Resource-Adaptive Neuro-Fuzzy Edge System for Wind Turbine Blade Defect Diagnosis

Abstract

Reliable UAV-based inspection of wind turbine blades requires diagnostic models that maintain acceptable decision quality under image degradation and limited edge-computing resources. This study proposes a resource-adaptive neuro-fuzzy edge system for wind turbine blade defect diagnosis. SqueezeNet transforms input images into four-dimensional feature descriptors, which are subsequently processed by three complementary neuro-fuzzy classifiers based on prototype learning, recurrent associative dynamics, and fuzzy rule induction. A controller selects a conditional cascade, a full heterogeneous ensemble, or a safe mode according to classification uncertainty and the available computational resources. The experimental evaluation was performed using a cleaned dataset of 744 unique images representing cracks, erosion, corrosion, and normal blade surfaces. The primary prototype-based neuro-fuzzy classifier achieved an accuracy of 0.7667 for original images and 0.7067 for both noisy and distorted images. The conditional cascade reduced neuro-fuzzy inference time by 37.75–39.98% compared with the full ensemble, while the difference in accuracy did not exceed 0.0133. Although the adaptive ensemble did not outperform the primary classifier in terms of accuracy, it provided better-calibrated class scores, with log-loss values ranging from 1.2349 to 1.3612. The results demonstrate that frame-wise routing can reduce computational costs while preserving classification performance.