Monitoring and Modeling the Space Station Electrical Power System

The International Space Station is an extremely large and complicated system that is composed of many different interacting devices. Currently, staff at mission control closely monitor the state of the Space Station's electrical system to ensure that enough power is available for safe operation and to meet mission objectives. However, the complexity of the electrical power system means that it is very difficult to understand, and human observers require several years of training to interpret the telemetry signals used for monitoring.

Nearly every device in the electrical power system is instrumented with sensors that measure physical quantities such as current, voltage, temperature, and device state. In total, there are about 50,000 sensors that can generate observations every 10 seconds. This results in a massive telemetry stream and clearly, human monitors can only watch a fraction of these signals and thus require computational aids to help detect anomalies in this complex system.

With collaborators at NASA Ames Research Center, we are developing methods to help humans deal with this information overload by providing means for filtering the time course data to find and present interesting anomalies. Our basic approach is to develop models of the power system at multiple levels of abstraction and then use these models to predict the behavior of the system over time. We can compare the model's predictions with telemetry observations to detect significant differences that may indicate a fault in a device.

We have developed a graphical monitoring environment that simultaneously displays models of the electrical power system and detected faults. The models are written in a quantitative process based simulation language that lets us represent physical processes and their effects on variables, include conditions on when each process will be active, and support both instantaneous and differential equations. For monitoring complex devices, we have extended the language to support subsystems and hierarchical models at multiple levels of abstraction.
 

A Monitoring Environment

Given models of Space Station components and a telemetry stream, our environment can detect potential faults and raise alerts to a human user. Figure 1 shows a screenshot of our monitoring enviroment and displays a hierarchical model of a single power channel in the electrical system. A power channel is an electrically isolated sub-system that is composed of three components representing power generation, storage, and consumption. In the figure, PowerGen represents the devices that are involved with power generation such as the solar array wings, electronic control units, and beta gimbal assemblies. PowerStorage represents the battery assemblies that store solar energy and release power when needed. Finally, SecondaryPower represents the devices that consume power on the station and the network needed to distribute this power.

The systems and variables in the display are color coded to indicate anomalies. Yellow indicates a substantial deviation that may eventually lead to severe problems. Red indicates a serious deviation that should be examined immediately. In Figure 1, the variable Power is yellow indicating that observed power levels differ substantially from the models prediction. Our environment can plot both the observed and predicted values for a variable, as shown in Figure 2.
 

Figure 1. The monitoring environment displaying a hierarchical model of a power channel.

 

Figure 2. Total power transfer in and out of the PowerStorage subsystem.

 

In Figure 1, the red border around PowerStore indicates that there is a serious problem within that system. Our models are hierarchical and can be displayed at varying levels of detail. For example, Figure 3 shows an expanded view of PowerStore that indicates the problem lies in battery assembly 1. Further expansion would reveal a problem with the battery's state of charge.
 

Figure 3. PowerStore decomposed into three battery assemblies ba1, ba2, and ba3.

 

Modeling Space Station Components

To effectively detect anomalies, our monitoring environment requires that we have accurate models of components in order to predict their behavior. However, developing accurate models of Space Station devices can be extremely difficult and time consuming for several reasons. First, many devices are only partially understood and an engineer may have a limited understanding of how different environmental conditions affect behavior. Second, the Space Station is an environment that is very different from a ground based system. Third, devices age and change over time so the models must be periodically updated to maintain accuracy. Finally, these issues coupled with the fact that there are many thousands of devices on the station means that engineers need computational aids for modeling.

To address this problem, we are developing computational tools to assist engineers in rapidly developing accurate models. We take a hybrid approach and combine knowledge from the engineer in the form of initial or partial models, observed data on the device's behavior, and search techniques from the field of equation discovery to find revised models that better explain the system.

For example, consider modeling the batteries on the Space Station. Batteries are complex electrochemical devices that are only partially understood, and engineers have models that explain some, but not all, of a battery's behavior. Figure 4 shows a simple model of a battery that an engineer might propose to represent the batteries on the Space Station. The battery is represented as an electric circuit with ideal components. The term Vcb represents the ideal voltage of the battery; Rs and Rp are resistors that represent losses during charging/discharging (Rs) and self discharge (Rp).

To a first approximation, one can model Vcb, Rs, and Rp as constants. However, in reality these terms are actually complex functions of other variables and a constant approximation may not lead to predictions that are accurate enough for our monitoring task. Generally, the engineer will be aware of these types of model deficiencies even if she does not know exactly how they should be corrected. In this particular case, the engineer may know that a more accurate model might have Vcb as a function of the batteries state of charge and possibly temperature, although the exact parameteric form might be unknown.

Our tools can take an initial model provided by an engineer and knowledge of that model's deficiencies, and automatically revise it by considering many different model changes and selecting those that best match observational data.
 
Figure 4. Initial model of a battery.

 

We applied our approach to modeling the batteries on the Space Station and we used our models to predict the voltage at the battery terminals. Figure 5a shows the actual battery voltage observed on the Space Station through telemetry. The signal to the left of the vertical line was used for developing a revised model and parameter fitting. The remainder of the signal was used for testing. Figure 5b shows the errors of our initial and revised models on the test data. The final error of our revised model is much lower and has a mean absolute error of about one volt. This result is competitive with hand constructed models of Space Station batteries.
 
(a)(b)
Figure 5. Voltage at the battery terminals. (a) Telemetry readings from the Space Station. (b) Error predicting voltage at the battery terminals on test data. The red line corresponds to a simple initial model. The blue corresponds to an automatically revised model.

 

Directions for Future Work

Although we have made much progress, there are many challenges in both monitoring and modeling of the Space Station that we still need to address. Currently we are investigating the following research questions:

Acknowledgements

This work has been supported by NASA grant NCC 2-1220, administered by the Human Centered Computing Program. Researchers involved in the effort include Stephen Bay, Pat Langley, Mei Wang Marker, Javier Sanchez, and Dan Shapiro.

Related Publications

Sanchez, J. N., & Langley, P. (2003). An interactive environment for scientific model construction. Proceedings of the Second International Conference on Knowledge Capture (pp. 138-145). Sanibel Island, FL: ACM Press.

Langley, P., George, D., Bay, S., & Saito, K. (2003). Robust induction of process models from time-series data. Proceedings of the Twentieth International Conference on Machine Learning (pp. 432-439). Washington, DC: AAAI Press.

Bay, S. D., Shapiro, D. G., & Langley, P. (2002). Revising engineering models: Combining computational discovery with knowledge. Proceedings of the Thirteenth European Conference on Machine Learning. Helsinki, Finland.


For more information, send electronic mail to sbay@apres.stanford.edu


© 2002 Institute for the Study of Learning and Expertise. All rights reserved.