Hyundai Predictive Maintenance Project: A Case Study

This project documents my process of attempting to build a predictive maintenance model for Hyundai vehicles. The initial goal was to predict whether a vehicle would have an "Anomaly Indication" based on a provided dataset of sensor readings and maintenance history.

Ultimately, this project serves as a practical example of a common and important scenario in data science: sometimes, the most valuable insight is discovering that the available data is not sufficient to solve the problem.

Project Structure

.
├── data/
│   └── cars_hyundai.csv
├── images/
│   ├── anomaly_distribution.png
│   ├── correlation_heatmap.png
│   └── feature_importance.png
│   └── ... (other plots)
├── models/
│   ├── best_random_forest_model.pkl
│   └── ... (other models)
├── src/
│   ├── 01_EDA_and_Preprocessing.py
│   ├── 02_model_training.py
│   ├── 03_hyperparameter_tuning.py
│   ├── 04_xgboost_modeling.py
│   └── 05_feature_importance.py
└── README.md

My Approach

My approach followed a standard data science workflow:

1. Exploratory Data Analysis (EDA): I first analyzed the dataset to understand its structure, check for missing values, and visualize the relationships between features. I found that the target variable, Anomaly Indication, was well-balanced, which was a promising start.

2. Modeling: I treated the problem as a supervised binary classification task and experimented with several models to find a predictive pattern:

   • A baseline RandomForestClassifier.
   • A hyperparameter-tuned RandomForestClassifier using GridSearchCV.
   • An XGBoost classifier, a more powerful gradient boosting model.

3. Evaluation: I evaluated each model on a dedicated test set using standard classification metrics, including accuracy, precision, recall, and F1-score, to ensure the performance measurement was robust.

Results and Final Conclusion

My modeling process confirmed the initial hypothesis that the data lacks predictive power. Despite a rigorous process of modeling and tuning, no model was able to achieve a performance significantly better than a random guess.

Here is a summary of the test set performance for each model:

Model	Accuracy	Weighted Avg F1-Score
Baseline Random Forest	57.0%	0.57
Tuned Random Forest	55.0%	0.55
XGBoost	49.1%	0.48

Notably, the performance slightly decreased as the model complexity increased. The most powerful model, XGBoost, produced the worst result. This is a strong indicator that the models were attempting to fit to noise rather than a true underlying signal.

The key insight came from a feature importance analysis I performed on the best model. The analysis clearly showed that none of the provided features (Engine Temperature, Brake Pad Thickness, Tire Pressure, Maintenance Type) had strong predictive power.

The Verdict

My conclusion is that the dataset, in its current form, is insufficient to build a reliable predictive maintenance model.

This is not a failure of the modeling process, but a crucial finding about the data itself. The features available do not contain a strong enough signal to allow a machine learning model to distinguish between anomalous and non-anomalous events.

Learning from this Project

This project highlights a fundamental truth in machine learning: the quality and predictive power of your data are more important than the complexity of your model. It demonstrates that a thorough and methodical process can lead to the important conclusion that a problem is not solvable without better data.

For this specific problem, I would recommend the following to move forward:

• Acquire Time-Series Data: Sensor readings over a period of time leading up to a maintenance event would be far more predictive.
• Collect Data from Additional Sensors: Information from components like oil viscosity sensors, vibration sensors, or exhaust readings could provide the necessary signal.
• Include Operational History: Vehicle age, mileage, and driving conditions are critical context that is currently missing.
• Use More Granular Labels: A more detailed definition of what constitutes an "anomaly" or failure would help clarify the target for the model.