Semi-automatic Condition Inspection of Lead Rubber Bearings based on Deep Learning

In different countries with seismic active zones such as Peru, Japan and the USA, seismic isolation have been used in buildings to improve their seismic resilience. Nonetheless, several devices have been operating for over ten years, possibly showing signs of deterioration. Due to their critical role in the seismic response, an efficient inspection system is important to ensure proper isolation conditions. In this context, international standards about the inspection of Lead Rubber Bearings (LRB) have been proposed, as well as more sophisticated noninvasive methods. After reviewing proposed methods, three limitations were identified: a) the lack of a detailed inspection protocol can cause biased and uncertain assessments, b) there is not clear deterioration levels of seismic isolators with parameter thresholds, and c) non-invasive methods are laborious, time consuming and costly for implementation in operational isolators. This study integrates ambient vibration signals with Neural Networks to propose an agile, accurate and cost-effective seismic isolator inspection system. The system is trained and developed with real data from existing LRBs. To this end, 14 LRB isolators from a case study, with nearly 11 years in operation, were inspected using traditional methods. Additionally, a four-level deterioration matrix was proposed. Then, ambient vibration signals were locally recorded using microtremors and relevant features were extracted for each isolator. Consequently, a Fully Connected Deep Neural Network (FCDNN) was trained with signal windows and other relevant variables. After training and hyperparameter tuning, the proposed FCDNN achieved an Accuracy and a ?1????? of 0.93 and 0.94 on the test set, respectively. Using the trained FCDNN and SHAP-based explanation, the following conclusions were drawn: a) the geometric and mechanical parameters of the LRB isolators define the susceptibility and ease of a device to maintain or change from a mild to severe level of deterioration; and b) the signal features define vibration patterns in the temporal and spectral domain in specific levels of deterioration. Finally, a web GUI is developed to predict the level of deterioration of a device after measuring ambient vibrations. The proposed system has the potential to revolutionize the frequent inspection and monitoring of isolation devices.