1 Miniature Mode Estimation Example ACS Subsystem GHe Pressure - - PDF document

SLIDE 1

1

Improving Model-based Mode Estimation Through Offline Compilation

Seung H. Chung John M. Van Eepoel Brian C. Williams Model-based Embedded and Robotic Systems Group Space Systems Laboratory Massachusetts Institute of Technology Thursday, June 21st, 2001

Outline

• Motivation
• Previous Mode Estimation Approaches
• Example ACS System
• Miniature Mode Estimation System
• Rule-based system comparison
• Conclusions

Motivation

Apollo 13 quintuple fault Mars Polar Lander failed due to a faulty sensor. Distant Explorers Mercury Orbiter

• Rule-based systems are becoming overly complex

as spacecraft systems grow.

• The model-based approach allows for a more direct

fault protection system.

• This model-based mode estimation system will

give:

• A set of inspectable rules used in diagnosis
• Faster response over previous systems.

General Diagnostic Engine

• Achieves a diagnosis through the divide and conquer approach
• Generates all diagnoses, including highly unlikely ones.

Candidate Generation Observations System Models Merge Diagnoses Partial Diagnoses Diagnoses

Sherlock / Livingstone

• Generates only most likely diagnosis that pertains to the current
• bservations through a generate and test approach.
• Problem : The test phase - SAT-SEARCH loop is expensive in time.

SAT Engine

Observations System Models Diagnosis

Conflict-directed Search

Checked Solution Conflicts Most-likely Candidate

Miniature Mode-Estimation

• Generates all diagnoses offline using dissents.

– Methods employed are resolution, enumeration and multi- resolution.

• The conflict directed search is utilized at run time to generate a

diagnosis.

System Models

Online Conflict-directed Search Offline Satisfiability Engine

Partial Diagnosis Rules Diagnosis Observations

SLIDE 2

2

Miniature Mode Estimation

• Offline satisfiability generates the partial diagnoses in two

steps.

• Online search performed by:

– Triggering the appropriate partial diagnoses – Search is directed by these conflicts

Mini-ME Offline Spacecraft Model Dissent Generator Compiled Model Online Partial Diagnosis Trigger Best-first Kernel Diagnosis Generator Conflicts Dissents Repair Manager Most Likely Diagnosis Monitors Continuous Observations Partial Diagnosis Rule Generator Partial Diagnosis Rules Discrete Observations

Example ACS Subsystem

• Pressure Transducer (S) Model

– Modes – {G, SH, SL, U}* – G(S) ⇒ ((TP = nom) ⇔ (P1 = nom)) ∧ ((TP = low) ⇔ (P1 = low)) – SH(S) ⇒ (P1 = nom) – SL(S) ⇒ (P1 = low) – U(S) ⇒ ( )

• Solenoid Valve (V) Model

– Modes – {O, C, U}* – O(V) ⇒ ((P1 = nom) ⇔ (P2 = nom)) ∧ ((P1 = low) ⇔ (P2 = low)) – C(V) ⇒ (P2 = low) – U(V) ⇒ ( )

• Catalyst Bed (C) Model

– Modes – {G, B, U}* – G(C) ⇒ ((P2 = nom) ⇔ (T = on)) ∧ ((P2 = low) ⇔ (T = off)) – B(C) ⇒ (T = off) – U(C) ⇒ ( )

• Observable Variables

– Pressure in Pipe 1 (P1) – {nom, low} – Engine Thrust (T) – {on, off}

• Unobservable Variables

– Tank Pressure (TP)– {nom, low} – Pressure in Pipe 2 (P2) – {nom, low} N2H4 GHe

P S

Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2

Hydrazine Thruster Inertial Sensor

* All modes have an associated probability.

Offline Satisfiability

• The satisfiability step is moved to an offline computation and dissents

are generated.

– Dissents link infeasible component modes to observations.

• (P1 = low) ∧ (T = on) ⇒ ¬( G(S) ∧ O(V) ∧ G(C) )
• Dissents are then transformed into partial diagnosis rules by making

an equivalent statement as:

– (P1 = low) ∧ (T = on) ⇒ ( SH(S) ∨ SL(S) ∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C) ) Offline Satisfiability

Spacecraft Model Dissent Generator Compiled Model Dissents Partial Diagnosis Rule Generator Partial Diagnosis Rules Online Conflict-directed Search Most Likely Diagnosis Observations

Online Conflict-directed Search

• Monitors discretize continuous data in two ways.

– Value : Sensor Voltage = 23 V Sensor Voltage = nominal – Time

• Partial diagnosis trigger
• (P1 = nom) ∧ (T = off) ⇒ SH(S) ∨ SL(S)

∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C)

• (P1 = nom) ⇒ G(S) ∨ SH(S) ∨ U(S)
• (P1 = low) ⇒ G(S) ∨ SL(S) ∨ U(S)
• …

G(S) ∨ SH(S) ∨ U(S)

P1 = nom Online Conflict-directed Search

Partial Diagnosis Trigger Best-first Kernel Diagnosis Generator Conflicts Repair Manager Monitors Continuous Observations M

• s

t L i k e l y D i a g n

• s

i s

Offline Satisfiability

Partial Diagnosis Rules Discrete Observations

Online Conflict-directed Search

• Best-first Kernel Diagnosis Generator then determines the set of

assignments that best resolves the conflicts.

G(S) SH(S) U(S) 0.865 0.017 0.002

• Repair Manager uses the

diagnosis and gives the appropriate action to the spacecraft to resolve failure.

Online Conflict-directed Search

Partial Diagnosis Trigger Best-first Kernel Diagnosis Generator Conflicts Repair Manager Monitors Continuous Observations Most Likely Diagnosis

Offline Satisfiability

Partial Diagnosis Rules Discrete Observations

Sample Diagnosis

N2H4 GHe

P S Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2 Hydrazine Thruster Inertial Sensor

• Observations
• P1 = low
• T = on

SLIDE 3

3

Sample Diagnosis

• Triggered Partial Diagnoses

– G(C) ∨ U(C) – O(V) ∨ U(V) – G(S) ∨ SL(S) ∨ U(S) – SH(S) ∨ SL(S) ∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C) N2H4 GHe

P S Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2 Hydrazine Thruster Inertial Sensor

• Observations
• P1 = low
• T = on
• Most-likely Diagnosis

U(C) G(C) 0.865 0.0009

Sample Diagnosis

• Triggered Partial Diagnoses

– G(C) ∨ U(C) – O(V) ∨ U(V) – G(S) ∨ SL(S) ∨ U(S) – SH(S) ∨ SL(S) ∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C) N2H4 GHe

P S Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2 Hydrazine Thruster Inertial Sensor

• Observations
• P1 = low
• T = on
• Most-likely Diagnosis

U(C) G(C) 0.0009 0.0846 O(V) U(V) 0.0009 0.865

Sample Diagnosis

• Triggered Partial Diagnoses

– G(C) ∨ U(C) – O(V) ∨ U(V) – G(S) ∨ SL(S) ∨ U(S) – SH(S) ∨ SL(S) ∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C) N2H4 GHe

P S Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2 Hydrazine Thruster Inertial Sensor

• Observations
• P1 = low
• T = on
• Most-likely Diagnosis

U(C) G(C) 0.0009 0.0846 O(V) U(V) 0.0009 0.865 0.0846 G(S) U(S) S L ( S ) 0.0002 0.0017

Sample Diagnosis

• Triggered Partial Diagnoses

– G(C) ∨ U(C) – O(V) ∨ U(V) – G(S) ∨ SL(S) ∨ U(S) – SH(S) ∨ SL(S) ∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C) N2H4 GHe

P S Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2 Hydrazine Thruster Inertial Sensor

• Observations
• P1 = low
• T = on
• Most-likely Diagnosis

U(C) G(C) 0.0009 0.0846 O(V) U(V) 0.0009 0.865 U(S) U(V) C ( V ) SL(S) B(C) U(C) SH(S) 0.0846 G(S) U(S) S L ( S ) 0.0002 0.0017

Sample Diagnosis

• Triggered Partial Diagnoses

– G(C) ∨ U(C) – O(V) ∨ U(V) – G(S) ∨ SL(S) ∨ U(S) – SH(S) ∨ SL(S) ∨ U(S) ∨ C(V) ∨ U(V) ∨ B(C) ∨ U(C) N2H4 GHe

P S Pressure Transducer Solenoid Valve Catalyst Bed Pipe 1 Pipe 2 Hydrazine Thruster Inertial Sensor

• Observations
• P1 = low
• T = on
• Most-likely Diagnosis
• Full Diagnosis : SL(S) O(V) G(C)

U(C) G(C) 0.0009 0.0846 O(V) U(V) 0.0009 0.865 U(S) U(V) C ( V ) SL(S) B(C) U(C) SH(S) 0.0846 G(S) U(S) S L ( S ) 0.0002 0.0017

Mini-ME vs. Rule-Engine

• Mini-ME makes explicit the deductive step of determining

component modes.

– Models allow mapping from symptoms to system states to recovery actions easier to specify.

Mini-ME Repair Manager

Mode Estimate

Spacecraft

Observations Action Rule Engine Spacecraft Observations Action

SLIDE 4

4

Application to Spacecraft

• Sample NEAR rule

– Symptom :

• (Charger current > 0.8 A) for

10 sec – Recovery :

• Switch to the redundant

charger and disengage the primary.

• Equivalent Mini-ME diagnosis

– (Charger-current = high) ⇒ (Primary-charger = broken) ∨ (Primary-charger = unknown) – Recovery action then specified based only on the mode.

Summary

• Conclusions

– Builds upon prior model-based mode estimation heritage. – The system is guaranteed to find a diagnosis. – Coverage of all rules in a rule-based system is possible. – Time performance is in line with a rule-based system

• Benefits

– Rule set is inspectable for correctness – More intuitive set of rules – Development of models is easier for engineers – Real-time performance

Mini-ME Offline Spacecraft Model Dissent Generator Compiled Model Online Partial Diagnosis Trigger Best-first Kernel Diagnosis Generator Conflicts Dissents Repair Manager Most Likely Diagnosis Monitors Continuous Observations Partial Diagnosis Rule Generator Partial Diagnosis Rules Discrete Observations