AI-Enhanced Predictive Maintenance for Alloy Material Fatigue Life.

Thesis Abstract

The increasing reliance on alloy components in critical engineering applications such as aerospace, automotive, and structural industries necessitates robust maintenance strategies to prevent unexpected failures attributable to material fatigue. Traditional maintenance approaches, often based on scheduled inspections or reactive repairs, frequently result in either unnecessary costs or catastrophic failures due to unforeseen fatigue damage accumulation. This study aims to develop and evaluate an AI-enhanced predictive maintenance framework specifically tailored for assessing the fatigue life of alloy materials, leveraging data-driven techniques to optimize maintenance schedules and extend component service life. The primary objectives include (1) identifying key material and operational variables influencing alloy fatigue life; (2) designing a comprehensive data acquisition protocol to gather real-time signal data from alloy components under simulated operational conditions; (3) developing machine learning models—specifically, convolutional neural networks (CNNs) and gradient boosting machines (GBMs)—to predict fatigue damage progression; and (4) validating the predictive models against empirical fatigue test data. The study also aims to compare the predictive accuracy and computational efficiency of these models and to formulate an integrated AI-based maintenance decision-support system. Employing a mixed-method research design, the study combines experimental fatigue testing with quantitative data analysis. The population comprises 150 specimens of high-strength aluminum alloy subjected to cyclic loading under laboratory-controlled conditions simulating operational stresses. A stratified random sampling technique selects 120 specimens for model training and validation, while the remaining 30 are reserved for independent testing. Data collection instruments include strain gauges, acoustic emission sensors, and digital image correlation systems, which record real-time deformation, crack initiation, and propagation signals during fatigue testing. The validity and reliability of the sensors are established through calibration procedures and statistical consistency checks, such as Cronbach’s alpha and intraclass correlation coefficients. Data analysis involves advanced statistical and machine learning techniques. Descriptive analysis summarizes the signal features and fatigue life measurements. Predictive models are developed using Python-based frameworks—TensorFlow for CNNs and XGBoost for GBMs—and evaluated using metrics like mean absolute error (MAE), root mean squared error (RMSE), and R-squared values. Model interpretability is enhanced through SHAP (SHapley Additive exPlanations) values, enabling insight into feature importance. Additionally, survival analysis models, rooted in the theory of reliability engineering, are employed to estimate fatigue life distributions. Ethical considerations involve ensuring data integrity, sensor calibration, and adherence to safety protocols during experimental procedures. Expected findings indicate that AI-driven models can accurately predict alloy fatigue life within 10% error margins, demonstrating superior performance over traditional statistical models. The models are anticipated to reveal significant predictors such as cyclic strain amplitude, acoustic emission activity, and microstructural features. Furthermore, the integration of these models into a decision-support system is projected to facilitate real-time maintenance scheduling, thereby reducing downtime and maintenance costs by up to 25%. It is also expected that the hybrid CNN-GBM approach will outperform standalone models in prediction accuracy and computational efficiency. This research significantly contributes to existing knowledge by advancing the application of artificial intelligence in materials fatigue assessment and maintenance optimization. It bridges the gap between theoretical machine learning models and practical maintenance strategies for alloy components. The study offers a novel, replicable framework for nondestructive evaluation and predictive maintenance that can be adapted to other metallic materials and operational contexts. The study concludes that AI-enhanced predictive maintenance systems are pivotal in transforming traditional maintenance paradigms in engineering applications. Recommendations include the integration of the developed models into industrial maintenance workflows, further validation in field conditions, and expansion of sensor arrays to encompass additional damage indicators. Future research should focus on real-time deployment automation, multi-material modeling, and the incorporation of industry 4.0 concepts such as IoT-enabled sensor networks to enable continuous, autonomous condition monitoring.