Comparison of CNN and MLP in Magnetic Positioning Applications

Magnetic positioning, a noncontact method with broad application potential in industry, medical robotics, and autonomous navigation, is limited by environmental interference, sensor noise, and high computational complexity. This study compares Convolutional Neural Networks (CNN) and Multi-Layer Perceptrons (MLP) for high-precision 3D magnetic field-based positioning. Datasets (simulating a 4×4 magnetic sensor array with Gaussian noise) were used to train and test both models: MLP processed 48-dimensional vectors, while CNN utilized 4×4×3 tensors to preserve spatial relationships. Evaluation metrics included Root Mean Square Error (RMSE) for position and average angle error for direction. Experimental results show CNN outperforms MLP: lower position RMSE (0.0062 m vs. 0.0063 m) and smaller angle error (5.6307° vs. 5.9658°), with stronger robustness to noise. These findings confirm CNN's suitability for magnetic positioning (due to superior local spatial feature extraction) and provide insights for optimizing neural network architectures in such systems.