Wind energy operations rely on high equipment availability, where unexpected turbine generator failures lead to costly downtime, reactive maintenance, and lost production. This project implements a neural network–based predictive maintenance pipeline to identify wind turbine generator failures from sensor data, enabling earlier intervention and improved operational reliability.
Beyond introducing neural networks, this work represents a transition from isolated model experimentation to end-to-end pipeline design, with explicit emphasis on decision support, false-negative minimization, and real maintenance impact.
Develop a predictive maintenance classification pipeline that:
- Detects potential turbine generator failures before breakdown
- Prioritizes minimizing false negatives to reduce unplanned downtime
- Supports proactive maintenance scheduling and cost reduction
The objective is to improve reliability and maintenance decision-making, not to optimize accuracy in isolation.
- Sensor-based operational data from wind turbine generators
- Numeric features representing equipment performance and operating conditions
- Binary classification target indicating failure vs. normal operation
- Strong class imbalance reflecting the rarity of failure events in production systems
- Data ingestion, validation, and preprocessing of sensor signals
- Handling of missing values and normalization of continuous inputs
- Feature preparation tailored for neural network architectures
- Establishment of baseline classification models for comparison
- Design and implementation of a feedforward neural network
- Experimentation with activation functions and network configuration
- Iterative refinement to balance sensitivity, stability, and generalization
This project marks the first instance where modeling decisions are treated as pipeline components, not standalone experiments.
- Evaluation focused on operational relevance, including:
- Recall and false-negative reduction
- Confusion matrix–driven failure analysis
- Overall classification reliability under imbalance
- Model selection guided by maintenance impact rather than accuracy alone
- Neural networks effectively capture nonlinear relationships in turbine sensor data.
- Prioritizing recall materially improves maintenance readiness and failure detection.
- Activation function choice has a significant effect on model sensitivity.
- Properly tuned neural networks provide meaningful gains over simpler classifiers in this domain.
- Model performance depends heavily on sensor data quality and coverage.
- Rare failure events constrain the number of positive-class examples.
- The pipeline is designed for decision support, not autonomous maintenance execution.
- Incorporation of temporal and sequential modeling approaches
- Cost-sensitive optimization aligned with maintenance economics
- Integration into predictive maintenance and monitoring workflows
- Continuous retraining as new operational data becomes available
notebooks/— End-to-end predictive maintenance pipeline implementation, including preprocessing, neural network modeling, and evaluationREADME.md— Project overview, methodology, findings, and pipeline framing
This project was completed as part of the Introduction to Neural Networks course in the
UT Austin / Great Learning Post Graduate Program in AI & ML (2025–2026) and has been refactored to reflect the transition from academic modeling to professional pipeline development.