Optimization of CNN Hyperparameters using bioinspired approaches for Photovoltaic Panel Defect Classification

Abstract

This paper presents a comprehensive and systematic study on optimizing the ResNet-50 convolutional neural network for photovoltaic (PV) panel defect detection by leveraging a diverse portfolio of xisteen recent bioinspired metaheuristic algorithms. Unlike prior studies that typically focus on one or two optimization methods, our approach rigorously explores and benchmarks a wide range of nature-inspired optimizers—encompassing both animal behavior-based and physics-inspired strategies—within the same experimental framework and data context.

Specifically, we implement and compare the performance of sixteen prominent bioinspired algorithms, including Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Grey Wolf Optimizer (GWO), Whale Optimization Algorithm (WOA), Harris Hawks Optimization (HHO),  Walrus Optimization Algorithm (WaOA), Equilibrium Optimization (EO), Fossa Optimization Algorithm (FOA), Prairie Dog Optimization (PDO), and Hare Escape Optimization Algorithm (HEOA), for the automatic hyperparameter tuning of ResNet-50. Each optimizer is systematically integrated into a deep learning pipeline targeting the multi-class PV panel defect classification problem, enabling fair and reproducible evaluation.

Our study not only benchmarks these algorithms on classification accuracy, convergence speed, and robustness, but also provides novel insights into their suitability for complex vision tasks involving real-world, high-dimensional datasets. As a result, the findings deliver a new reference point for both the Photovoltaic panel analytics and deep learning optimization communities, guiding future development and application of bioinspired methods for defect detection and beyond.