Page 1
LOCKHEED MARTIN PROPRIETARY INFORMATION
James Waltner
Data Scientist / Analyst
Gaithersburg, MD, May 2019
Predictive Asset Management and Applications to Manufacturing
NIST STANDARDS REQUIREMENTS WORKSHOP
With contributions from: Greg Kacprzynski, Mike Koelemay, Sam Friedman, John Labarga,
Matt Trudeau, Hari Khanal, et al
©Copyright 2019 Lockheed Martin Corporation.
Rotary & Mission System (RMS)Analytics, Prognostics & Health Management, and Artificial Intelligence (APAI) Innovations
Page 2
2
©Copyright 2019 Lockheed Martin Corporation.
LM produces some of the most sophisticated equipment on the planet…• Planes
• Helicopters
• Satellites
• UAVs
• Rockets
• Ships
• Radars
• Lasers
• Fusion reactors
• and more..
Page 3
3
©Copyright 2019 Lockheed Martin Corporation.
…with some very important and complex missions
• Humanitarian Assistance & Disaster Relief
• Global materiel transport
• Human Mission to Mars
• Even tracking space junk!
Page 4
4
©Copyright 2019 Lockheed Martin Corporation.
…so we collect lots of data (variety & volume) to help us sustain these fleets.
• Platform Operational data (e.g. Health & Usage Monitoring Systems (HUMS)
• Flight Tests
• Supply chain & logistics
• Maintenance
• Safety
• Operator meta-data
• Engineering
Page 5
5
©Copyright 2019 Lockheed Martin Corporation.
Equipment Prognostics & Health Management (PHM)
Page 6
6
©Copyright 2019 Lockheed Martin Corporation.©Copyright 2018 Lockheed Martin Corporation.
6
Downtime AvoidanceHUMS trend monitoring
Level-of-Repair OptimizationOpportunity to reduce removals
Extending Time-on-WingLeveraged HUMS and maintenance data to identify candidates for TBO extension
RepairabilityDeveloped repairs to reduce scrap rates
Reducing Costs and Improving Reliability
Page 7
7
©Copyright 2019 Lockheed Martin Corporation.©Copyright 2018 Lockheed Martin Corporation.
7
Which one is not like
the others?
Are there patterns of behavior?
What insight can we generate about
operations ?
What happened and has it occurred before?
Heat map of differences in
operator reported vs. calculated Flight Hours
Ad Hoc Analyses
Page 8
8
©Copyright 2019 Lockheed Martin Corporation.©Copyright 2018 Lockheed Martin Corporation.
8
Rotor & Propulsion SystemRotor State - Load & Motion Sensors
Virtual Monitoring Loads
Rotor Component Health Monitoring
Rotor System Diagnostic Reasoner
Blade Impact Detection & Characterization
Automated Rotor Track & Balance
Drive SystemLoads Monitoring
Adv. Dynamic Load-Based Diagnostics
Drive System Diagnostic Reasoner
Oil condition & debris monitoring
Electric Power & WiringSmart solid state power
Distribution components
Wire fault detection & isolation
LRU diagnostic reasoner
AirframeGW/CG monitoring
Global/Local Loads & Impact Monitoring
Environment Monitoring & Risk Assessment
Fatigue, Corrosion, Impact/Battle Damage Detection
Structural Integrity Reasoner
EngineAuto Power Assurance
LRU Fault Diagnostics
Propulsion System Diagnostic Reasoner
Engine Prognostics Flight Controls & HydraulicsHydraulic Leak Detection
Hydraulic Pump Diagnostics
Servo Diagnostics
Adaptive ControlsLoad limiting controls
Damage adaptive controls
Fleet ManagementUsage-Based Maintenance
Condition-Based Maintenance
Damage Tolerance
Maintenance-Free Operation Period
Total System Health Management
Page 9
9
©Copyright 2019 Lockheed Martin Corporation.
Fleet Management
©Copyright 2018 Lockheed Martin Corporation.
Page 10
LOCKHEED MARTIN PROPRIETARY INFORMATION
Dealing With The Data
©Copyright 2018 Lockheed Martin Corporation.
Page 11
11
©Copyright 2019 Lockheed Martin Corporation.
• Hundreds to thousands of raw parameters sampled continuously throughout operation
• Thousands of discrete event types
• Thousands of indicators calculated from raw data
• Thousands of usage metrics
• and more…
Platform Operational Data e.g. Health & Usage Monitoring Systems (HUMS)
Page 12
12
©Copyright 2019 Lockheed Martin Corporation.
Using all of the data…
Amount of chronological history
# o
f as
sets
day
s
year
s
1 tail
Fleet•Usage based component lifing• Fleet-based threshold calc.•Operator flight characterization• Fleet history event searches•Data driven root cause analysis• Fleet-based risk assessment
• Fleet monitoring
•operator / battalion DB• alerts, trends
• “near-asset” / ground station• alerts, trends
•Component life extensions
Page 13
13
©Copyright 2019 Lockheed Martin Corporation.
Data Problems
Store
Process
Retrieve
Ingest
Page 14
LOCKHEED MARTIN PROPRIETARY INFORMATION
Applications Leveraging GPUs
©Copyright 2018 Lockheed Martin Corporation.
Page 15
15
©Copyright 2019 Lockheed Martin Corporation.
Multivariate Timeseries Data
• Classification
• Anomaly Detection
• Signal reconstruction
• Virtual monitoring
Page 16
16
©Copyright 2019 Lockheed Martin Corporation.
Telemetry Data Classification
Top 20 Most Frequent Classes After resampling and SMOTE
• Top 5 regimes encompass 90% of data, others are very rare• Rare classes are difficult to learn
Page 17
17
©Copyright 2019 Lockheed Martin Corporation.
Parameter Inference and Virtual Monitoring
• Using deep networks to generate missing signals
Autoencoder Approach
Sequence Modeling (RNN-LSTM) Approach
Page 18
18
©Copyright 2019 Lockheed Martin Corporation.©Copyright 2018 Lockheed Martin Corporation.
18
NLP For Text Classification
FaultR/H PILOT FLOOR MIC SWITCH STUCK IN THE ON POSTION. CONSTANT HOT MIC.ActionREPLACED FLOOR MIC SWITCH.
EmbeddingModel
19A06(FOOTSWITCH)
Fault & Action text →Maintenance codeFree form text →Malfunction codeMaintenance data →Machine summaryetc
RNN
Page 19
19
©Copyright 2019 Lockheed Martin Corporation.
FRACAS Code Scoring
Fault Action
R/H PILOT FLOOR MIC SWITCH STUCK IN THE ON POSTION. CONSTANT HOT MIC.
REPLACED FLOOR MIC SWITCH.
WUC Model Confidence
Most Indicative Words
19A06(FOOTSWITCH)
96.2% • ‘MIC’• ‘FLOOR’• ‘HOT’• ‘SWITCH’• ‘STUCK’
Proportion >90% confident: 53%
FRACAS: Failure Reporting, Analysis, and Corrective Action SystemWUC: Work Unit Code
Page 20
20
©Copyright 2019 Lockheed Martin Corporation.
tSNE showing semantic separability between two WUCs in the maintenance records
FRACAS Code Scoring
Page 21
LOCKHEED MARTIN PROPRIETARY INFORMATION
Applications to the Factory
©Copyright 2018 Lockheed Martin Corporation.
Page 22
22
©Copyright 2019 Lockheed Martin Corporation.©Copyright 2018 Lockheed Martin Corporation.
22
Drive SystemLoads Monitoring
Adv. Dynamic Load-Based Diagnostics
Drive System Diagnostic Reasoner
Oil condition & debris monitoring
Electric Power & WiringSmart solid state power
Distribution components
Wire fault detection & isolation
Fleet ManagementUsage-Based Maintenance
Condition-Based Maintenance
Maintenance-Free Operation Period
Factory Sustainment
Fleet Sustainment to Factory Sustainment
Page 23
LOCKHEED MARTIN PROPRIETARY INFORMATION
Takeaways
©Copyright 2018 Lockheed Martin Corporation.
Page 24
24
©Copyright 2019 Lockheed Martin Corporation.
Frontiers must be pushed and GPUs are a critical enabling technology that has allowed us to pioneer.
Takeaways
Systems Engineering
AI/ML & Analytics
Data
Sustainment analytics requires relevant data, informed application of analytics tools and engineering expertise.
Page 25
25
©Copyright 2019 Lockheed Martin Corporation.
Takeaways
Data is Data. The challenges of maintaining Fleets of Aerospace vehicles are the same as the challenges of maintain “Fleets” of Factory Machines
Designing in high-quality contextualized data is an enabler for providing high value Predictive Asset Management Solutions.
©Copyright 2018 Lockheed Martin Corporation.