05 / ANOMALY DETECTION / ROBOTICS
Anomaly Detection in Robotics
A statistical learning project focused on distinguishing anomalous robotic behavior from normal operating patterns through supervised and unsupervised approaches.
The problem
Robotic systems generate patterns of measurements that describe normal operation. When those patterns change, the deviation can indicate an unexpected movement or an anomalous condition.
The project studies this as a normal-vs-anomaly learning problem , exploring multiple statistical approaches to understand which methods are most suitable for separating abnormal behavior from ordinary operation.
Exploration and modeling
Explore the data
Use exploratory analysis to understand variable distributions, relationships, and how normal and anomalous observations differ.
Study lower-dimensional structure
Apply PCA to investigate whether the dominant directions of variation help reveal structure in the data and improve interpretability.
Supervised experiments
Train models with labeled normal/anomaly examples and evaluate their ability to separate the two classes.
Unsupervised experiments
Investigate approaches that do not rely on the same level of labeled supervision, comparing their behavior with the supervised baseline.
Neural-network exploration
Extend the comparison with neural-network experiments to assess a more flexible nonlinear modeling approach.
What I focused on
- Understanding anomaly detection from both classification and unsupervised-learning perspectives.
- Using PCA as an exploratory and dimensionality-reduction tool.
- Comparing approaches under the same dataset and evaluation context.
- Evaluating not only predictive performance but also the practical complexity of different methods.
The main takeaway was that anomaly detection is rarely a one-algorithm problem: the right solution depends heavily on how much labeled anomaly data is available and on the trade-off between interpretability, flexibility, and detection performance.