Joseph Thompson is a Ph.D. student in joint supervision with Polytechnique Montréal and IMT Atlantique. The topic of his research is the incorporation of machine learning and metaheuristics to solve mixed-model assembly line balancing problems with walking workers (MMABP-W). This problem involves the planning of production lines with multiple product types, commonly found in the automotive and electronics industries. The sheer number of possible configurations of workers, equipment, and tasks make MMABP-W difficult to solve with exact methods, such as mixed integer linear programming. Joseph is working on approximative methods that can solve large problem instances with limited memory and processing.
In the context of airborne Anti-Submarine Warfare (ASW), sonar constitutes one of the most effective means in the available arsenal for searching, tracking and detecting underwater threats. The objective of this thesis work in collaboration with Thales Defence and Mission Systems (DMS) is to optimise the spatial and temporal configuration of a sonar network in order to maximise the probability of detecting an underwater target (i.e. ideal insonification of the area). More specifically, this thesis focuses on Multistatic Sonar Networks (MSNs) made up of a set of active sonar systems, i.e. with emission and reception of an acoustic wave. Such systems may be in a monostatic and/or bistatic configuration. A sonar system is said to be monostatic when the transmitter and receiver are co-located and bistatic when they are delocalised, i.e. positioned at two distinct geographical locations and potentially on two disjoint immersion planes. Finally, the retained application concerns acoustic buoys, also known as "sonobuoys". These are deployed and parachuted from an airborne carrier such as a Maritime Patrol Aircraft (MPA), a helicopter or even a Unmanned Aerial Vehicle (UAV) over the Area of Interest (AoI). Upon impact with the water surface, they submerge to a predetermined depth and begin their life cycle for a given time period (transmission/reception and UHF/VHF communication with the aircraft). Furthermore, they can be transmitter-only (Tx), receiver-only (Rx) or transmitter-receiver (TxRx).
In recent days, the usage of learning algorithm to improve optimization methods have become increasingly interesting. For example, the Vehicle Routing Problem (VRP) that logistic companies might face daily. The main problem is arising whenever the delivery routes could not be optimal, which causes an increase in delivery costs. To reduce these costs, we need optimize the delivery route. However, most optimization algorithm still solves the problem from scratch, even for the same problem type, and nothing useful is extracted from prior solutions. Meanwhile, the historical data could be useful to gain solutions efficiently and effectively. In term of optimization algorithm, the use of artificial intelligence (AI) for solving VRP promise to learn from past solutions or in real-time and then to guide the algorithm to solve the problem. Moreover, the optimization algorithm could learn from its own decisions and adjust its behaviour accordingly to gain better result. Therefore, the objective of this research is divided into two goals: (1) to get an understanding of the connection between the quality of the solutions, their features, and the associated problem instances, and (2) to develop an efficient learning process consolidated with a powerful optimization algorithm to solve the problems quickly and effectively.
The district heating network (DHN) facilitates the integration of renewable heat producers and enables efficient energy transport with an optimal control. To take advantage of these advantages, it is necessary to optimize and simulate the network, which is computationally expensive. Dubon joined the Lab-STICC DECIDE team in October 2022 and is also affiliated to the GEPEA OSE team. From the formulation of the DHN as a connected graph, in his thesis, Dubon studies a new method to reduce this simulation cost by aggregating similar consumer nodes using the latest artificial intelligence architectures. His objective is to develop two complementary models: an unsupervised model to identify similar consumer nodes and a time-dependent supervised model to learn the physical dynamics of these consumer nodes.
Modern ships are much more efficient than their 19th or 20th century ancestors. Like factories, these improvements have been made possible by the high degree of systems automation. These systems, hybrids between computers and mechanics, are called cyber physical systems. Industrial control systems (ICS / SCADA) that drive machines, navigation systems, such as AIS are some examples of these CPS. Like all computer systems, they are vulnerable to cyber attacks, but a malfunction on these systems can lead to a disaster in the real world. In addition, it is very difficult to modify these systems to make them "cyber by design". This is why it is important to be able to detect quickly the occurrence of a cyber attack on these systems. There are several methods to do this: signature databases, expert rules or AI. The research questions identified at this stage are the following: What is the place of AI for the detection of attacks on physical cyber systems of ships? How to evaluate the performance of a cyber attack detection algorithm in a maritime SPC context? How to allow the decision maker to choose the most efficient algorithm or combination of algorithms in this context?
The objective of Antoine Mallégol's thesis work is to propose an optimization method for a coupled energy system, composed of an electrical network and a heat network that can interact with each other. The main challenge of this work is to solve a complex model with a long time horizon. Different modeling and optimization methods are proposed: mixed integer linear programming, heuristics, and metaheuristics. A multi-objective approach allows to improve the costs and the environmental impact of the system.