052 LLM based Health Aware Control Learning System for Reusable Engines

Job title:

052 LLM based Health Aware Control Learning System for Reusable Engines

Company:

Centre National d’Etudes Spatiales

Job description

Your application must include a recommendation letter from your Ph.D. supervisor, a detailed CV including university education and work experience, a list of publications, a 2-page description of the work undertaken during the course of your PhD.For more Information, contact : Directeur de Recherche :Submit the complete application online (Apply) before March 14th, 2025 Midnight Paris timeThe demand for space launch capabilities in Europe is increasing, with a focus on reducing costs, improving performance, offering flexible solutions, and managing the carbon footprint. Reusing rocket stages addresses these challenges, necessitating the development of reusable engines, such as the PROMETHEUS engine, which requires effective regulation during operation. Research by CNES and CRAN (* & **) has explored engine regulation within a reuse context, accounting for material wear/degradation over successive flights to extend in a safety manner the useful life of components. The reuse of an engine throughout its nominal lifecycle necessitates integrating unknown variable dynamics into the control loop, which leads to intelligent reconfiguration of the control law by forecasting the health state evolution and estimating the defective component’s lifespan. This postdoctoral research project aligns with these efforts to consider both Health Monitoring Systems based on the recent approach on Large Language Models learning with optimal control for Reusable Engines.Large Language Models (LLMs) like GPT-3 and LLAMA-3, based on “Transformers,” have shown remarkable ability in capturing sequential and semantic relationships in complex datasets. Building on LLMs, we aim to develop a prognostic module/approach. Component degradation results in data output shifts or changes relative to input data. LLMs, trained to recognize and predict degradation dynamics in sequential data, can effectively model these changes as recently illustrated in [1]. This research aims to develop a prognostic module for LPRE (Liquid Propellant Rocket Engines), capable of estimating current health status using incoming data and predicting Remaining Useful Life (RUL) through advanced learning techniques using LLMs.Additionally, a second major objective is to design a control reconfiguration loop that incorporates prognostic-based information to ensure optimal system operation under degradation. This approach aims to extend the system’s RUL while maintaining stability. Our prior collaboration with CNES will be leveraged for this aspect, utilizing existing controller knowledge which performs well under nominal conditions.While the first part involves developing an efficient prognostic module based on simulator input/output data, the second part focuses on control design that ensures stability and optimal performance under system degradation. This involves incorporating non-linearity from degradation mechanisms (aging) and dynamic changes with operational points within the learning paradigm using LLMs.Such an innovative project based on two major objectives has an expected duration of 2 years. To achieve these two objectives, an exploration of the existing state-of-the-art in LLMs is required. Existing methods must be studied. The goal, over a 16-month period, is to develop an LLM-based strategy to estimate system health state and predict future health indicators and RUL. This approach should integrate both model-based insights and available experimental databases (e.g., sensor data, historical data). Activities include designing LLM architectures specifically for time series analysis, extending standard transformer models for effective long-sequence handling, generate data using a CNES simulator across various degradation scenarios, build data processing pipelines, exploring fine-tuning methodologies suited for engine health data to detect degradation trends accurately, implement a data processing system leveraging transfer learning to expedite learning with pre-trained models on similar sequential data types.Over an 8-month period, the main goal is to adapt/develop an advanced controller with a prognostic module; a controller combined with an optimization procedure inspired from [2] and considered in PROMETHEUS engine simulator, searching for an optimal balance between system performance and the desired RUL extension. The challenge is integrating the controller with the LLM-based prognostic module without defined an estimation of a degraded system.This post-doctoral project will target dissemination of the work in high quality scientific journals (Q1) where the candidate will be expected to publish. LLMs and simulation platform will require time computing for learning step, consequently Carbon Footprint index will be estimated during all project.

• Dr. MS JHA: :

** Prof. D. THEILLIOL :References[1] Wang, H., Li, Y. F., & Xie, M. (2023). Empowering ChatGPT-Like Large-Scale Language Models with Local Knowledge Base for Industrial Prognostics and Health Management. arXiv preprint arXiv:2312.14945.[2] Thuillier, J., Jha, M. S., Le Martelot, S., & Theilliol, D. (2024). Prognostics Aware Control Design for Extended Remaining Useful Life: Application to Liquid Propellant Reusable Rocket Engine. International Journal of PHM, 15(1).ProfilPh.D in computer science, AI based, control theory, Prognostics and health management, natural language processing, or LLMs is expected. Python and Matlab language are required, C1 English level will be appreciated. Some backgrounds on Deep learning based sequential predictions, time series forecasting using Deep neural networks, preferred previous experience with LLMs, NLP, transformers for time series prediction are appreciated.

Expected salary

Location

Vandoeuvre-lès-Nancy, Meurthe-et-Moselle

Job date

Wed, 05 Feb 2025 07:52:12 GMT

To help us track our recruitment effort, please indicate in your email/cover letter where (vacanciesin.eu) you saw this job posting.