NASA Logo, National Aeronautics and Space Administration

Case Study

Electrical power systems (EPSs) play a critical role in spacecraft and aircraft. EPS loads in an aerospace vehicle include the following crucial subsystems: avionics, propulsion, life support, and thermal management systems. Loss of electrical power to any of these subsystems could lead to serious consequences for personnel or for the vehicle.

There are, from the point of view of vehicle health management, several technical challenges associated with electrical power systems. First, electrical power systems often have a large number of distinct modes due to mode-inducing components such as relays, circuit breakers, and loads. Second, while much EPS behavior is deterministic, there is both sensor noise and system state uncertainty in EPSs. Sensor noise is due to the imperfections of sensing, while system state uncertainty is due to failures of EPS components and sensors. These two technical challenges have been two of our main concerns. Our use of Bayesian networks and arithmetic circuits, rather than other approaches to technical diagnosis, is motivated by the need to construct EPS diagnostic models that capture both deterministic and uncertain behavior when many modes are present.

The Advanced Diagnostic and Prognostic Testbed (ADAPT) is an electrical power system testbed developed at the NASA Ames Research Center. ADAPT provides: (i) a standard testbed for evaluating diagnostic algorithms and software; (ii) a capability for controlled insertion of faults, giving repeatable failure scenarios; and (iii) a mechanism for maturing and transitioning diagnostic technologies onto manned and unmanned vehicles. Here, we present our development of a diagnostic capability for ADAPT, using probabilistic techniques.

The EPS functions of ADAPT are as follows:

  • For power generation, ADAPT currently uses utility power.
  • For power storage, ADAPT contains 3 sets of 24 VDC 100 Amp-hr sealed lead acid batteries.
  • Power distribution is aided by electromechanical relays and two load banks with ac and dc outputs; there are also several circuit breakers.
  • ADAPT loads include pumps, fans, and light bulbs.
There are sensors of several types, specifically for measuring voltage, current, relay position, temperature, light, and liquid flow. Control and monitoring of ADAPT takes place through programmable automation controllers. With the sensors included, ADAPT contains a few hundred components and is representative of EPSs used in aerospace.

As a simple example reflecting ADAPT, we consider the following BN with five nodes that represent an EPS relay along with its feedback mechanism.

pca approach v3 portable

Our Bayesian diagnostic process has as input sensor readings for sensor nodes and observed commands for command nodes, and as output query nodes that provide the health status of sensors and EPS components.

In this example BN, let us consider health nodes (or output) Health_Relay1 and Health_Feedback1 and suppose that evidence (or input) is as follows: Command_Relay1 = cmdClose, Sensor_Feedback1 = readClosed. Using computation of marginals, as illustrated in the figure above, we obtain most likely values (MLVs) Health_Relay1 = healthy, and Health_Feedback1 = healthy. In other words, given a command to close Relay1, and a confirming feedback message from Feedback1, it is inferred that both the relay and the feedback mechanism are healthy (as is to be expected).

First Gov logo
NASA Logo -