Feature selection for neural network based defect classification of ceramic components using high frequency...

Current methods for inspecting ceramic armor components often lack the precision needed to detect microstructural flaws. Non-destructive testing is essential to ensure component reliability. Prior research has shown that ultrasonic techniques can identify internal defects. However, these methods often require manual interpretation and lack automation. Feature selection remains a challenge in classification systems. High-dimensional data can reduce model efficiency and accuracy. Dimensionality reduction techniques are widely used but may not always optimize classification outcomes. This gap motivated the need for a more effective automated inspection system. The research addresses the need for accurate and efficient defect detection in ceramic materials.

This study aims to develop an automated classification system for ceramic components using high-frequency ultrasound. The goal is to detect and locate defects before deployment. The system must reduce manual inspection and improve reliability. A key problem is selecting optimal features for classification. The motivation is to enhance inspection accuracy and speed. The study compares two feature selection techniques. The focus is on reaction-sintered silicon carbide components. The objective is to identify the most effective method for defect classification.

The study uses wavelet-based feature extraction from ultrasonic signals. Features are extracted from the region of interest in the signals. An artificial neural network is used to evaluate these features. Genetic algorithms are applied for feature selection optimization. Principal Component Analysis is also used as a comparison method. Both techniques are evaluated for classification accuracy. The performance of each method is tested on defect data. The results are compared in terms of feature count and accuracy.

Principal Component Analysis achieved 96% classification accuracy. Genetic algorithms reached 94% accuracy in the same tests. PCA selected fewer features than the genetic algorithm. Both methods improved classification over random feature sets. The study found PCA to be more efficient in feature selection. The results suggest PCA is better suited for this application. The neural network performed well with both methods. The data supports the use of PCA for optimal classification.

The study shows that PCA outperforms genetic algorithms in this context. The authors propose PCA as the preferred method for feature selection. The results align with the goal of improving classification accuracy. The findings suggest that PCA reduces feature count effectively. The system supports automated inspection of ceramic components. The authors suggest that this method could be applied to other materials. The study confirms the feasibility of automated ultrasonic inspection. The results support the use of PCA in future classification systems.