[STAGE MASTER 2 ]

Description Scientific fields: Artificial Intelligence, Trustworthy AI, Uncertainty Modelling, System Engineering, Information Fusion. Keywords: Artificial Intelligence, Belief Functions Theory, Cyber-Physical Systems, Smart building, Data Fusion, Heterogenuous Sources. Abstract  Cyber-Physical Systems (CPS) are increasingly deployed in safety-critical settings such as manufacturing, robotics, autonomous infrastructures, and intelligent buildings [1]. Smart Building systems, in particular, rely on dense networks of heterogeneous sensors to monitor energy consumption, indoor environmental quality, and occupant comfort. These systems must operate under heterogeneous data sources, incomplete information,sensor unreliability, and unpredictable disturbances [2]. Traditional AI-based decision systems provide deterministic outputs, often without expressing confidence levels or handling conflicting evidence [3]. FUTURE project aims to develop an uncertainty-aware and trust-oriented decision-making framework combining uncertainty modelling, AI-based predictions, and Dempster–Shafer theory. By integrating belief-function reasoning with multi-source evidence fusion, FUTURE enhances the resilience of CPS decision pipelines. A realistic CPS scenario specifically, a Smart Building environment will serve as the experimental validation context, leveraging sensor data to evaluate uncertainty, reliability, and evidence-based decision performance. Work Program/ Objectives : The work program of FUTURE is structured around eight major tasks, each corresponding to a key component of the proposed uncertainty-aware decision-making framework. These tasks form a coherent pipeline, from data importation to experimental validation, and are described below. – As a first step, we identify a publicly accessible and representative dataset that reflects real-world smart-building sensor deployments. Such datasets must capture multiple sensing modalities relevant to CPS applications, enabling the evaluation of uncertainty-aware learning, reliability assessment, and evidence fusion in realistic operating conditions. For the FUTURE project, the selected test dataset is the CU-BEMS Smart Building Energy and IAQ Data1, which provides multi-sensor recordings of energy consumption, indoor air quality, and environmental variables across an operational smart- building environment. 2. Data Cleaning, Normalisation, and Harmonisation – Preprocess the raw data through noise filtering, unit conversion, timestamp alignment, and signal normalisation. 3. Uncertainty Modeling of Data – Estimate the data uncertainty of multi- heterogeneous sources. 4. Source-Aware Categorisation of Data – Organise the processed data into meaningful categories. 5. Development of an Uncertainty-Aware AI Model – Implement and train uncertain Neural Network or an evidential deep learning model capable to produce both predictions and associated uncertainty estimates. 6. Transformation of Model Outputs into Belief Masses – Convert the outputs of the AI model into to representative class-specific supports together with residual ignorance caused by uncertainty. 7. Evidence Fusion – Combine belief masses using different composition strategies depending the detected level of conflict. 8. Decision-Making under Uncertainty – Evaluate how uncertainty and conflicting evidence influence the robustness of the final decision. ExpectedScientific/TechnicalOutput : Scientific Contributions A unified uncertainty-aware decision framework for CPS. A novel integration of belief-function theory. Experimental analysis of DS-based fusion in real CPS environments, particularly in Smart Building systems. A reproducible methodology supporting trustworthy AI research. A scientific research report. Technical Contributions : A fully implemented Python prototype of FUTURE. A validated CPS case study. References: 1. Samir Ouchani. Secure and Reliable Smart Cyber-Physical Systems. . HDR Thesis, CNAM Paris, France. 2. Safa Ben Ayed, Malika Ben Khalifa, and Samir Ouchani. Modeling distributed and flexible PHM framework based on the belief function theory. In Artificial Intelligence Applications and Innovations - 20th IFIP WG 12.5 International Conference, AIAI , Corfu, Greece, June 27-30, , Proceedings, Part I, volume , pages –. Springer, . 3. Md Mehedi Hasan, Moloud Abdar, Abbas Khosravi, Uwe Aickelin, Pietro Lio, Ibrahim Hossain, Ashikur Rahman, and Saeid Nahavandi. Survey on leveraging uncertainty estimation towards trustworthy deep neural networks: The case of reject option and post-training processing. ACM Comput. Surv., 57, . 4. Souhila Badra Guendouzi, Samir Ouchani, Hiba Al Assaad, and Madeleine El-Zaher. Ensuring the federation correctness: Formal verification of federated learning in industrial cyber-physical systems. Future Gener. Comput. Syst., :, . Profile Skills : The candidate should possess a Master student or equivalent in Software Engineering or Computer Science. She/He should have some knowledge and experience in a number of the following points: Scientic skills Probability, uncertainty modelling, and evidence-based reasoning. Fundamentals of machine learning and classification. Understanding of CPS architectures. Technical Skills Programming environments. Python programming (NumPy, SciPy, Pandas). PyTorch or TensorFlow for deep learning. Dempster–Shafer libraries (e.g., pyds). Data cleaning techniques. Git version control. Interpersonal Skills Analytical thinking and autonomy. Problem-solving. Ability to work in interdisciplinary teams. Scientific communication and reporting. Starting date Dès que possible