Adaptation Approach for Inertial Measurement Unit Model Parameters
Arunabh Chattopadhyay, Vibhor L. Bageshwar, and Srivatsan Varadarajan, Honeywell Aerospace, Honeywell International
Location: Big Sur
Alternate Number 1
Vehicle Navigation systems are typically designed using a sensor set; kinematic models, measurement models, and sensor measurement models (SMMs); and a filter that fuses the sensor measurements to estimate the vehicle kinematic state vector statistics relative to one or more selected navigation frames. The sensor set typically includes one or more inertial measurements units (IMUs) and aiding sensors such as GPS, magnetometers, odometers, and, in certain applications, sensors or signals such as vision based sensors or signals-of-opportunity. The filter relies on the selected models to capture the vehicle kinematic motion, to capture the relationship between the sensor measurements and the vehicle kinematics, and to capture the effects of sensor measurement errors on the sensor measurements. The vehicle kinematic state vector is typically augmented by a sensor specific state vector to enable the filter to estimate sensor specific statistics to more accurately estimate the vehicle kinematic statistics. The filter takes advantage of the complimentary statistics of the sensors to estimate the sensor specific state vector statistics.
The SMMs govern the time evolution of the sensor specific state vector. These SMMs include both stochastic and deterministic components; these components are designed using parameters that quantify the grade of the sensor and the effect of these components on the sensor measurements. For example, an IMU measurement model can include models of turn-on bias, in-run bias, scale factor, axes misalignment, and measurement noise. The parameters for these models are selected from calibration testing and sensor specification sheets. The navigation filter relies on the proper selection of these parameters. Improper selection of these parameters can cause the filer to converge to a biased state mean vector with an inconsistent state covariance matrix. In other words, the filter can converge to an incorrect vehicle kinematic state vector (estimate an incorrect navigation solution) with a state covariance matrix that is much too small indicating that the filter has over-confidence in its incorrect estimate of the vehicle kinematic state vector.
Typically, low-grade MEMS IMUs are not calibrated across a wide temperature range and a large vibration environment. Further, these low-grade MEMS IMUs are not subject to rigorous lifetime testing. As a result, the parameters used in the IMU measurement models may not reflect actual sensor performance in operation due to calibration errors, temperature variation, vibration environment, and length of service. The results of these incorrect parameters and model mismatch on the navigation system are that, first, the navigation system will not be able to satisfy its performance requirements and, second, the IMU will need to be replaced more frequently. These issues with incorrect IMU model parameters can be mitigated if an adaptation system can be implemented that can, first, determine whether the parameters have varied from their nominal values, second, determine the amount of parameter variation and, third, adapt to the parameter variation within the filter design.
In this paper, we propose and develop an adaption approach called the sensor off- and on-line calibration and monitoring (SOLCAM) system that can refine the IMU measurement model to increase navigation system robustness to changing IMU sensor performance and improve navigation performance. The SOLCAM system is built on an algorithm designed to generate, monitor, and update an SMM in off-line (factory calibration) and on-line (real-time) applications. The algorithm compares the power spectral densities (PSDs) of two residuals generated from reference measurements, sensor measurements, and SMMs. The geodesic distance between the PSDs is a metric that indicates whether the SMM accurately captures the actual sensor performance. The algorithm calibrates, verifies, and updates the SMM by morphing one PSD toward the other PSD. The morphed PSD that minimizes the geodesic distance between PSDs is the optimal SMM that best approximates the actual sensor performance.
Our development of the SOLCAM adaptation approach involves the following steps. First, we will select an IMU measurement model for low-grade MEMS IMUs. Second, we will develop geodesic distance thresholds that indicate whether the IMU measurement model parameters accurately capture the actual sensor performance in simulation. These geodesic distance thresholds can be considered sensor integrity metrics and will help determine whether the IMU measurement model requires adaptation. Third, we will implement the SOLCAM system in conjunction a navigation system to adapt the IMU measurement model parameters to the correct values. Fourth, we will demonstrate the performance of the SOLCAM system in conjunction with a navigation system in simulation using both singular and multiple simultaneous parameter variations of the IMU measurement model and different vehicle maneuvers. These parameter variations will be informed using the same bounds as those used in the ongoing DARPA Building Resource Adaptive Software Systems (BRASS) Program.