Sustained output from photovoltaic installations depends heavily on the condition of individual panel surfaces, yet routine inspection across large solar farms remains a predominantly manual and resource-intensive task. This paper presents the design, implementation, and field evaluation of an autonomous ground rover developed for systematic photovoltaic panel inspection. The rover traverses solar panel rows by tracking a physical guide line and executes a structured inspection sequence at each panel stop without operator intervention. A triggered OV3660 camera captures a surface image that is classified on-device by a Convolutional Neural Network trained through the Edge Impulse TinyML platform on a dataset spanning three surface conditions: clean, dust-covered, and contaminated by bird droppings. This inference runs entirely on an ESP32-S3 microcontroller, requiring no cloud connectivity. Complementary dual-LDR sensing quantifies surface reflectance loss, and integrated voltage-current monitoring records electrical output at each panel. Results from all three measurement channels are consolidated and transmitted to an inspection analytics dashboard, which presents panel-level health status in a format accessible to non-specialist operators. Across a controlled field evaluation, the system achieved strong classification accuracy, consistent stopping precision, and a throughput suitable for regular inspection scheduling. The system was able to classify them with a high accuracy of 92-95%, and this was done in just 2-3 seconds, indicating that it is efficient enough to be used in real-time checks. The rover was assembled from commercially available components at a modest hardware cost, demonstrating that capable automated inspection is achievable without expensive infrastructure.
Jagadeeswaran et al. (Thu,) studied this question.