Machine Learning for Metal Fatigue Crack Prediction Using Lamb Wave Signals

Overview of Metal Fatigue Crack Experiments

This package is based on a comprehensive experimental dataset collected from fatigue testing of riveted aluminum lap-joint specimens instrumented with piezoelectric (PZT) sensors. Each specimen consists of two 1.6 mm-thick aircraft-grade 2024-T3 aluminum sheets joined by rivets. They were subjected to cyclic loading until metal fatigue crack initiation and growth occurred.

To track the damage, PZT actuators and sensors were arranged in a pitch–catch configuration. This setup generated and captured Lamb wave signals during each loading stage. The repeated measurements helped reduce uncertainty. The dataset includes eight specimens (T1–T8). T1–T6 are used for training, while T7–T8 serve as validation samples. Both constant-amplitude and variable-amplitude loading conditions were included.

Applying Machine Learning Crack Detection Techniques

This package teaches participants how to use machine learning crack detection methods to analyze Lamb wave signals and predict fatigue crack growth. By using the PHM 2019 Aluminum Lap Joint Fatigue Dataset, the package covers feature extraction, preprocessing, ensemble model training, and evaluation in Python. As a result, users gain hands-on experience in data-driven structural health monitoring (SHM) and predictive maintenance of metallic structures.

Because the package focuses on crack detection using machine learning, it demonstrates how algorithms interpret Lamb wave responses to identify crack initiation and progression. These techniques enhance early detection and improve predictive accuracy in practical SHM applications.

Lamb Wave Signals for Metal Fatigue Crack Monitoring

Fatigue tests on aluminum lap-joint specimens were monitored using Lamb wave signals recorded at multiple fatigue cycles. These signals correlate strongly with crack propagation behavior. Ground-truth crack lengths were collected through optical measurements.

The dataset is divided into training and validation subsets to support model development. This structure helps users build reliable predictive models for fatigue life estimation. Therefore, the package enables a smooth connection between physics-based understanding and data-driven algorithms for metal fatigue crack detection and prediction.