Abstract
A fast method to simultaneously calibrate multiple MEMS Magnetic Inertial Measurement Units (MIMUs) accurately in the field is needed in many application areas. The MEMS MIMUs require calibration of systematic errors of bias, sensitivity, non-orthogonality and misalignment, which vary with temperature and use. Even after calibration, the sensors undergo stochastic errors in static and dynamic conditions and thus uncertainty of output must also be modeled. We propose a method for easy and fast calibration of multiple MIMUs together, while mounted on a single platform. The precise alignment of sensors is not assumed. Our method calibrates both fixed array of MEMS MIMUs or many independent MIMUs simultaneously using kinematic constraints. The novelty of our approach is that the uncertainty of sensors output is also learned as part of our model. Compared with existing state-of-art methods, our algorithm gives more consistent readings of all MIMUs and our framework also predicts the associated uncertainty of the sensor output. The uncertainty prediction of individual sensors is particularly helpful in the sensor fusion.
Introduction
In the world of robotics, wearables, and navigation, we rely heavily on MIMUs (Magnetic Inertial Measurement Units). These tiny chips—containing accelerometers, gyroscopes, and magnetometers—tell a device how it’s moving and where it’s pointing. However, MEMS-based sensors are notoriously finicky; they drift, they’re sensitive to temperature, and they rarely come perfectly calibrated out of the box.
A new framework proposes a more elegant solution: a fast, simultaneous calibration method that doesn't just fix errors, but also tells you exactly how much to trust the data. By using a two-tier deep learning framework, this approach learns to predict sensor uncertainty in real-time, significantly improving the consistency and reliability of sensor arrays used in motion capture and navigation.
Methodology
The researchers developed a two-tier calibration model that combines classical physics with modern deep learning. The framework simultaneously calibrates multiple MIMUs mounted on a single rigid platform using mutual kinematic constraints.
The process involves four main stages of implementation:
- Data Collection (I): Sensors are mounted on a rigid platform (such as a 20-position icosahedron) and moved through specific protocols to capture a wide range of motion data.
- Weak Model Initialization (II): Simple, fast protocols (like a quick free-fall for the accelerometer and a level rotation for the gyro) create basic linear calibration models. These "weak learners" are used to initialize a neural network (an LSTM), allowing it to converge much faster than with random weights.
- Kinematic Constraints (III): Because all sensors are mounted on the same rigid platform, they must experience the same magnitude of rotation and acceleration at any given moment. The model uses these physical "rules" as a cost function to fine-tune the calibration.
- Uncertainty Learning: The network is trained using a Negative Log Likelihood (NLL) loss function. This allows the model to output both a calibrated value and a variance ($\sigma$), representing the aleatoric uncertainty (noise inherent in the data).
Key Insights and Ideas
- Magnitude is Key: You don't need to know the exact 3D alignment of every sensor to calibrate them. By focusing on the norm (magnitude) of the acceleration and rotation vectors, you can force multiple sensors into agreement using simple rigid-body physics.
- Physics-Informed Initialization: By seeding the neural network with the parameters of a physical linear model, the authors bypassed the need for the massive datasets usually required by deep learning.
- Dynamic vs. Static Uncertainty: Sensor uncertainty isn't a flat line; it spikes during high-dynamic movements. By learning this relationship, the system provides a much better input for downstream "sensor fusion" algorithms like Kalman Filters.
Key Results
The proposed method consistently outperformed state-of-the-art field calibration techniques:
- Accuracy: On synthetic data, the model reduced the Mean Absolute Error (MAE) for gyroscopes significantly compared to standard level-rotation methods.
- Consistency: In real-world tests with seven sensors, the Mean Absolute Deviation (MAD)—a measure of how much the sensors disagreed with each other—dropped from 3.3 to 0.3 for gyroscopes in dynamic conditions.
- Uncertainty Correlation: The predicted uncertainty was shown to correlate perfectly with actual error spikes during movement, providing a reliable "trust metric" for the sensor output.
Conclusion
While the method is fast (taking only 2-3 minutes of data collection), it does require the sensors to be mounted on a rigid platform during the calibration phase. Future work involves integrating this "uncertainty-aware" data directly into orientation estimation filters. If a filter knows exactly when a sensor is becoming unreliable, it can temporarily "ignore" that sensor, leading to much more robust navigation in challenging environments.
Summary updated