Implementación de técnicas avanzadas de análisis con infrarrojo cercano en conjunto con Deep clustering para el estudio de la corrosión.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La corrosión marina representa uno de los desafíos más críticos para la infraestructura costera e offshore, con costos globales superiores a 2.5 billones de dólares anuales (3.4% del PIB mundial). En Chile, este fenómeno genera pérdidas estimadas de 11.229 millones de dólares anuales, afectando sectores como la acuicultura y la industria naval. La Corrosión Influenciada por Microorganismos (MIC) representa entre el 20% y el 40% de la corrosión submarina.
Los métodos de inspección tradicionales presentan limitaciones fundamentales: son reactivos, subjetivos y resultan incapaces de detectar corrosión en etapas incipientes bajo recubrimientos protectores. Los enfoques de aprendizaje profundo supervisado, aunque efectivos, enfrentan un obstáculo crítico: la escasez de grandes conjuntos de datos públicos etiquetados, cuya generación es costosa y propensa a sesgos humanos.
Este trabajo desarrolla una metodología de visión artificial basada en Deep Clustering no supervisado para la segmentación automática de estados de corrosión mediante análisis de imágenes hiperespectrales en el infrarrojo cercano (NIR, 895-1718 nm). El sistema integra un pipeline de cuatro etapas: (1) adquisición con cámara hiperespectral Pika IR+-1060 (344 bandas espectrales), (2) reducción dimensional con PCA (344 a 3 componentes, ¿99% varianza), (3) extracción de características con ResNet-50 pre-entrenada usando SimCLR (embeddings 2048D), y (4) clustering mediante UMAP y HDBSCAN optimizados por GridSearchCV.
El sistema identificó automáticamente 12 clústeres en 103 probetas expuestas durante 156 días, correspondientes a diferentes modos de fallo (oxidación, metal desnudo, ampollamiento, agrietamiento, recubrimiento intacto). Solo el 5.23% de muestras fueron clasificadas como ruido. La validación cuantitativa confirmó la calidad de la segmentación: Silhouette Score = 0.569, Davies- Bouldin Index = 1.19, Calinski-Harabasz Score = 16,059.53. La consistencia entre conjuntos de entrenamiento y testeo demostró capacidad de generalización.
Una ventaja fundamental del pipeline reside en su escalabilidad. Al implementar un enfoque no supervisado, el sistema no requiere etiquetado manual de datos, eliminando un cuello de botella crítico. El modelo puede ser mejorado mediante el aumento de la base de datos connuevas mediciones hiperespectrales NIR, sin necesidad de costosas campañas de anotación por expertos. Este trabajo demuestra la viabilidad de utilizar imágenes hiperespectrales y Deep Clustering para clasificación automática de estados de corrosión, estableciendo una base para sistemas de inspección que revolucionen el mantenimiento predictivo de estructuras metálicas.
Marine corrosion represents one of the most critical challenges for coastal and offshore infrastructure, with global costs exceeding 2.5 trillion USD annually (3.4% of global GDP). In Chile, this phenomenon generates estimated losses of 11.229 billion dollars per year, affecting strategic sectors such as aquaculture and the naval industry. Microbiologically Influenced Corrosion (MIC) accounts for 20-40% of subsea corrosion. Traditional inspection methods present fundamental limitations: they are reactive, subjective, and unable to detect incipient corrosion under protective coatings. Supervised deep learning approaches, although effective, face a critical obstacle: the scarcity of large labeled public datasets, whose generation is costly and prone to human biases. This work develops a computer vision methodology based on unsupervised Deep Clustering for automatic segmentation of corrosion states through near-infrared (NIR, 895-1718 nm) hyperspectral image analysis. The system integrates a four-stage pipeline: (1) acquisition with Pika IR+-1060 hyperspectral camera (344 spectral bands), (2) dimensional reduction with PCA (344 to 3 components, ¿99% variance), (3) feature extraction with ResNet-50 pre-trained using SimCLR (2048D embeddings), and (4) clustering using UMAP and HDBSCAN optimized by GridSearchCV. The system automatically identified 12 clusters in 103 specimens exposed for 156 days, corresponding to different failure modes (oxidation, bare metal, blistering, cracking, intact coating). Only 5.23% of samples were classified as noise. Quantitative validation confirmed segmentation quality: Silhouette Score = 0.569, Davies-Bouldin Index = 1.19, Calinski-Harabasz Score = 16,059.53. Consistency between training and testing sets demonstrated generalization capability. A fundamental advantage of the pipeline lies in its scalability. By implementing an unsupervised approach, the system does not require manual data labeling, eliminating a critical bottleneck. The model can be improved by expanding the database with new NIR hyperspectral measurements, without costly expert annotation campaigns. This work demonstrates the viability of using hyperspectral imaging and Deep Clustering for automatic classification of corrosion states, establishing a foundation for inspection systems that revolutionize predictive maintenance of metallic structures.
Marine corrosion represents one of the most critical challenges for coastal and offshore infrastructure, with global costs exceeding 2.5 trillion USD annually (3.4% of global GDP). In Chile, this phenomenon generates estimated losses of 11.229 billion dollars per year, affecting strategic sectors such as aquaculture and the naval industry. Microbiologically Influenced Corrosion (MIC) accounts for 20-40% of subsea corrosion. Traditional inspection methods present fundamental limitations: they are reactive, subjective, and unable to detect incipient corrosion under protective coatings. Supervised deep learning approaches, although effective, face a critical obstacle: the scarcity of large labeled public datasets, whose generation is costly and prone to human biases. This work develops a computer vision methodology based on unsupervised Deep Clustering for automatic segmentation of corrosion states through near-infrared (NIR, 895-1718 nm) hyperspectral image analysis. The system integrates a four-stage pipeline: (1) acquisition with Pika IR+-1060 hyperspectral camera (344 spectral bands), (2) dimensional reduction with PCA (344 to 3 components, ¿99% variance), (3) feature extraction with ResNet-50 pre-trained using SimCLR (2048D embeddings), and (4) clustering using UMAP and HDBSCAN optimized by GridSearchCV. The system automatically identified 12 clusters in 103 specimens exposed for 156 days, corresponding to different failure modes (oxidation, bare metal, blistering, cracking, intact coating). Only 5.23% of samples were classified as noise. Quantitative validation confirmed segmentation quality: Silhouette Score = 0.569, Davies-Bouldin Index = 1.19, Calinski-Harabasz Score = 16,059.53. Consistency between training and testing sets demonstrated generalization capability. A fundamental advantage of the pipeline lies in its scalability. By implementing an unsupervised approach, the system does not require manual data labeling, eliminating a critical bottleneck. The model can be improved by expanding the database with new NIR hyperspectral measurements, without costly expert annotation campaigns. This work demonstrates the viability of using hyperspectral imaging and Deep Clustering for automatic classification of corrosion states, establishing a foundation for inspection systems that revolutionize predictive maintenance of metallic structures.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Químico/a.
Keywords
Corrosión marina, Corrosión influenciada microbiológicamente, Teoría de los agregados (Física nuclear)