Proposal of inspection method for isolation and vibration control devices using deep learning (Anomaly detection based on force-displacement relationship images of normal data)

Abstract

This paper proposes a novel inspection system for seismic isolation and vibration control devices using unsupervised deep learning to enhance evaluation reliability and objectivity. Conventional force and stiffness assessments through loading tests require human inspectors, creating potential subjective bias and necessitating impartial third-party evaluation. The proposed deep learning system minimizes human intervention, significantly improving test result consistency while eliminating operator bias. The unsupervised learning approach enables the model to learn exclusively from normal operational data, facilitating detection of anomalies in previously unseen patterns with high sensitivity. This paper presents a comprehensive framework encompassing data generation, preprocessing, and model inference. Experimental validation using oil dampers and laminated rubber bearings, representative components in seismic isolation technology, demonstrates the method's effectiveness with approximately 98% classification accuracy for oil dampers and 100% for rubber bearings in distinguishing normal from anomaly conditions. These results confirm the system's viability for large-scale manufacturing deployment. Furthermore, anomaly visualization capabilities provide valuable insights for manufacturers and regulatory bodies, reinforcing the importance of objective and transparent evaluation. This inspection system establishes a robust foundation for quality control in seismic isolation and vibration control technologies, with significant potential for broader adoption toward ensuring safer and more reliable infrastructure.