Synthetic Data Generation of Surgical Drills Using Physics-Constrained GAN

Abstract

Class imbalance in time series data occurs when some classes have far fewer training samples than others. Training a neural network classifier on such an imbalanced dataset results in bias and poor performance. In this paper, we focus on the class imbalance issue in a surgical bone-drilling dataset, where the drill traverses certain bone regions faster than others, leading to far fewer recorded samples in those regions. We propose a physics-constrained Generative Adversarial Network (P-GAN) that generates synthetic time-series data for underrepresented classes. Unlike standard generative models, the proposed P-GAN enforces physics constraints during the generation of synthetic samples. These constraints include ensuring depth and force values are within sensor-valid ranges and that depth sequences progress monotonically. Furthermore, these constraints enforce that the synthetic samples behave like real drill measurements. We evaluate the proposed method on the drill dataset by adding 150 synthetic sequences per underrepresented class and measure the effect on the model’s performance. Results on a drill dataset show consistent improvements in F1 scores for both GRU and CNN-based classifiers, suggesting that physical constraints are essential for generating useful synthetic data on small, domain-specific datasets.