Main problems and approach

The Meteosat experience

After the launch of Meteosat-1 there were reported a number of unexplained anomalies. For the launch of Meteosat-2 there was built a spacecraft environment monitor (SEM-1) that would measure the electron flux in the 30 eV to 20 keV energy range. The results from this showed that the electron flux in this energy range did not correlate well with the occurrence of the anomalies (Johnstone et al., 1985). Therefore on Meteosat-3 a second electron monitor (SEM-2) was developed to look at 43 keV to 300 keV electrons. The flux of electrons in this energy range was shown to be correlated with the radiometer anomalies, and that anomalies associated with high flux were seen in the morning, while those associated with low flux in the afternoon (Rodgers, 1991). Both these studies relates the anomalies to the local electron environment as measured with on-board instruments. These studies lead to considerable improvement of the design of Meteosat spacecraft which then experienced much less anomalies.

Andersson et al. (1998) analysed the local plasma environment as measured by SEM that cause Meteosat-3 spacecraft anomalies with the help of Neural network techniques. The data base consists of 724 anomalies over the 7 year life time of the satellite, i.e. on average 2 anomalies per week. From this, a model for the prediction of anomalies from local environment data was suggested and the best input parameters were identified. It was shown that 53% of the anomalies were predicted, when at the same time 84% of the non-anomalies were predicted. The predicted anomalies were associated with times of high electron flux.

The use of geospace environment information in anomaly forecasting

A change of the spacecraft local environment is often related to a change on a broader scale in the magnetosphere. Therefore can information on the environment as derived from other spacecraft or geomagnetic observations be useful. Lopez Honrubia and Hilgers (1997) developed a new technique based on pattern classification methods for the prediction of spacecraft anomalies. They used a set of 40 anomalies that occurred on Meteosat 3, 4, and 5 and the space environment was characterized by the electron flux with energy above 2 MeV measured on the GOES satellites. Although limited by the small amount of data, the study demonstrated that a correlation could be established between the MeV electron environment as measured on GOES and the anomaly experienced by other spacecraft. Furthermore, the technique could be used for making anomaly prediction one day in advance or for extracting characteristic patterns of the space environment preceding anomaly occurrence.

Another study was made to relate the geomagnetic indices Kp and Dst to spacecraft anomalies (Wu et al., 1998). A data base of anomalies was created and several neural networks were trained to make daily predictions of  anomalies. Two geostationary satellites were included in the data base: Satellite-1 (1988-1995) and Satellite-2 (1989-1998). The data base consists of reported anomalies on the two satellites, energetic electron flux (E>2MeV) from the GOES satellites, and planetary indices Kp and Dst. All data were averaged into daily values. Several different neural networks were trained by using different combinations of input parameters and varying the time delay line from 1 day to 10 days. The best parameter for daily predictions of these satellite anomalies was found to be Kp. The prediction accuracy was 80% in this case. This suggests that for the type of anomalies investigated plasma processes in an energy  range lower than MeV are significant.

Artificial neural networks and fuzzy systems

As proved in many other applications artificial neural networks (ANNs) and fuzzy systems are powerful models for the prediction and analysis of complex non-linear systems. When large amounts of data are available the artificial neural networks can be trained and tested with a high confidence. The number of parameters and the flexibility of the data base are important factors to consider before the networks are trained. The data set should be divided into three separate sets: training set, validation set, and test set. The free parameters of the network are found through training on the training set, and then the optimal network architecture is determined from the validation set. Finally the performance of the network is calculated from the test set.

When some knowledge about the system exists or when data is more sparse the fuzzy systems can be developed with a higher reliability than the ANNs. Existing expert knowledge is coded in to the fuzzy system, which then can be further trained on the data. E.g. from Rodgers (1991) it was shown that there was a high probability for anomalies in the morning sector when the electron flux was high. The fuzzy rule could thus be: If high electron flux and morning sector then anomaly. This looks similar to an expert systems, however, there are important differences. In fuzzy logic it is allowed with partial contradictions, whereas in expert systems the logic is strict. The fuzzy system can also be trained on the data as the rules are coded into functions and numerical values. In this example the measured electron flux can belong to some degree to the class `high electron flux' and at the same time to the class `low electron flux'. The rules are converted into numbers that are weighted together and the output are the classes `anomaly' and `non-anomaly' to different degrees.

The artificial neural networks and fuzzy systems can then be converted into neuro fuzzy systems that can be trained on the data to further improve the model.