Presentation on theme: "1 Relational Data Mining Applied to Virtual Engineering of Product Designs Monika Žáková 1, Filip Železný 1, Javier A. Garcia-Sedano 2, Cyril Masia Tissot."— Presentation transcript:
1 Relational Data Mining Applied to Virtual Engineering of Product Designs Monika Žáková 1, Filip Železný 1, Javier A. Garcia-Sedano 2, Cyril Masia Tissot 2 and Nada Lavrač 3,4 1 Department of Cybernetics, CTU Prague, 2 Semantic Systems, Derio, Spain, 3 Jozef Stefan Institute, Ljubljana, Slovenia 4 University of Nova Gorica, Nova Gorica, Slovenia
ILP 2006 2 / 17 Outline 1. Motivation 2. Semantic Virtual Engineering 3. Annotation of CAD designs 4. Challenges for ILP 5. Our approach 6. Preliminary results 7. Conclusions and future work
ILP 2006 3 / 17 Motivation Engineering is one of the most knowledge-intensive activities Knowledge in form of CAD designs, documents, simulation models and ERP data bases Goal: Making implicit knowledge contained in CAD designs explicit useful for reuse, training, quality control No industrial software employing ILP techniques in real-life regular use we are aware of
ILP 2006 4 / 17 Project More specific motivation: SEVENPRO: Semantic Virtual Engineering for Product Design project IST-027473(2006-2008) funded under 6th Framework Programme of the European Commission.
ILP 2006 7 / 17 Design Annotation the information available in CAD files and other data sources formalized and integrated by means of semantic annotation based on ontologies Semantic annotation of CAD designs generated automatically from the commands history available via the API of CAD tools based on a CAD ontology developed in SEVENPRO available in RDF format annotation including ontology of CAD items and axioms defining core relations automatically translated into Prolog
ILP 2006 9 / 17 Challenges to ILP There are three main challenges for ILP due to ontolgies in the background knowledge: hierarchies of term sorts induced by subclassOf relation hierarchies of relations induced by subpropertyOf relation representation conversion between Prolog and other knowledge representation languages (SWRL)
ILP 2006 10 / 17 Our Baseline Approach Our Baseline Approach based on sorted refinement operator (Frisch 1999) sorted subsumption relation combines θ-subsumption with taxonomies on terms Tasks: Currently: Propositionalization Finding maximal patterns Clustering of designs Other: Classification Requirement/design matching Outlier detection
ILP 2006 12 / 17 Propositionalization propositionalized representation of classified relational data generated by constructing first-order features during the feature generation a table of mutual feature subsumptions maintained this subsumption is exploited in propositional search pruning any conjunctions of subsumer with its subsumee specializing a conjunction not only by extending it, but also by replacing an included feature with its subsumee.
ILP 2006 13 / 17 Finding maximal patterns for non-classified data maximal patterns emerging patterns of some limited length covering the minimum set amount of examples can be used for: Discovering repetitive patterns Finding typical ways some type of item is designed Creating templates that can be reused
ILP 2006 14 / 17 Preliminary Results Preliminary Results the system tested on a set of 35 CAD designs one design ~ 100 predicates Language bias imposed based on maximum depth and max. number of relations with the same input variable the dataset
ILP 2006 15 / 17 Extracted Features Examples of extracted features f(X1:cADFileRevision) = hasCADEntity(X1:cADFileRevision,X2:cADPart), hasBody(X2:cADPart,X3:body),hasFeature(X3:body,X4:extrude),…, hasFeature(X3:body,X7:extrude),hasFeature(X3:body,X8:pocket),... hasFeature(X3:body,X12:pocket),hasFeature(X3:body,X13:fillet),... hasFeature(X3:body, X16:fillet), hasFeature(X3:body,X17:cADFeature). f(X1:cADFileRevision) = hasCADEntity(X1:cADFileRevision,X2:cADPart), hasBody(X2:cADPart,X3:body),hasFeature(X3:body,X4:extrude),…, hasFeature(X3:body,X7:extrude),hasSketch(X7:extrude,X8:circular Sketch),hasGeomElement(X8:circularSketch,X9:circle).
ILP 2006 16 / 17 Future work Include taxonomy on predicates Improve efficiency using graph search techniques For closer integration of more complex hierarchical background knowledge the following approaches considered Integration of subsumption operator with proven properties Use of hybrid languages AL-log, CARIN Use of more complex representational formalism ψ -terms, antecedent description grammars