This project analyzes sensor data from a semiconductor manufacturing process to detect faulty products using unsupervised anomaly detection. The goal is to navigate a complex, real-world dataset, handle challenges like missing values and class imbalance, and evaluate machine learning models. A final interactive dashboard was created to present the findings.
An interactive dashboard summarizing the project's key findings and model performance was created using Tableau.
View the Interactive Dashboard on Tableau Public
- Dataset: SECOM Dataset from UCI ML Repository
- Technologies: Python, Pandas, Scikit-learn, Matplotlib, Jupyter Notebook, Tableau.
- Data Cleaning: Addressed significant missing data by removing features with over 50% null values and imputing the rest with the median.
- Exploratory Data Analysis (EDA): Discovered a severe class imbalance, with only ~7% of products labeled as 'Pass' (+1). This highlighted that
Accuracywould be a misleading metric. - Data Preparation: The data was split into stratified training (70%) and test (30%) sets. Features were then scaled using
StandardScaler. - Modeling: Trained two unsupervised models,
Isolation ForestandOne-Class SVM.
The project followed an iterative modeling process. A baseline was established with unsupervised models, followed by advanced techniques to address the severe class imbalance.
Performance was evaluated on Recall for the minority 'Pass' class.
| Approach | Model | 'Pass' Class Recall | 'Pass' Class F1-Score |
|---|---|---|---|
| Baseline | 1. Isolation Forest | 16% | 0.17 |
| Attempt 1 | 2. XGBoost + SMOTE | ~10% | ~0.16 |
| Attempt 2 | 3. XGBoost + PCA + SMOTE | 16% | 0.18 |
The analysis revealed that even advanced techniques like SMOTE and PCA did not significantly improve performance over the simple baseline model. This highlights the inherent difficulty and noisy nature of this high-dimensional dataset. The project concludes that achieving a high recall rate would likely require extensive hyperparameter tuning or domain-specific feature engineering beyond the scope of this initial analysis.
This iterative process itself is a key finding, demonstrating a realistic approach to complex data science problems where solutions are not always straightforward.
- Clone the repository and install dependencies:
pip install -r requirements.txt - Run the Jupyter Notebooks in order.