Deployment of an Intelligent Robotic Diagnostic System for Outdoor Switchgear Equipment Based on Multispectral Analysis and AI

A central premise of the study is that decision-grade diagnostics is constrained not only by model accuracy, but also by information sufficiency, traceability, and asset-to-data consistency.

In practical applications, diagnostic evidence is often fragmented across temporal data streams, relational asset records, and file-based inspection repositories, which limits its suitability for time-dependent decision making.

The paper demonstrates how modern information systems, AI-based analytics, and distributed computing infrastructure can be combined to address asset monitoring challenges in the power industry by enforcing consistent metadata, precise asset binding, and structured diagnostic data management. A methodology for forecasting defect development is introduced, based on the dynamics of thermal and electrical diagnostic indicators derived from multispectral robotic inspections. Approximation and extrapolation of parameters such as surface temperature distribution and intensity of surface partial discharges enable prediction of defect progression and support maintenance decision making.

The developed cyber-physical system autonomously inspects outdoor switchgear and collects multispectral diagnostic data in visible (RGB), infrared (IR), and ultraviolet (UV) ranges through coordinated UAV and UGV missions. During industrial deployment, 812 equipment units were inspected from 1,705 viewpoints over 21 hours, achieving 95 percent repeatability.

This repeatability is ensured by precise association of each diagnostic observation with a specific equipment unit and inspection event, enabling consistent comparison of diagnostic parameters over time. Depending on switchgear layout, the ratio of data collected by UGV and

UAV platforms ranged from 85:15 to 75:25.

Manual inspection workflows are prone to errors such as image misassignment and viewpoint variability, which limit the reliability of temporal analysis. In contrast, the robotic system enables structured data acquisition with a pointing accuracy of ±1 degree and eliminates misassignment at the data ingestion stage. Multispectral image data are processed using convolutional neural network models for equipment recognition and defect localization, trained on an author-collected dataset of approximately five thousand RGB, IR, and UV images acquired under diverse environmental conditions. The trained models and forecasting methodology were implemented in the developed software platform and validated through industrial deployment. The main results include stable automated diagnostics with high repeatability, multispectral defect localization, and forecast-driven decision support, with achieved F1-scores of 84 percent for thermal defects and 87 percent for electrical defects.