High Accuracy Passive Magnetic Field-Based Localization for Feedback Control Using Principal Component Analysis

Shaohui Foong, Zhenglong Sun

TL;DR

This paper introduces a novel magnetic field-based localization system that uses a network of sensors and Principal Component Analysis (PCA) to accurately track the position of a moving object. By reducing the dimensionality of sensor data with PCA before feeding it into an Artificial Neural Network (ANN), the system achieves high precision and computational efficiency, rivaling expensive optical encoders for feedback control.

What Problem Does This Paper Address?

Precise position measurement is crucial for many automated systems, especially for controlling linear actuators over large distances. While magnetic field-based sensing offers advantages in harsh environments (e.g., immunity to temperature, radiation, and no line-of-sight requirement), a major challenge is establishing a reliable and unique correspondence between magnetic field measurements and actual position. This is due to the inherent non-linearity and non-uniqueness of magnetic field models. Furthermore, processing data from multiple sensors for real-time feedback control can be computationally intensive, especially when using complex Artificial Neural Networks (ANNs). This paper aims to overcome these limitations by developing a robust, accurate, and computationally efficient magnetic localization system.

Methods

The core of the proposed system lies in a two-pronged approach:

  1. Multi-Sensor Network for Bijectivity: A spatial network of magnetic sensors is employed to capture magnetic flux density (MFD) at multiple locations simultaneously. By strategically arranging these sensors, the system aims to create a unique relationship between the combined sensor readings and the object's position, overcoming the non-uniqueness issues of individual sensors.
  2. PCA-Optimized ANN Mapping:
    • Principal Component Analysis (PCA): Due to the spatial proximity of sensors, their readings are often correlated. PCA is used as a "pseudo-filter" to decorrelate these measurements and reduce their dimensionality. It identifies the principal components (eigenvectors) that capture the most variance in the data, effectively creating a more compact and informative representation. This significantly reduces the number of inputs required for the ANN.
  3. Artificial Neural Networks (ANNs): A feedforward ANN with a single hidden layer is trained using backpropagation to map the PCA-transformed sensor data to the object's precise position. This model-free approach allows for accurate mapping of the complex, non-linear relationship between magnetic fields and position.

The paper also details numerical simulations to investigate the impact of geometric parameters (PM dimensions, sensor spacing) and Gaussian noise on the system's accuracy.

Datasets

The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, it relies on:

  • Simulated Data: Numerical simulations were used to generate magnetic field data under various conditions, including different geometric configurations and the addition of artificial Gaussian noise. This simulated data was used to train and evaluate the PCA-ANN model.
  • Experimental Data: A custom experimental setup was built using a linear actuator, a permanent magnet (PM), and a network of nine Hall-effect sensors. The system's performance was evaluated by recording magnetic field measurements as the PM moved along a linear track. This experimental data was used to validate the simulation results and compare the proposed system against a high-resolution optical encoder.

Key Insights and Ideas

  • PCA as a Dimensionality Reduction and Noise Filter: PCA is not just for reducing computational load; it also acts as an effective noise filter. By projecting the data onto principal components, the system becomes more resilient to sensor noise, as demonstrated in the simulations.
  • Bijectivity through Sensor Network Design: The spatial arrangement of sensors is critical. Even with individual sensors exhibiting non-bijective behavior, a carefully designed network can collectively provide a unique mapping to position.
  • Sub-micron Accuracy Achieved with Cost-Effective Sensors: The system demonstrates that high-precision localization (comparable to expensive optical encoders) can be achieved using relatively inexpensive Hall-effect sensors, making it a more accessible solution.
  • Insensitivity to Separation Distance: PCA-assisted mapping shows a remarkable insensitivity to the separation distance between the sensors and the magnet, a significant advantage for practical deployment.
  • Computational Efficiency for Real-Time Control: By reducing the number of ANN inputs via PCA, the system achieves faster computation, crucial for real-time feedback control loops.

Key Results

  • High Accuracy: The proposed magnetic sensing system, using a 9-sensor network and PCA-optimized ANN, achieved micron-level accuracy.
  • Comparative Performance: In closed-loop tracking experiments, the system demonstrated performance comparable to a 1 µm resolution optical encoder. The average tracking error was 0.074 mm for the magnetic system versus 0.085 mm for the optical encoder.
  • Computational Advantage: PCA significantly reduces the number of inputs to the ANN. For instance, using only the first three principal components was sufficient for accurate mapping, compared to using all six raw sensor inputs in a multi-sensor configuration. This leads to a substantial reduction in computational operations.
  • Noise Resilience: In the presence of Gaussian noise (1% and 10%), the RMSE of the ANN using PCA inputs was an order of magnitude lower than that of an ANN using raw sensor measurements.

Limitations and Future Work

  • Saturation Effects: At very small separation distances between sensors and the PM, sensor saturation can occur, leading to reduced SNR and potentially impacting accuracy. This is noted as a limitation, particularly for the closest sensor configurations.
  • Geometric Parameter Dependence: While PCA helps mitigate some issues, the system's performance is still influenced by geometric parameters (sensor spacing, PM dimensions). The paper explores these trade-offs but doesn't eliminate their impact entirely.
  • Future Work: The paper suggests that PCA-transformed outputs and ANNs can be realized with simple analog summing circuits, hinting at the potential for hardware implementation of these complex algorithms, even for large-scale sensor networks. Further exploration into optimizing the number of principal components and ANN architecture for specific applications could be a future direction.

Summary updated

Abstract

In this paper, a novel magnetic field-based sensing system employing statistically optimized concurrent multiple sensor outputs for precise field-position association and localization is presented. This method capitalizes on the independence between simultaneous spatial field measurements at multiple locations to induce unique correspondences between field and position. This single-source-multi-sensor configuration is able to achieve accurate and precise localization and tracking of translational motion without contact over large travel distances for feedback control. Principal component analysis (PCA) is used as a pseudo-linear filter to optimally reduce the dimensions of the multi-sensor output space for computationally efficient field-position mapping with artificial neural networks (ANNs). Numerical simulations are employed to investigate the effects of geometric parameters and Gaussian noise corruption on PCA assisted ANN mapping performance. Using a 9-sensor network, the sensing accuracy and closed-loop tracking performance of the proposed optimal field-based sensing system is experimentally evaluated on a linear actuator with a significantly more expensive optical encoder as a comparison.
BibTeX
@inproceedings{foong2016high,
  author = {Shaohui Foong and Zhenglong Sun},
  title = {{High Accuracy Passive Magnetic Field-Based Localization for Feedback Control Using Principal Component Analysis}},
  booktitle = {Sensors},
  year = {2016},
  url = {https://openreview.net/forum?id=FGMCKKZksz},
}
Figure