images into a single data repository has enabled the training of AI models with over 90% accuracy.

Context

Elia Group is a founding member of the Cross -Industry Innovators ecosystem – selected companies that have come together to explore and implement technological innovations in asset management and operations. The ecosystem addresses shared challenges by implementing joint Proof-of-Concept projects, exemplified by the award-winning Smart Image Database (SID) project.

For more information, visit Cross Industry Innovators.

Traditionally, monitoring the state of the grid and assets was done manually. Today, drones and robots are increasingly used to collect images for analysis. Within the Smart Image Database (SID) project, which has won the Maintainer Award in 2024, shared images from these drones and robots are utilized to train AI models for rapid fault detection.

Approach

The innovation mediator Infront from the ecosystem acted as project coordinator in the Smart Image Database project. After multiple workshops in 2023 it was decided that a new and promising startup would be used to train the AI models. Three project phases were devised.

In the first phase the focus was on collecting as many good images as possible of isolator chains. A common taxonomy of assets was made for labelling, and the first images were labelled, and some basic AI models were trained.
In the second phase a lot more images were labelled, more partners joined and the AI models were perfected.
In the third and final phase, in 2025, synthetic data will be tested to see its added benefit in closing some development gaps and training for new assets.

Find out more on how to join our cross-industry ecosystem or the SID project through the following video:

Results

The initial dataset for the SID project has grown significantly, now containing 12,000 inspection images. The curated dataset includes 3,000 labeled images representing 280,000 samples or labels, allowing for more comprehensive training of AI models. As a result, the number of detection and classification models has increased from 17 to 25 detection models and 13 classification models, achieving an accuracy of over 90%. The EL.AI team at Elia Group is currently deploying these models.

In 2025, the SID project will be launched on Kubernetes, enhancing scalability and security. The models will be tested by the LA&S AM team at ETB and OGZ at 50Hertz using representative images, with qualitative and quantitative analyses to be done in the beginning of 2025. This collaboration demonstrates the strong commitment within Elia Group, with business teams providing expertise and the digital department operationalizing products (EL.AI, IRIS).

Next steps

The SID project will continue throughout 2025, with plans to create a new version of the dataset that includes more assets identified by AI models. The exploration of synthetic data—AI-generated images of damages—will further optimize model performance. Although the SID project is nearing its final phase, it has already yielded significant results, prompting the consortium to consider how to sustain these outcomes beyond the project's conclusion.