🔧 Steel Plate Fault Detection Dashboard

TL;DR

• Built an interactive dashboard around the UCI Steel Plates Faults dataset to analyze and classify different types of surface defects.
• Implemented both unsupervised (e.g., clustering / anomaly detection) and supervised (e.g., tree-based classifiers) models to compare approaches.
• Exposed key evaluation metrics and visualizations (class distribution, decision boundaries, feature importance) in a Streamlit app.
• Deployed the dashboard on Hugging Face Spaces so others can explore the dataset and models in the browser.

Problem & Context

Steel plates used in manufacturing can exhibit various surface faults (scratches, stains, bumps, etc.) that impact quality and downstream processes. Automatically detecting and classifying these faults from sensor measurements can help automate quality control and reduce human inspection workload.

The goal of this project is not only to train a single classifier, but to provide an interactive environment where users can:

• Understand the dataset and class distribution.
• Compare unsupervised vs supervised modeling strategies.
• Visually inspect how models behave on different subsets of the data.

Data & Inputs

• UCI Steel Plates Faults dataset, containing numerical features describing geometry, edge wave, surface properties, and more.
• Each row is labeled with one of several fault types (or no fault), treated as a multi-class classification problem.
• Data preprocessing includes feature scaling, basic cleaning, and handling of class imbalance for supervised models.

The dashboard provides an overview of feature distributions, correlations, and class frequencies, helping users understand why certain faults might be harder to classify than others.

Approach & Models

The dashboard is structured around two complementary modeling strategies:

• Unsupervised / semi-supervised: clustering or anomaly detection to see whether fault types form natural groups in feature space without using labels.
• Supervised classification: training models such as Random Forests or Gradient Boosting to predict fault type directly from features.

Users can switch between model types, adjust basic parameters, and immediately see the impact on performance metrics and visualizations.

Results & Insights

The supervised models achieve solid performance on the more common fault classes, with decreasing accuracy on rare fault types due to limited examples and overlapping feature patterns. Unsupervised methods provide some structure but cannot fully separate all classes.

• Supervised classifiers (e.g., Random Forest) show good overall accuracy and balanced performance across major classes.
• Confusion matrices reveal which fault types are often confused and highlight where more data or features would help.
• Feature importance plots show which measurements contribute most to distinguishing specific fault types.

The dashboard makes these trade-offs visible and encourages exploration rather than hiding everything inside a notebook.

Implementation

• Implemented in Python using Pandas for data handling and scikit-learn for modeling.
• Dashboard built with Streamlit, with separate sections for data overview, model selection, and evaluation.
• Deployed on Hugging Face Spaces so the entire pipeline runs in the cloud without local setup.

The codebase is structured to keep data loading, model definitions, and Streamlit layout modular, allowing new models or visualizations to be added with minimal changes.

Challenges & Lessons Learned

• Handling class imbalance is critical for fair evaluation; naive accuracy can be misleading when some fault types are much rarer.
• Unsupervised approaches provide interesting structure but are not a drop-in replacement for supervised models when labels are available.
• Building a dashboard forces you to think about how others will interact with your models and what questions they might want to answer.

This project strengthened my understanding of industrial tabular ML problems and how to present them in an interactive, exploratory way.

Links

Live Dashboard on Hugging Face Spaces  ·  GitHub Repository  ·  Back to all projects