Ph. D. Project
Title:
Knowledge discovery and formalisation for Additive manufacturing through Artificial Intelligence and Information
Retrieval methods
Dates:
2020/10/01 - 2023/09/30
Supervisor(s):
Other supervisor(s):
Dr. Yan LU (yan.lu@nist.gov)
, Dr. Albert T. Jones (albert.jones@nist.gov)
Description:
The engineering of inter-operating systems is based on different types and levels of abstraction and models. The
proposed technological and scientific context focuses on several systems that need to interoperate. The AM
processes have multiple characteristics that need to be controlled and optimized and that are in relation between
them. These systems have to be modelled and the results must express not only the "structural" aspect of the
components of the system but also their behaviour. One of the core problems of this work is to study model
interoperability problems by cooperative model-driven systems engineering (David, 2005).
The scientific challenge is thus to provide languages and modelling tools adapted to each part of the modelling
project, despite the data heterogeneity and the variety of the process-es. This challenge has two dimensions: on
the one hand, the capacity of modelling to equip the processes, through the definition and the formalization of
their invariants; On the other hand, the study of the conditions of use of models in practice, always evolutive and
uncertain.
The Formal Concepts Analysis (FCA) (Ganter, 2004) is a useful and powerful tool to formally describe the links
between any objects. This couple forms a formal context. This method is based on the lattice theory (Wille, 2009),
which can be used to solve problems of extracting tacit knowledge from formalized systems. An extension of the
FCA mechanisms has been introduced in (Rouane, 2013) and called Relational Concept Analysis (RCA), where the
focus is on data sets compatible with the Relative Entity Model (ER) or, alternatively, with the RDF (Resource
Description Framework). This is a method for extracting conceptual knowledge from multi-relational data. This
kind of approach inscribes itself in the Multi Relational Data Mining (MRDM) domain.
The RCA method is not limited to the extraction of knowledge of separate contexts: it aims to express knowledge
by inter-operating the semantics of different contexts, that is to say that in addition to extracting the knowledge
of a context, the data contained in the other contexts are used in order to enrich the extraction of knowledge.
Scientific proposal:
Our proposed methodology is based on applying FCA and ML algorithms to predict the impact of AM process
parameters on the eventual quality of the manufactured AM part. Such a prediction would give the actionable
information about the factors that come into play while producing an AM part. That actionable information will
be based on a variety of in-process sensor data that must be fused before running the ML algorithm/tool or the
RCA method. By using ML algorithms, we would train data coming from AM process sensors to create that
actionable information. The input sensor data could be in different formats such as images, metrics, 3D models.
Other input information would include material properties, current process parameters, and scan patterns among
others.
This study will explore various types of algorithms including descriptive, diagnostic, predictive, and prescriptive.
Based on the input type, RCA methods and advanced ML tech-niques such as Artificial Neural Network (especially
Deep Neural Network and Convolutional Neural Networks) would be applied to analyse input information and
extract association rules. These association rules would be set as knowledge and represented in various formats.
This output would be stored in a knowledge base so we can keep track of any changes to existing knowledge. This
knowledge base could be then queried through a tool created to get prediction and make decisions while creating
parts using AM processes.
The AM input data is characterized by multi-modal (CAD model, material property, pro-cess control, and
monitoring data), multi-rate, and multi-scale. It is vital that the input data is understandable by the algorithms
that will extract the AM rules. A methodology would be developed in order to first represent the data, then fuse it
when necessary, and finally integrate it depending on the input type. Methodologies would also be developed to
represent the extracted rules into knowledge.
Mathematics and physics-based models would be used as a foundation for managing AM data for the use of AI
algorithms. To do so, the following outputs would be made:
● Mathematical representation of the different type of data
● Methodology to represent fuse and integrate AM data for the use of AI algorithms and RCA methods
● Tools and techniques using AI/RCA to extract rules from AM data
● Methodology to transform the extracted rules into knowledge
● Methodology to represent, formalize and integrate knowledge data
● Tools with physics-based AM models, advancing the prediction capability
● AI solutions enabling decision making for AM parts based on the predictions
Objectifs :
L'objectif de cette recherche serait d'identifier, de formaliser, de développer et de valider des méthodologies
basées sur l'IA pour découvrir et formaliser les connaissances basées sur les données de l'AM afin de faire des
prédictions lors de la création de la pièce. Les prédictions permettront d'améliorer les décisions de fabrication, de
produire des pièces de haute qualité, et aussi de minimiser les coûts de production. La méthodologie proposée
pour l'amélioration de la qualité des produits AM guidera les industries AM dans la direction de la prise de
décision et de l'adoption de l'AM.
proposed technological and scientific context focuses on several systems that need to interoperate. The AM
processes have multiple characteristics that need to be controlled and optimized and that are in relation between
them. These systems have to be modelled and the results must express not only the "structural" aspect of the
components of the system but also their behaviour. One of the core problems of this work is to study model
interoperability problems by cooperative model-driven systems engineering (David, 2005).
The scientific challenge is thus to provide languages and modelling tools adapted to each part of the modelling
project, despite the data heterogeneity and the variety of the process-es. This challenge has two dimensions: on
the one hand, the capacity of modelling to equip the processes, through the definition and the formalization of
their invariants; On the other hand, the study of the conditions of use of models in practice, always evolutive and
uncertain.
The Formal Concepts Analysis (FCA) (Ganter, 2004) is a useful and powerful tool to formally describe the links
between any objects. This couple forms a formal context. This method is based on the lattice theory (Wille, 2009),
which can be used to solve problems of extracting tacit knowledge from formalized systems. An extension of the
FCA mechanisms has been introduced in (Rouane, 2013) and called Relational Concept Analysis (RCA), where the
focus is on data sets compatible with the Relative Entity Model (ER) or, alternatively, with the RDF (Resource
Description Framework). This is a method for extracting conceptual knowledge from multi-relational data. This
kind of approach inscribes itself in the Multi Relational Data Mining (MRDM) domain.
The RCA method is not limited to the extraction of knowledge of separate contexts: it aims to express knowledge
by inter-operating the semantics of different contexts, that is to say that in addition to extracting the knowledge
of a context, the data contained in the other contexts are used in order to enrich the extraction of knowledge.
Scientific proposal:
Our proposed methodology is based on applying FCA and ML algorithms to predict the impact of AM process
parameters on the eventual quality of the manufactured AM part. Such a prediction would give the actionable
information about the factors that come into play while producing an AM part. That actionable information will
be based on a variety of in-process sensor data that must be fused before running the ML algorithm/tool or the
RCA method. By using ML algorithms, we would train data coming from AM process sensors to create that
actionable information. The input sensor data could be in different formats such as images, metrics, 3D models.
Other input information would include material properties, current process parameters, and scan patterns among
others.
This study will explore various types of algorithms including descriptive, diagnostic, predictive, and prescriptive.
Based on the input type, RCA methods and advanced ML tech-niques such as Artificial Neural Network (especially
Deep Neural Network and Convolutional Neural Networks) would be applied to analyse input information and
extract association rules. These association rules would be set as knowledge and represented in various formats.
This output would be stored in a knowledge base so we can keep track of any changes to existing knowledge. This
knowledge base could be then queried through a tool created to get prediction and make decisions while creating
parts using AM processes.
The AM input data is characterized by multi-modal (CAD model, material property, pro-cess control, and
monitoring data), multi-rate, and multi-scale. It is vital that the input data is understandable by the algorithms
that will extract the AM rules. A methodology would be developed in order to first represent the data, then fuse it
when necessary, and finally integrate it depending on the input type. Methodologies would also be developed to
represent the extracted rules into knowledge.
Mathematics and physics-based models would be used as a foundation for managing AM data for the use of AI
algorithms. To do so, the following outputs would be made:
● Mathematical representation of the different type of data
● Methodology to represent fuse and integrate AM data for the use of AI algorithms and RCA methods
● Tools and techniques using AI/RCA to extract rules from AM data
● Methodology to transform the extracted rules into knowledge
● Methodology to represent, formalize and integrate knowledge data
● Tools with physics-based AM models, advancing the prediction capability
● AI solutions enabling decision making for AM parts based on the predictions
Objectifs :
L'objectif de cette recherche serait d'identifier, de formaliser, de développer et de valider des méthodologies
basées sur l'IA pour découvrir et formaliser les connaissances basées sur les données de l'AM afin de faire des
prédictions lors de la création de la pièce. Les prédictions permettront d'améliorer les décisions de fabrication, de
produire des pièces de haute qualité, et aussi de minimiser les coûts de production. La méthodologie proposée
pour l'amélioration de la qualité des produits AM guidera les industries AM dans la direction de la prise de
décision et de l'adoption de l'AM.
Keywords:
Artificial Intelligence, Machine Learning, Knowledge, Formal Concept Analysis
Conditions:
3 years, NIST, USA and France (CRAN), 34000 euros per year, Computer Engineer who has a very good knowledge
of Machine Learning methods and good mathematical skills.
of Machine Learning methods and good mathematical skills.
Department(s):
| Modeling and Control of Industrial Systems |
Funds:
4000 US$ each month for three years payed by the NIST.
Publications:
