Learning rules for temporal fault diagnosis in satellites

Contents and links

Introduction to temporal fault diagnosis

We have applied ILP to a problem within the aerospace industry, namely the diagnosis of power-supply failures in a communications satellite.

The satellite recharges its batteries using solar energy. As the it orbits the Earth, its position relative to the Sun changes, taking its power-supply subsystem through four distinct stages: battery charging, solstice, eclipse, and battery reconditioning. Using qualitative modelling, we constructed a simulation with which the behaviour of the power supply in each stage could be predicted. By provoking simulated faults in the components, we used the simulation to generate examples of how each fault would affect the supply's behaviour. From these examples, the ILP program Golem induced a set of rules for diagnosing power supply failures. In generating the examples, faults were provoked in all possible components, thus guaranteeing that the rules are complete and correct for all single faults.

Because the power-supply's behaviour changes with time, the formalism used to describe the examples must be able to express a fault's dependence on time. We have developed such a formalism, which is based on temporal logic and which is suitable for ILP learning.

Qualitative modelling

Using qualitative models to learn diagnostic rules was first demonstrated by Kardio [Bratko I., Mozetic I. and Lavrac N.(1989)]. The basic idea is simple. Use a qualitative model of a system to generate behaviours of the system. If there is a fault in the system, it will be reflected in the behaviour generated by the model. This ``faulty'' behaviour can be used to provide examples for a learning program, which can then learn diagnostic rules for the particular fault. The resulting rules are thus guaranteed complete and correct for all single faults since all examples of these were generated from the original model. This technique is applicable to problems other than the original one considered in Kardio. An example is the power subsystem of a satellite. The basic structure of one such system is shown below:

The arrows indicate the flow of supply. The qualitative model of this system consists of about 40 components and 29 sensors [Pearce D.A.(1988)].

An important feature of this system is that the diagnosis of some faults is not possible with simple classification rules. These faults usually require the history of related components. For example, a simple check of power from the battery is not sufficient to determine if it is faulty. We need to know the history of other components in the circuit before this can be decided. Although many different faults can be simulated, the concern here is with battery faults. The information used is as follows:

Examples are are recorded in time points. So, a positive example fault(32) states that there was a fault at time instance 32.
Background knowledge records the status of the 29 sensors at each time point. The list of sensors is attached.
During orbit, the basic operation of the power subsystem is divided into phases. The time instant that each phase commences is also provided as background knowledge.
The succession of time instants is represented by the predicate succ/2.

The sensors in the power subsystem are as follows:

tm040_switch          : system switch status indicator.
tm018_switch          : system switch status indicator.
tm031_switch          : system switch status indicator.
m038_switch           : system switch status indicator.
tm022_switch          : system switch status indicator.
tm043_switch          : system switch status indicator.
tm013_switch          : system switch status indicator.
tm042_switch          : system switch status indicator.
tm007_switch          : system switch status indicator.
tm222_charging        : battery charging status indicator.
tm071_asr_or_switch_20: system switch status indicator.
tm070_supply_3c       : power supply from solar.
tm058_asr_or_switch_10: system switch status indicator.
tm057_supply_2c       : power supply from solar.
tm257_battery_voltage : system battery status indicator.
tm017_switch          : system switch status indicator.
tm009_switch          : system switch status indicator.
tm220_supply_1c       : power supply from solar.
tm055_supply_1b       : power supply from solar.
tm054_supply_1a       : power supply from solar.
tm029_ovt_disabled    : battery over-temperature blocking signal.
tm021_eoc_disabled    : battery end-of-charging indicator.
tm004_eoc_signaled    : battery end-of-charging blocking  signal.
tm002_battov_temp     : battery over-temperature sensor.
tm039_eod_disabled    : battery end-of-discharging blocking signal.
tm015_eod_signaled    : battery end-of-discharging indicator.
tm011_eod_override    : battery end-of-discharging overwriting signal.
tm001_eod_relay       : battery end-of-discharging relay switch.
tm211_bus_voltage     : system power bus status.

The Golem dataset

The dataset is stored as one compressed TAR file. Within that, the data files are as used in the original Golem experiments. That is, background knowledge files have a ``.b'' suffix, positive example files have a ``.f'' suffix, and negative example files have a ``.n'' suffix.

Bibliography

Feng, C.(1991).
Inducing Temporal Fault Diagnostic Rules from a Qualitative Model
In Proceedings Eighth International Workshop on Machine Learning ,
pp 403 - 406, Morgan Kaufmann, San Mateo, C.A..

Bratko I., Mozetic I. and Lavrac N. (1989).
Kardio: a Study in Deep and Qualitative Knowledge for Expert Systems.
MIT Press, Cambridge Mass.

Pearce D.A. (1988).
The induction of fault diagnosis systems from qualitative models.
In Proceedings Seventh National Conference on Artificial Intelligence,
Saint Paul, Minnesota.

Up to applications main page.

Machine Learning Group Home Page