The CEA List develops AI methodologies based on Deep-Learning to extend the implementation of usual approaches to complex industrial applications (little/no data, potentially of poor quality). A significant part of the work concerns data.
o Synthetic image creation methodologies in order to quickly build up a learning base to optimise the Deep-Learning modules in specific versions for the targeted application case.
- Synthetic image generation (GAN)
- Style Transfer Tools
o Unsupervised learning/clustering methods for data exploration and analysis.
- Dataset analysis for a more in-depth understanding of the different characteristics of the Bases and which could be appropriate for classification.
.
- Selects groups of images with similar visual and geometric characteristics so that data-mining can then be performed, e.g. selecting the most relevant images for annotations and learning, instead of selecting images randomly
.
Projet en cours :
MONITORING (2021)
RECARIA (2021)
The CEA List develops AI methodologies based on Deep-Learning to extend the implementation of usual approaches to complex industrial applications (little/no data, potentially of poor quality). A significant part of the work concerns the algorithms.
o Approaches such as Transfer Learning allowing the optimisation of tools dedicated to a given type of production by considering as a baseline modules previously optimised on other types of parts. These approaches have the advantage of requiring less data than for "from scratch" development and of having a significant saving in the time required for optimisation.
o Development of methodologies for progressive deployment of AI approaches according to the number of data available and their level of qualification by business experts. For this, modules developed at CEA List and integrating innovative AI tools will be integrated into a common framework (One-Class > Zero Shot / Few Shot > Multi Class). The expert operator would qualify the level of performance of the different stages, notably via the expertise of the data from the modules. This business feedback, coupled with statistical analysis of the images generated during the checks (using CEA data qualification tools), would make it possible to validate the change of stage.
o Travaux amont sur des approches IA innovantes pour le développement de modules à partir de bibliothèques de modèles disponibles intrinsèquement explicables (intégrant des outils issus de la physiques théorique) judicieusement sélectionnés par des opérateurs experts des contrôles à effectuer
AR consists of mixing the real and virtual worlds in such a way as to provide the user with digital information to guide him or help him understand his environment. In the manufacturing field, the use of AR technologies offers many advantages to facilitate various operations: maintenance, training, assembly assistance, etc. To achieve this, the systems used must offer high levels of performance (precision, stability, latency, operation in difficult environments, etc.).
The List develops SLAM tools coupled with model-based constraints to meet industrial requirements. These methodologies are applicable if a 3D model of a part of the environment is known - in the case of AR applications, the CAD model is used as the basis for the development of the SLAM tools.
Since 2011, the modules of SLAM constraint Model are transferred to the company DIOTA which offers within its DiotaPlayer a set of Augmented Reality functionalities allowing to superimpose working instructions with precision on real objects. SIALV's work enables DIOTA products to meet industrial challenges such as the complexity of scenes, the required robustness, the constraints of on-board systems, etc.
Since 2017, numerous computer vision works for picking/control applications by the List have led to strong advances and the acquisition of first-rate technological know-how and bricks.
Among these works, some of them dealing with optical/sensor developments within the framework of the SIALV/TRIDIMEO joint laboratory have enabled TRIDIMEO's 3D/MS cameras to be upgraded (sensitivity, resolution, etc.) in order to meet industrial criteria (acquisition time, precision, etc.). At the same time, the CEA has developed innovative 3D registration approaches (3 patents in progress / 5 others to come). This work has enabled a major milestone to be reached, enabling these solutions to be implemented from 2018 in an automotive production site for a picking application - the interfacing of the assembly with a robot arm was notably addressed in this context, making it possible to have a complete picking system.
In its current version, the complete picking system (TRIDIMEO sensor + CEA Algos + robot arm) is at a perennial level of development, both in terms of maturity and stability.
Beyond the current level, work in progress allows us to envisage very promising prospects. Some of this work relates to the implementation of AI methodologies for localisation and/or control applications, thus considerably increasing the possible applications to objects with very diverse shapes/aspects, for which the currently implemented geometric approaches are deficient.
Since 2017, these alignment modules have been transferred to the company TRIDIMEO, which designs and markets a new generation of high-performance multispectral 3D cameras and develops vision software solutions that enable various robotic guidance systems with short cycle times and high-performance quality controls.
• Current projects :
PICK A FUTURE (2020-2021)
Work is being carried out to integrate data-mining tools into Electronic Document Management (EDM) technologies.
◦ Discovery: extraction of business terminology, document matching.
◦ Structuring: document categorisation, structuring of business concepts.
◦ Restitution: semantic search engine, synthetic business view of documents.
◦ Easy adaptation to the domain for greater user autonomy.
The technological stakes are multiple and ambitious:
◦ Making CLIMA evolve for the needs of EDM 2.0.
◦ Capitalise the evolutions of the search engine according to the cases of use treated.
- Search by keywords, entities, relations and cross-document exploration.
◦ Develop the automatic terminology structuring theme.
Production, maintenance or inspection processes require semi-structured reports containing metadata, text and images. They serve as a support for the transmission or capitalisation of information.
Due to the lack of available time and the working environment, the reports are poorly written and/or incomplete, making them difficult to exploit through automatic processing. In addition, reports are written "on the way back to the office", which often results in a loss of information.
NLP tools have been implemented to simplify the report writing process by automatically structuring dictated reports for integration into automated processes.
Financed as part of Factory Lab's innovative projects, the DIVORA prototype aims to simplify the process of writing structured reports in the industrial context by adapting automatic written language processing tools for the analysis and interpretation of voice dictation. The project has benefited from the input of industrialists through the field truth of their use cases: Bureau Veritas, Vinci Construction, Safran, TechnipFMC and PSA.
Below a demo (in French):