This capstone project explores whether a convolutional neural network can recover the physical parameters of a second-order dynamical system - damping ratio and natural frequency - directly from a photograph of its step response curve. Rather than training a single network to predict parameters end-to-end, the project uses a structured three-stage pipeline that mirrors how a control engineer would reason about the problem.
Motivation
System identification traditionally requires clean numerical time-series data and domain expertise to select an appropriate model structure. This project investigates whether visual inspection of a step response - something engineers do informally all the time - can be formalized into an automated pipeline that is more interpretable than a direct regression network.
Technical Approach
The pipeline has three stages. First, a classifier CNN determines the damping regime (underdamped, critically damped, overdamped) so that downstream stages can apply regime-appropriate geometry. Second, a regression network estimates geometric observables from the image: overshoot percentage, peak time, and settling time. Third, closed-form physics equations convert those observables into the underlying system parameters. Grad-CAM visualizations confirm the networks attend to the correct regions of the response curve.
Results
The hierarchical structure yields more interpretable failure modes than a direct end-to-end network: errors can be attributed to misclassification, geometric estimation error, or noise in the image rather than being entangled in a single opaque regression head. The pipeline generalizes well to step responses rendered with realistic axis noise and label occlusion.