A Thesaurus-guided Method for Smart Manufacturing Diagnostics Farhad Ameri May 2019, NIST
July 9, 2019 1 Landgrebe, J., & Smith, B. (2019). Making AI meaningful again. arXiv preprint arXiv:1901.02918.
#MakingAIMeaningfulAgain July 9, 2019 2 General Limitations of Machine Learning
• Need for huge amount of data for training • Need for stable environment • Unstable against variations in input • Lack of transparency (blackbox operation mode) • Output of all stochastic models is approximative
“It is naïve to believe stochastic models alone can mimic human’s decisions without benefiting form prior knowledge” “AI applications today can simulate only a small fragment of the semantics underlying human language use”.
July 9, 2019 3 Proposed Solution
• Incorporating Prior Knowledge into AI Algorithms • Combining formal logic (semantic representation) and stochastic models in AI applications.
July 9, 2019 4 Minimal Requirements for AI language processing systems
• Exactness • Information security • Robustness • Data parsimony • Semantic fidelity • Inference • Prior knowledge usage
July 9, 2019 5 Semantic Spectrum of Knowledge Organization Systems
Ontology Conceptual Model OWL, Common Logic Taxonomy
RDFS, UML Expressivity Thesaurus RDFS Semantic interoperability Informal Syntactic/Structural Hierarchy SKOS interoperability List XML
Glossary, Dictionary, Controlled Vocabulary Development Cost
July 9, 2019 6 Thesaurus vs. Ontology
Ontology Thesaurus
Logic-based semantics Relational semantics Structural semantics Lexical semantics
Text → Logic-based Formulae Text → Concept Vector
July 9, 2019 7 Research Objectives
• To develop a minimalistic knowledge graph (thesaurus) that can represent relational and lexical semantics in the domain of industrial maintenance. • To develop models and tools for: • Text summarization and tagging • Maintenance diagnosis
July 9, 2019 8 Prior Maintenance Knowledge (Relational Semantics)
System
Failure Mode 1 Failure mode 2
Root cause 1-1 Root cause 2-1
Root cause 2-2
Treatment 1 Treatment 2
July 9, 2019 9 Gear Failure modes and Treatments
Artifact Gear
Modes of failure Scuffing
Gear Bending Fatigue Gear macropitting Gear micropitting wear Gear noise
Treatment Replace gear Gear grinding Gear Lubrication
July 9, 2019 10 One knowledge graph for the entire maintenance domain
July 9, 2019 11 Company-specific or Industry-specific Knowledge Graphs
Boeing GM BP
Integration
Aerospace Industry Automotive Industry Oil & Gas Industry
Integration Integration Integration Integration
Integration
July 9, 2019 12 SKOS model (Simple Knowledge Organization System)
SKOS is a common data model for knowledge organization systems such as thesauri, classification schemes, subject heading systems, and taxonomies. • W3C Standard Galvanization Chemical Reaction • Web-native syntax Coating
• Machine-readable SKOS:Broader Rust • Simple semantics SKOS:Related
Oxidation SKOS:AltLabel Rusting
SKOS:Narrower
SKOS:Definition Homolytic Oxidation
Heterolytic Oxidation
Oxidation is the loss of electrons during a July 9, 2019 reaction by a molecule, atom or ion. 13 Concept Schemes (types) in Maintenance Diagnosis Thesaurus
July 9, 2019 14 MD Thesaurus
(property)
July 9, 2019 15 Alternative labels
July 9, 2019 16 Relatedness
skos:related A B Part of cause of symptom of Reason for Impact of Input for Output of
July 9, 2019 17 Relatedness
skos:related When an artifact is part of another artifact or share a common parent Artifact Artifact
When certain artifacts are prone to certain problems (failures) Artifact Problem
When certain treatments are effective for certain artifacts Artifact Treatment
When problems have known treatments Problem Treatment
When Certain problems are caused by other problems (causality) Problem Problem
When certain treatments complement each other Treatment Treatment
July 9, 2019 18 Part-Problem subgraph
Undesirable
Behavior
r
e
d
a
o
r b Part Maintenance Problem related leaking Gasket
Leaking Leaking Leaking air oil water
July 9, 2019 19 Graph of “related” problems
related
related
related
related
related
July 9, 2019 20 Artifacts related to Problems
Artifacts Related Problems
July 9, 2019 21 Problems related to Treatments
Problems Related Treatment
July 9, 2019 22 How MD Thesaurus Helps ?
• Summarizing and tagging maintenance logs. • Decision support in maintenance diagnosis systems • From symptoms to root-causes to treatments • Document classification
July 9, 2019 23 Data Tagging and Summarizing
July 9, 2019 24 Maintenance Logs
July 9, 2019 25 SKOS Tool
July 9, 2019 26 Tagged Text
July 9, 2019 27 Extracted Concepts
Exported Table
July 9, 2019 28 Sufficient Tagging
Problem Description: Table will not index. Resolution: Table indexing valve coil found bad and replaced
Concept Schema Top Concept Concept (preferred label) Freq. Artifact Part coil 1 Artifact Part indexing valve 1 Functional Maintenance Problem Maintenance Problem not indexing 1 Maintenance Treatment Treatment Action replace 1 Artifact Part valve 1
July 9, 2019 29 Insufficient Tagging
Problem Description: Damaged stock tubes (x3) Resolution: New Hush Tubes ordered 9/8 -Installed 9/27 x3
Concept Schema Top Concept Concept (preferred label) Freq. Non-Functional Maintenance Problem Maintenance Problem damaged part 1 Artifact Part tube 2
July 9, 2019 30 Decision Support Tool for Maintenance Diagnosis
July 9, 2019 31 APT Graph
Artifact (System/Part/Machine)
Problem
Treatment
July 9, 2019 32 ATP Graph
Artifact (System/Part/Machine) A
Problem 1 2 3
Treatment 1 2 3 4
July 9, 2019 33 July 9, 2019 34 July 9, 2019 35 Future Directions
• Automated creation of Bayesian Network • Improving the expressivity of the knowledge graph • Expanding the graph
July 9, 2019 36 Questions ?
July 9, 2019 37