Real-time localization method for magnetic capsule robot using transfer learning in physics-informed neural network

Magnetic manipulation provides a promising approach to enhancing the maneuverability and safety of capsule robots (CR). Due to its advantages of human harmlessness and low latency, magnetic signal-based positioning technology is regarded as the most promising CR localization method. However, traditional magnetic localization approaches based on optimization algorithms have initial values sensitivity and error accumulation issues especially in environments with multiple magnetic sources. To address this challenge, a neural network based on magnetic positioning is proposed, which integrates physics-based principles with transfer learning. The method builds a ResNet-based prediction model with multi-scale convolution and channel attention mechanism to bolster system resilience. A loss function based on physical information and designed with a magnetic field analysis model is provided, which enhances the interpretability of model inference. A transfer training strategy involving pre-training with physical model data and fine-tuning with real data is presented, which achieves high positioning accuracy even when data is limited. The efficacy of the localization methodology is validated in an ex vivo pig stomach, resulting in an average position error and angle error of 2.90 mm and 1.26°, respectively, with a single positioning time of 0.34 ms. The results demonstrate the promising potential of this technology for translation into clinical practice.