Machine learning now tracks magnetic objects with a diamond quantum sensor

Researchers at the University of Melbourne and RMIT University have developed a machine learning method that predicts the position of a magnetic target in real-time, trained solely on experimental data and bypassing the need for pre-existing physical models. The work uses a quantum diamond magnetometer to track an elevator’s movements, achieving over 80% accuracy at a rate of 10 Hz with a positional error tolerance of 30 centimeters, a precise measurement compared to the 4-meter spacing between parking levels.

This approach offers a new path for applications including ground transport monitoring, underground infrastructure mapping and hazardous detection, as the results suggest this machine learning method can be applied more generally for real-time monitoring of magnetic objects.

Machine Learning Tracks Magnetic Objects with Diamond Quantum Sensors

This approach allows for tracking within a noisy environment, a significant hurdle for conventional magnetic sensing techniques that rely on detailed system knowledge. The core of this work is the algorithm’s capacity to learn directly from sensor readings, rather than requiring a prior understanding of the magnetic field’s behavior.

The team used a quantum diamond magnetometer to track variations of a few hundred nanoteslas produced by an elevator moving along a single axis. This work represents a step towards broader applications, building on the success of tracking one-dimensional movement, which mirrors realistic environments like high buildings and train circuits.

Quantum Diamond Magnetometry for Weak Magnetic Field Detection

The quantum diamond magnetometer was used to track variations of a few hundred nanoteslas produced by the monitored elevator, despite the presence of a much stronger Earth’s magnetic field of approximately 50 microteslas. By measuring magnetic field projections along four NV axes within the diamond, any external magnetic field vector can be reconstructed. This reconstruction process relies on an inverse matrix calculation, transforming measured frequency shifts into a vector magnetic field and temperature shift, enabling the precise determination of the elevator’s position.

The experimental magnetic dataset, comprising five segments, was then used to train the machine learning algorithm without any reliance on pre-existing physical models of the elevator or its environment. This approach differs from traditional magnetic sensing methods, which often require detailed knowledge of the system under study. The algorithm’s ability to learn directly from experimental data opens up possibilities for deployment in scenarios where such models are unavailable or impractical, such as complex underground infrastructure or rapidly changing environments.

Remote Magnetic Sensing Limitations in GPS-Denied Environments

The ability to accurately pinpoint a moving magnetic source without prior knowledge of its characteristics represents a shift in remote sensing capabilities. This approach relies on experimental data alone, differing from methods often hampered by the difficulty of accurately modeling real-world magnetic environments. The system’s predictive power was validated using an elevator as a magnetic target, with the machine learning algorithm trained to track its one-dimensional movement.

This simplified scenario, resembling environments like high-rise buildings and tunnels, allowed for robust data collection and algorithm refinement. The machine learning algorithm can be trained in approximately 40 minutes, achieving over 80% accuracy in predicting the target’s position at a rate of 10 Hz, with a positional error tolerance of 30 centimeters.

The researchers emphasize that the target object can be any source of magnetic signal above the floor noise, broadening the potential applications beyond elevators to include ground transport, underground infrastructure mapping and even hazardous detection. The core of the system is a quantum diamond magnetometer, using a dense ensemble of nitrogen-vacancy defects to sense external magnetic fields. These results suggest this machine learning method can be applied more generally for real-time monitoring of magnetic objects, which will broaden the scope of magnetic detection applications.

Real-Time Elevator Tracking with Machine Learning Algorithms

This approach allows for real-time prediction of a magnetic target’s position, a capability demonstrated using an elevator as the test subject, without needing detailed prior knowledge of the elevator’s mechanics or magnetic properties. The system processes magnetic information into time windows, creating a matrix input for the algorithm where the number of magnetic components used impacts the data’s dimensionality.

Optimization of the machine learning hyperparameters involved testing multiple options for each parameter, training several algorithms and computing tracking accuracy with a one-meter tolerance, with the lowest results counting as the final score; the researchers note this reflects the progressive stages of optimization used to achieve the reported performance.

This approach relies on supervised learning, where the algorithm learns from a dataset pairing magnetic signals with corresponding elevator positions, comparing predictions to target values during training. These windows are then associated with the corresponding elevator position during the training phase.

Magnetic Signal Complexity from Elevator and External Sources

These contributing factors create a complex magnetic environment that necessitates a robust analytical approach, as the magnetometer is positioned roughly 10 meters from the elevator doors. The system’s design, detailed in Figure S3, captures variations in the magnetic field on the order of a few hundred nanoteslas when the elevator is in motion. Analysis of the collected data revealed that the magnetic signal’s correlation with the elevator’s position was most pronounced along one specific axis, while environmental noise significantly impacted the other two.

This fortunate alignment, though potentially arbitrary, allowed for easier identification of the relationship between magnetic signals and the elevator’s location; the researchers note that any other coordinate frame could have obscured this correlation. The team’s ability to identify a clear correlation between magnetic signals and elevator position occurred without any prior knowledge of the system’s magnetic behavior.

More like this

Artificial IntelligenceResearchers analysed 179 incidents to map AI memory ‘negative space’

Quantum PhysicsHelper flip suppresses offset with a 27% information cost.

Quantum Machine LearningEngineered Learning Problems Show Performance Gap

Quantum SensorsHouston Team Tunes Photon Exchange Using Light Pulses