Autonomous defect estimation in aluminum plate and prognosis through stochastic process modeling

Area of Science:

• Materials Science and Engineering
• Mechanical Engineering
• Data Science

Background:

• Aluminum alloys are crucial for lightweight engineering, demanding robust methods for damage assessment.
• Existing supervised machine learning (ML) approaches for damage detection face challenges like large data requirements, overfitting, and high computational expense.

Purpose of the Study:

• To develop an unsupervised learning framework for accurate and efficient damage detection and prognosis in aluminum alloy components.
• To overcome limitations of traditional supervised ML methods in structural health monitoring.

Main Methods:

• Utilized Scanning Acoustic Microscopy (SAM) for data acquisition.
• Extracted features in time, frequency, and time-frequency domains via Short-Time Fourier Transform (STFT).
• Combined k-means clustering for defect localization/sizing with a multi-phase gamma process for damage progression modeling.

Main Results:

• Achieved accurate localization and sizing of surface defects without labeled data.
• Successfully modeled the stochastic progression of damage over time.
• Demonstrated high accuracy in estimating defect geometry and prognostic trajectories with low computational complexity.

Conclusions:

• The proposed unsupervised framework offers an interpretable and generalizable solution for damage detection and prognosis.
• This method shows significant potential for real-time structural health monitoring (SHM) in safety-critical applications.
• The approach is adaptable to various materials, advancing the field of non-destructive testing.