Sample Acquisition, Processing and Handling Systems for Future Mars Missions
- Paper number
IAC-04-Q.3.B.09
- Author
Mr. Paul Fulford, MDA, Canada
- Year
2004
- Abstract
This paper proposes a new method to monitor a bias instability of an onboard attitude rate sensor, Fiber Optical Gyro, using a support vector machine (SVM) technique as one of autonomy application to small satellite systems. To validate the concept of the method, this paper shows a result using real down-linked telemetry data of a small science satellite. The satellite, called REIMEI, was launched into a near-sun-syncronous orbit in 2005 and has been operated for over two years [1]. It has a three-axis bias-momentum attitude control system consisting of one momentum wheel, three magnetic torques, one star tracker, one magnetometer, and three fiber optical gyros.
The attitude determination system of REIMEI depends mainly on the star tracker whose nominal accuracy is 3 arc minutes. If the star tracker is not available, the fiber optical gyros take the role of the principal sensor to know the current satellite attitude by propagation. Two situations that the star tracker is available and not available come one after another since the earth comes into the field of view of the star tracker when the satellite attitude is controlled to be fixed in a inertial coordinate system. The duration times for the two situations are 67 minutes for the star tracker available period and 30 minutes for not available period, respectively. The accuracy of the satellite attitude estimation depends only on the fiber optical gyro data during the 30 minutes periods, which means a bias stability of the fiber optical gyro is important to maintain the accuracy. The bias is estimated by an onboard Kalman filter. These fiber optical gyros were well-calibrated and their thermal environment has been controlled stable by heaters. How precise the bias has been estimated can be evaluated by the attitude error data that are obtained just after the star tracker has turned available. The attitude error at this moment when the star tracker becomes available can be regarded as equal to the accumulation of propagation error caused by the bias estimation error.
The bias estimation error can be divided into two parts, the Kalman filter performance and the bias instability. The bias was modeled as a time variant and it is possible to tune any parameters of the Kalman filter using both simulators and onboard software. Like other systems, these parameters were carefully tuned using a flight software simulator and actual flight telemetry data and the result was good enough to have been operated without any problems until about one year and half passed from the launch. Since then, there has been some minor problem in the bias estimation. The most probable reason is that a fiber transparency degradation by radiation became large enough to cause the bias instability. But, however, there must be still some margin against such degradation according to the results of radiation test that were performed before launch. Re-tuning the parameters has been tried continuously in the same fashion before, but it has not been succeeded yet, in practice. The model of the bias seems to have changed. In case that, re-tuning the parameters has little effects than ever. As an practical solution, the following process has been performed: select two tuning parameters and send commands to changed them until an acceptable result is obtained from the telemetry data. If the parameters become unsuitable, repeat this process again.
Trying to search for temporal optimal parameters in this way is so heuristic that its result is rather unstable compared with that were obtained before the bias model had changed. In order to keep the bias estimation error smaller, it is necessary to monitor some variables that are related to the attitude estimation error and to notice any sign of the bias change before the attitude error goes out of the range of requirements. Support vector machine (SVM) is one of the most practical methods to monitor such signs. It is a popular discrimination scheme used in many fields, and the author also have been trying to apply it to actuator failure detection problems especially for onboard flight software of small satellite systems [2]. Thus, this paper focuses on a monitor of the fiber optical gyro bias instability and shows several results using both flight telemetry data of REIMEI and its onboard software simulator.
[Methodology and Result] The supervisory learning process of SVM was performed as follows: in Step 1, prepare a telemetry data set that was obtained before the bias model had changed. This data set will be treated as a no-error case data set. In Step 2, prepare a telemetry data set that contains some signs indicating the parameters are unsuitable. This data set will be treated as a error case data set. In Step 3, transform these data sets into other data sets expressed in a feature space to make mathematical calculation as simple as possible. Using these data sets as input, support vectors are calculated. The kernel function was a Gaussian Radial Basis in this paper. What kind of variable transformations to adopt and what kernel should be used will make crucial difference in efficiency and reliability of classification, which will be described in details also in this paper. In Step 4, using the support vectors, an on-line telemetry data set will be discriminated whether the current parameters have been tuned suitable or not.
A no-error case data set was selected from the telemetry data of one week which had been down-linked in 2005, and unsuitable parameter case data sets from 2007 telemetry of several weeks. The support vectors have been used to monitor the bias errors for several weeks down-linked data with successful results, in those data, It was found that there were some kinds of errors caused by not only the bias instability but by an alignment of the fiber optical gyros and some misidentified star tracker data, that had not been included in the supervisory training data sets.
[Conclusion] A SVM was designed to monitor the instability of the fiber optical gyros by finding any sign implying off-tuned parameters were in the rate estimation filter. Moreover, the SVM suggested some causes that have not been given in the supervised learning phase. However this discriminator was implemented on a telemetry monitor system on ground segment, the classification part of the SVM logic can be migrated on flight software segment without any large efforts, which means that these kinds of state monitor can be a standard and one of key software components of autonomy not only for telemetry data analysis tools but for onboard software even in small satellite systems.
[Reference]
[1] Fukushima, Y., Sakai, S., and Saito, K., "Flight Performance of the REIMEI Microsatellite Attitude Determination System," Proceedings of the Small Satellite Systems and Services, ESA-SP-625, Sardignia, Italy, 2006.
[2] Fukushima, Y., "Onboard Sensor and Actuator Failure Detection using SVM for Autonomy of Small Satellite Systems," Proceedings of the 9th International Symposium on Artificial Intelligence and Robotic Automation in Space: iSairas, Los Angels, CA, 2008.
- Abstract document
- Manuscript document
IAC-04-Q.3.B.09.pdf (🔒 authorized access only).
To get the manuscript, please contact IAF Secretariat.