Structural damage detection for a small population of nominally equal beams using PSO-optimized Convolutional Neural Networks

Autor(es)

Ruiz, Dianelys Vega ; de Braganca, Cassio Scarpelli Cabral ; Poncetti, Bernardo Lopes ; Bittencourt, Tulio Nogueira ; Futai, Marcos Massao

• Tipo de documento: Artigo Científico
• Fonte: MECHANICAL SYSTEMS AND SIGNAL PROCESSING; v. 225, p. 20-pg., 2024-12-28.

Resumo

This paper investigates the application of one-dimensional convolutional neural networks (1D CNNs) for damage detection in standardized structural components, addressing the limitations of existing methods that typically focus on individual structures. To assess damage detection capabilities across populations, particularly in situations where no prior damage data from the structures are available, a 1D CNN is trained and validated on simulated damage data and then tested using experimental data from a population of beams. The data consists of vibration responses from five aluminum beams with damage introduced by rectangular notches at different locations. A finite element model that incorporates beam-to-beam variability through random coefficients is employed to generate the training data. To optimize feature extraction across the population, the number and size of kernels in the convolutional layers are fine-tuned using Particle Swarm Optimization (PSO). The method's robustness is evaluated by comparing its damage detection accuracy with additional deep learning models and conventional machine learning approaches that rely on manual feature extraction. The proposed method achieves near-perfect accuracy-up to 100% in training and 99.6% in validation using simulated data, outperforming the other models in accuracy and computational efficiency. However, when applied to the experimental population, its accuracy drops, particularly in identifying the precise damage state. Nevertheless, it successfully distinguishes between damaged and undamaged samples with an accuracy of 76.3%. These results extend the applicability of 1D CNNs for detecting damage across experimental populations of similar structural components, requiring only a single CNN to identify the presence of damage across the population. (AU)

• Processo FAPESP: 22/13045-1 - Desenvolvimento de metodologias para a identificação de danos em pontes ferroviárias com base em monitoramento indireto e aprendizado de máquina
• Beneficiário: Cassio Scarpelli Cabral de Braganca
• Modalidade de apoio: Bolsas no Brasil - Doutorado