From Panel Image to Wiring Compliance: Building Explainable AI for Utility Verification

Frequently Asked Questions

Explainable AI Framework for Utility Verification

What challenge does the explainable AI framework address in utility verification?

The framework tackles the difficulty of verifying that field installations in energy utilities match approved engineering designs. Traditional compliance verification is often slow, manual, and prone to human error. By automating the comparison between field images and schematics, the framework improves accuracy and efficiency in grid safety and reliability checks. Source

How does the graph-based approach work for comparing installations to schematics?

The framework represents both the approved schematic and the completed installation as attributed graphs built on a shared ontology. It creates structured representations of components and their connections, enabling direct comparison between what was designed and what was built. This method allows for nuanced detection of discrepancies beyond simple image recognition. Source

What are the two layers of compliance evaluation in the framework?

The framework evaluates compliance in two layers: First, deterministic rules check universal installation requirements consistent across projects. Second, a severity-weighted graph edit distance compares the field installation to its specific approved design, generating a compliance score and a detailed list of potential discrepancies. Each finding identifies the affected component, its severity, and the governing standard. Source

How does the framework ensure explainability and trust for engineers?

Instead of returning a black-box score, the framework provides detailed findings that show what may be wrong, where the issue appears, and why it matters. Every discrepancy is linked to the affected component and the relevant standard, enabling engineers and field teams to evaluate findings, apply their expertise, and determine next steps. Source

What types of infrastructure and environments has the framework been demonstrated on?

The framework was demonstrated using medium-voltage underground switching cabinets, evaluating installations against industry and manufacturer specifications. Schematic graphs were automatically extracted from production CAD drawings, and a computer vision system provided structured data from field photographs. All development followed NDA-governed data-handling requirements and remained on local infrastructure, ensuring sensitive utility data was not exposed to cloud services. Source

Can this explainable AI framework be applied to other regulated industries?

Yes, while the initial work focuses on energy infrastructure, the framework can be adapted to any environment where teams must compare a completed physical system with an approved design governed by structural rules. This includes construction, commissioning, or maintenance in regulated industries, helping reduce costly rework and prevent equipment failures or safety incidents. Source

Where can I learn more about the project and its technical details?

You can explore further technical details and project information at https://wyrevern.netlify.app/ and by reading the full article at Data Society's website. The framework was presented at the 4th Data Science & AI Summit during London Data Science Week 2026 by Principal Data Scientist Yekaterina “Katya” Mijatovic and collaborators. Source

As energy utilities expand and modernize the grid, teams must verify that a growing number of field installations match their approved engineering designs. That work is essential to grid safety and reliability, but traditional compliance verification can be slow, manual, and vulnerable to human error.

Data Society Group Principal Data Scientist Yekaterina “Katya” Mijatovic will explore a new approach to this challenge at the 4th Data Science & AI Summit during London Data Science Week 2026. Her oral presentation, “From Panel Image to Wiring Compliance: An Explainable Graph-Based Verification Framework,” is part of the Applied Data Science in Industry track.

Cassandra Rounds, an energy utility partner subject matter expert, will co-present the session. The Data Society Group team also developed the research with Ishita Jain, Poornima Joshi, Alexander Berry, and Ethan Gueck.

Moving Beyond Digital Checklists

Current digital tools can help teams manage inspection checklists, but they do not typically compare images of completed field installations directly against engineering schematics. Building that capability requires more than an image-recognition model.

The system must understand the components shown in an installation, how those components are connected, and which standards apply. It must also account for normal variation in field photographs while producing findings engineers can inspect and trust.

The framework Katya and the team developed represents both the approved schematic and the completed installation as attributed graphs built on a shared ontology. The system creates structured representations of the components and connections found in each source, making it possible to compare what was designed with what was built.

A Two-Layer Approach to Compliance

The framework evaluates compliance in two layers. First, deterministic rules check universal installation requirements that should remain consistent across projects.

Next, a severity-weighted graph edit distance compares the field installation with its specific approved design. The resulting distance serves as a compliance score, while the edit path becomes a detailed list of potential discrepancies.

Every finding identifies the affected component, its severity, and the standard governing the requirement. Instead of returning a black-box score, the framework shows users what may be wrong, where the issue appears, and why it matters.

In highly regulated environments, engineers and field teams need enough information to evaluate a finding, apply their expertise, and determine the next step.

Designed for Sensitive Infrastructure

The team demonstrated the framework using medium-voltage underground switching cabinets and evaluated installations against applicable industry and manufacturer specifications. Schematic graphs were automatically extracted from production CAD drawings, while a fixed interface allows a computer vision system to provide structured data from field photographs.

Because the work involved critical energy infrastructure information, all development followed NDA-governed data-handling requirements and remained on local infrastructure, without using cloud services. This let the team build a graph-based model without exposing any sensitive utility data.

Broader Potential for Regulated Industries

Although the initial work focuses on energy infrastructure, the framework can apply to other installations governed by structural rules. Any environment where teams must compare a completed physical system with an approved design could benefit from a similar approach.

Identifying discrepancies during construction, commissioning, or maintenance can reduce costly rework and lower the risk that an overlooked defect becomes an equipment failure or safety incident. The framework’s deterministic and explainable findings could also support future agentic workflows while keeping qualified professionals involved in review and decision-making.

Katya brings extensive experience in scientific computing, machine learning, and applied AI to this work. As a Principal Data Scientist at Data Society Group, she leads technical teams and develops custom solutions and data products designed around the needs of engineers and business stakeholders, with a particular focus on the energy and utilities industry. Learn more about the project here: https://wyrevern.netlify.app/

The 4th Data Science & AI Summit will take place October 1–2, 2026, in London. Explore the full speaker list and event information, or connect with Katya on LinkedIn.