•  
  •  
 

Abstract

This paper discusses the motivation behind turbine migration and addresses the challenges of using NN (neural network) computing systems. Moreover, it focuses on high-dimensional data from WT (wind turbine) blades. The three key aspects addressed in this study are turbine migration, overfitting, and strict feature selection. To evaluate the performance of the machine learning system, the study considers the characteristics of WT blades, specifically the similarity in blade color and the differences in shape. The authors apply pre-processing techniques, in particular a bilateral filter, in conjunction with the SS (selective synthesizer) of blade fouling patterns. The SS method adopts the framework of ResNet50 to evaluate the computational efficiency. The experimental results show that the introduction of the SS method for feature selection improves the accuracy rate of the NN model to over 92%. For data validation, the study employs the YOLO (You Only Look Once) deep learning framework. Specifically, YOLOv4-Tiny is used due to its trade-off between recognition speed and accuracy. In addition, YOLOv4-Tiny was integrated with the Nvidia Jetson Nano edge computing hardware. Overall, the article focuses on the use of machine learning techniques, such as preprocessing and feature selection, to improve the performance of NN computing systems in analyzing high-dimensional data from WT blades. The authors validate their approach using the YOLO framework, specifically YOLOv4-Tiny, and highlight the integration with Nvidia Jetson Nano for edge computing.

Download
COinS