TL;DR
This paper presents a novel Kalman filter-based approach to fuse data from accelerometers, GNSS, and rotational sensors to achieve a more accurate six-component (6C) motion tracking. This method is crucial for structural health monitoring and earthquake analysis, as it overcomes the limitations of individual sensors by integrating their strengths, particularly by accounting for rotational effects on GNSS data.
What Problem Does This Paper Address?
Monitoring the motion of structures, especially during events like earthquakes or under wind loads, requires understanding not just translational movement but also rotational motion. Traditional sensors like accelerometers and GNSS receivers only capture translational data, missing crucial rotational information. While some sensors can measure rotation, fusing these disparate data streams (with different sampling rates, noise characteristics, and inherent biases) into a coherent 6C motion estimate is a significant challenge. This paper addresses the non-trivial problem of fusing these heterogeneous sensor data to provide a more complete and accurate picture of structural dynamics.
Methods
The core of this paper is a six-component (6C) Kalman filter (KF) designed for sensor fusion. The authors employ a linear KF, which is a standard technique for state estimation. What's novel here is its application to fuse three distinct sensor types (accelerometer, GNSS, and rotational sensor) and its specific formulation to account for the influence of rotations on both accelerometer and GNSS measurements.
Key methodological aspects include:
- Pre-computation: Angular velocities are integrated to obtain rotation angles, and acceleration data is rotated into the local coordinate system using these angles.
- Rotation Correction: The filter explicitly accounts for how rotations can induce apparent translations in GNSS data due to the antenna phase center offset and how they affect accelerometer readings (gravitational leakage).
- State and Measurement Equations: The KF's state vector includes displacement and velocity, while the input vector is formulated using pre-rotated acceleration and corrected GNSS antenna positions. The GNSS positions serve as the primary observation for the filter.
- Multi-rate Fusion: The KF is designed to handle sensors with different sampling rates (GNSS at 100 Hz, IMU/accelerometer at 250 Hz) through interpolation.
Datasets
The paper utilizes data from an experimental setup designed to simulate structural motion. This setup involves:
- An industrial six-axis robot arm to generate controlled 6C motions.
- KVH 1750 IMU: A high-quality inertial measurement unit providing acceleration and angular velocity.
- Javad GrAnt G3T GNSS Antenna and Septentrio Receiver: For precise positioning.
- EpiSensor and Centaur Digitizer: For acceleration measurements.
The robot's feedback system is used as the ground truth for evaluating the performance of the fused sensor data. Three experimental sets (TLRL, TLRS, TSRS) with varying amplitude sets were performed to test the system under different motion conditions.
Key Insights and Ideas
- Rotations significantly distort GNSS data: Contrary to common assumptions, rotations can induce noticeable translational errors in GNSS positions, especially for slender structures or during seismic events. This effect is directly linked to the antenna phase center offset and the point of rotation.
- Fusion improves accuracy across the board: Combining accelerometer, GNSS, and rotational data through the KF significantly reduces the Root Mean Squared Error (RMSE) compared to using individual sensors, demonstrating the synergistic benefit of fusion.
- Rotational correction is vital for both sensors: Accounting for rotations is crucial for both accelerometer (correcting gravitational leakage) and GNSS (correcting induced translations) to achieve accurate 6C motion estimates.
- GNSS is key for long-period motion, accelerometers for short-period: The study highlights that GNSS is better suited for resolving low-frequency, long-period motions, while accelerometers excel at capturing high-frequency, short-period dynamics. The KF leverages these complementary strengths.
- Vertical component challenges: The vertical (z) component of GNSS is often too noisy to be useful for displacement measurements, requiring reliance on accelerometers for this axis.
Key Results
- The proposed KF-based fusion scheme significantly reduces the RMSE of motion estimates compared to individual sensor data. For instance, correcting for rotational effects on GNSS data in the TLRL experiment halved the RMSE.
- The system demonstrated good performance across different motion amplitudes (TLRL, TLRS, TSRS experiments), with RMSE values for translational axes generally below 0.35 mm.
- The spectral analysis shows that the KF estimates closely match the robot feedback across a broad frequency range, particularly for horizontal displacements and velocities.
- The study successfully demonstrated the ability to correct for rotationally induced errors in both accelerometer and GNSS data, leading to more precise 6C motion tracking.
Limitations and Future Work
- Robot limitations: The robot arm has limitations in generating very small amplitude and high-frequency motions, which are characteristic of distant earthquakes or stiff structures. This necessitated simulating wind turbine-like motions.
- Synchronization challenges: Achieving perfect time synchronization between all instruments was a hurdle, requiring post-processing corrections.
- GNSS noise in the vertical component: The high noise level in GNSS vertical positions limits its utility for this axis, requiring reliance on accelerometers. The paper notes that setting the GNSS vertical component to zero is a pragmatic solution but not universally applicable.
- Future work: The authors suggest extending this methodology to real-world structural monitoring applications, including more complex structures like bridges and buildings. They also propose developing more sophisticated KF formulations that can adapt to varying sensor qualities and potentially handle situations where GNSS data is insufficient for displacement estimation. Further research could explore real-time implementation and the use of more advanced sensors for broader dynamic range and higher sensitivity.
Summary updated