Bayesian-Based Hyperparameter Optimization of 1D-CNN for Structural Anomaly Detection.

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Massive structural health monitoring data necessitate advanced analytical techniques.
• Deep learning models show promise for diagnosing structural anomalies but require complex hyperparameter tuning.
• Current methods for hyperparameter adjustment are often subjective and experience-based.

Purpose of the Study:

• To propose a novel strategy for building and optimizing 1D-CNN models for diverse structural damage diagnosis.
• To enhance model applicability across different structure detection scenarios.
• To overcome limitations of traditional, subjective hyperparameter adjustment methods.

Main Methods:

• Developed a strategy for building and optimizing 1D-CNN models using Bayesian optimization for hyperparameters.
• Integrated data fusion technology to improve model recognition accuracy.
• Applied the method to monitor entire structures even with sparse sensor measurement points.

Main Results:

• Achieved efficient and accurate identification of parameter changes in small local elements in a simply supported beam test case.
• Verified method robustness on publicly available structural datasets, reaching 99.85% identification accuracy.
• Demonstrated significant advantages over existing methods in sensor occupancy rate, computational cost, and identification accuracy.

Conclusions:

• The proposed strategy offers a robust and accurate approach for structural damage diagnosis.
• The method enhances the applicability of deep learning models in structural health monitoring.
• This approach provides a more objective and efficient alternative to traditional hyperparameter tuning.