Transverse projects

MAP/INF630 (3 ECTS): Students will work half a day a week on a transverse project, ie. a case study corresponding to a challenging question either raised by an industrial partner or by a researcher in the domain spanned by the graduate degree.

Presentation of transverse projects will occur on 25th September at 2pm in Alan Turing building (on Polytechnique campus) in room Sophie Germain.

Enedis - contact Frédéric Boutaud frederic.boutaud@enedis.fr

  • Dataposte: Détection automatique du type de matériel en place à partir de photos, pour l'aide à l'intervention lors des missions de maintenance du réseau. - contact Séverine MULATIER - severine.mulatier@enedis.fr

Numerous pictures campaigns had been organized nationwide by Enedis inside secondary substations in order to collect informations about material (manufacturer, model….). During this summer, the first version of an algorithm have been developed to identify texts in label printings (OCR processing) for hight voltage cell unit, LV switchboard and fault detector. The aim of this project is to :

- Identify if AI can permess us to identify label printings of transformers (more complexe than others materials).

- Set up a deep learning process to classify a big amont of pictures

- Join a transverse task force organised arround this topic

  • Predictive maintenance - contact: Eric TEYSSEDRE - eric.teyssedre@enedis.fr

Context. Enedis is deploying Linky communicating meter and the program will be ended by 2022. More than 13 million meters are yet installed. Data recorded in each Linky meter provide valuable information about events which appear on electricity grid, which is a new opportunity for Enedis. On the other hand, failures affect the low voltage network leading to power outages, about 40000 a year.

Problem. The problem is the following: how to take advantage of new available data (Linky meter data, weather, network load,…) to detect anomalies on the electrical network and avoid power outages ? In other words, how to develop predictive maintenance to optimize our resources?

Description. The idea is to use data registered by Linky meter (short outages, voltage excursions, surges) and other data considered relevant which need to be identified, in order to build an algorithm based on Artificial Intelligence (AI) allowing to predict and characterize failures. It comes specifically to search correlations between data available and network outages to define failures “signatures”. These signatures would be so recognized by the algorithm and would lead to recommendations for action on the field to correct the anomaly before the incident. More precisely, the study will focus on: - Quick benchmark of use of AI to predict failures in other electrical companies - Critical analysis of the Enedis works undertaken on the subject (method, algorithm,…) - Research of new correlations between some types of failures and data available - Design of an algorithm or improve existing algorithm for predicting failures The student would work in collaboration with Enedis data scientist who started a study on the subject. The student would work physically on the Enedis site of Nanterre with data scientist during the phase of data analysis and data processing (sensitive data). For other part of the study or to analyse non-sensitive data, he will have the possibility to work at Polytechnique. Depending on the first results, this study could lead to a project of several months in order to continue works engaged.

  • Customer relationship: Création de chatbot conviviaux pour répondre aux demandes des utilisateurs - contact: Richard BAVARIN - richard.bavarin@enedis.fr

But : apprécier comment l'IA pourrait prendre en compte le volet “émotionnel” des appels clients reçus sur notre Centre d'Appel Dépannage. Des enregistrements réels seront utilisés pour l'apprentissage. Le premier volet de se projet consistera en un état de l'art sur le sujet pour évaluer une première faisabilité sur le cas d'usage proposé.

Google contact Damien Henry damienhenry@google.com, project Google Arts and Culture.

  • Automatic detection of Art Style in paintings: Il existe de nombreuses bases de données non structurée, pour lesquels il pourrait être interessant de detecter automatiquement des méta donnée, en particulier le mouvement artistique, l'auteur, etc… (voir ici, et ).
  • Image generation: La generation d'image grâce au ML est en plein essor avec de nombreuses applications possibles. La technique classique est basée sur des Generative Adversarial Network. Une technique plus récente et prometteuse est basée sur les Normalizing Flow. (GAN et GLOW). L'objectif de ce projet est de comparer plusieurs approches pour générer des images de visage à partir d'une base d'images d'apprentissage. Dans les deux cas, des données peuvent être trouvée ici.

Idemia contact Stéphane Gentric stephane.gentric@idemia.com

  • Semi-supervised learning for a localization task. (possible continuation in an internship, and eventually a CIFRE PhD) Using a DCNN (Deep Convolution neural Network), we want to learn the absolute position, scale and rotation of an object in an image. Standard methods rely on annotated data and are limited by the precision of those annotations. We want to study the feasibility and performance of a learning process without any annotations, using only the fact that when applying a given similarity to the image, the expected changes in position, scale and rotation are known. We will start with a toy problem and hopefully move on to real objects and more complex scenes.
  • Building an image-based algorithm selector for face recognition based on speed and performance of candidate algorithms (possible continuation in an internship). The increasingly ubiquitous presence of biometric solutions and face recognition in particular in everyday life requires Idemia to adapt its solutions for practical requirements, may they be memory space, speed or performances. Idemia has developed several solutions, but where global decisions can be made, they are far less efficient then tailoring such decisions to the complexity of each image, which allows for the best compromise between constraints such as speed and performances. We would like to build a DCNN (Deep Convolution neural Network) selector of the best suited solution to each input image. For this purpose, we will lend a coding/matching software suite capable of generating different options.

Ynsect (start-up; not official partners yet) contact Arturo Escaroz Cetina arturo.escarozcetina@ynsect.com

  • Conduite d’élevage 4.0 : Automated Insects Physiological Data Retrival from Insect Population pictures. @Ynsect (www.ynsect.com), our insects are raised into trays of various size : Various pictures of them are taken regularly to perform quality control operations. We would like to enhance our data collection methods to get significant improvement on our insect population modeling tools. The objective of this project is to convert pictures into already know data of interest (visual computing) & to make data driven R&D into the discoveries of any observable Behavior patterns through pictures (AI). Already known data of interest can be picked in : insect Size, stage, density, color, number of rings, defects, behavior, Amount of feed / top layer description, population distribution pattern; etc. The project might include: data collection methods’ revision; Image characterization and classification; Pattern recognition & Predictive tools’; Any other methods tools that could be of interest and that we don’t know of yet !