Architec(ng Digital Twins for Model-Centric Engineering: Seman(c and - - PowerPoint PPT Presentation

SSRR 2019 November 19, 2019 1

Architec(ng Digital Twins for Model-Centric Engineering: Seman(c and Machine Learning Approach

Sponsor: OUSD(R&E) | CCDC

By

• Dr. Mark Aus(n (PI), Maria Coelho, Dr. Mark Blackburn

11th Annual SERC Sponsor Research Review November 19, 2019 FHI 360 CONFERENCE CENTER 1825 Connec(cut Avenue NW, 8th Floor Washington, DC 20009 www.sercuarc.org

SSRR 2019 November 19, 2019 2

Incubator Project

Basic Idea: Explore design of digital twin architectures that support AI and ML formalisms working side-by-side as a team, providing complementary and supportive roles in collection of data, identification of events, and automated decision making. Research Challenge: How to design digital twin elements and their interactions so that collectively they can support a wide variety of systems engineering methods and processes? Incubator Goals: Understand the range of possibilities for which machine learning

• f large-scale graphs and their attributes support activities in model-centric

engineering.

Learn Structure and Sequence Military Drone (Physical) System actions Digital Twin (Cyber) Semantic Modeling Machine Learning Knowledge Representation

DRONE OPERATING SYSTEM

Reasoning data Remember Identify Objects, Events

SSRR 2019 November 19, 2019 3

Architecture for Mul(-Domain Seman(c Modeling

Multi−domain Semantic Modeling

Data Source A Data Source B Ontology A Ontology B Rules A Rules B import import visit visit Data Models / Sources Domain−Specific Rules Domain design flow Ontology classes and properties design flow

Domain B Domain A 1 Events !!!

Semantic Graph Attach Rules Engine Revisions to semantic graph import import

Executable Processing of Events

Step 1: Data-Ontology-Rule Footing (Work at UMD / NIST / SERC in 2017).

SSRR 2019 November 19, 2019 4

Mul(-Domain Seman(c Modeling

Weather.rules Weather model Occupant model Environment Environment Occupant.owl Weather.owl Rules Domain Occupant.rules and Properties Ontology Classes Engineering Building model Sensor model Equipment model FDD model Engineering Building.owl Sensor.owl Equipment.owl FDD.owl Engineering Building.rules Sensor.rules Equipment.rules FDD.rules Sources of Data (XML data files) visit Framework for Executable Processing of Events load Semantic Graphs load load Reasoner graph transformation ........... Spatial.rules Spatial.owl load Environment Meta−Domain Ontology & Rules design flow design flow Framework for Concurrent Data−Driven Development of Domain Models, Ontologies and Rules Domain Data Models and

Example: Detection and Diagnostic Analysis of Faults in HVAC Equipment.

SSRR 2019 November 19, 2019 5

Mul(-Domain Seman(c Modeling

F10 F2

Rule 01: Occupant location. Rule 02: Occupant expected comfort. Rule 03: Occupant current comfort. Rule 04: A fault has occurred. Rule 05: Valve is shut.

F19

Rule 07: AHU failed.

F17 F18

Rule 06: Coil failed. Rule 08: Evidence 4 is true and, hence, Hypothesis 3 is valid. Equipment Domain FDD Domain

F6 F3 F1 F5 F8 F11 F12 F14 F13 F15 F16 F20 F4

Domain Building Domain Occupant

F9 F7 F4

AND AND AND AND AND AND AND AND

SSRR 2019 November 19, 2019 6

Combined Seman(cs + Data Mining

2

Data Source A Data Source B Ontology A Ontology B Rules A Rules B import import visit visit Data Models / Sources Domain−Specific Rules Domain design flow Ontology classes and properties design flow

Multi−domain Semantic Modeling

• ntologies

refinement of refinement of rules

Machine Learning / Data Mining

Classification Clustering Association decision tree Group A Group B Group A implies Group B

• ntologies

domain data domain

Domain B Domain A Semantic Feature Engineering 1

Step 2: Work at UMD / Building Energy Group at NIST / NCI, 2018-2019 Research Question: How can semantic modeling + machine learning / data mining work together as a team?

SSRR 2019 November 19, 2019 7

Incubator Project

relationships. Data Source A Data Source B Ontology A Ontology B Rules A Rules B import import visit visit Data Models / Sources Domain−Specific Rules Domain design flow Ontology classes and properties design flow

Multi−domain Semantic Modeling

• ntologies

refinement of refinement of rules

Machine Learning / Data Mining

Classification Clustering Association decision tree Group A Group B Group A implies Group B

• ntologies

domain data domain

Domain B Domain A Semantic Feature Engineering 1 2 3

4 1 5 2 8 Weighted Directed Undirected

Teaching Machines to Understand Graphs

Predictions: graph nodes and labels, dependency

Observation: A lot of model-centric engineering boils down to representation

• f systems as graphs and sequences of

graph transformations punctuated by decision making and work / actions. Hence: Explore opportunities for teaching machines to understand graphs.

Step 3: Focus on Machine Learning of Graphs and Model-Centric Engineering.

Autoencoder encoder decoder minimize loss [ −2.0, 0.5, 1.0 ..... −0.5 ] representation vector lower−dimensional decompress Input: System graph topology and attributes Output: reconstruction of system graph compress

SSRR 2019 November 19, 2019 8

Incubator Research Ques(ons

• What types of graphs (e.g., undirected, directed, weighted, multi-graph) are

easy for the ML to learn?

• How well do these techniques work with graph topology and attributes that are

dynamic?

• What can the ML do that is outside the capability of semantic modeling? And

vice-versa?

• How can the ML improve the semantic modeling? And vice-versa?
• How to design the red arrows connecting layers 1, 2 and 3?
• How to represent and reason with uncertainties?
• How does the difficulty of these challenges increase with graph size?
• How to map AI-ML capability to state-of-the-art engineering views?

Contact Information Mark Austin (PI): austin@isr.umd.edu Maria Coelho: mecoelho@terpmail.umd.edu Mark Blackburn: mblackbu@stevens.edu

Semantic Modeling Machine Learning

.........

Statechart View: Sequences of Digital Twin Control Actions ... Digital Twin (Cyber)