Abstract

International audience; The modes of reasoning which are used in the context of safety analysis and the very nature of knowledge about safety mean that a conventional computing solution is unsuitable and the utilization of artificial intelligence techniques would seem to be more appropriate. Our research has involved three specific aspects of artificial intelligence: knowledge acquisition, machine learning and knowledge based systems (KBS). Development of the knowledge base in a KBS requires the use of knowledge acquisition techniques in order to collect, structure and formalizes knowledge. It has not been possible with knowledge acquisition to extract effectively some types of expert knowledge. Therefore, the use of knowledge acquisition in combination with machine learning appears to be a very promising solution. This paper presents the result of these two research activities which are involved in the methodology of safety analysis of guided rail transport systems.

Abstract

International audience; Despite of the success of artificial intelligent (AI) methods in many domains, there is big dilemma for AI when applying to air traffic management (ATM). That is AI researchers have long stated their AI methods are effective and reliable enough to handle many ATM problems, while human controllers still refuse to adopt such AI methods. We believe the dilemma is not about whether AI methods is effective or reliable enough, but about why human controllers should be replaced by AI methods. In other words, as long as an AI method aims to compete and replace human controllers, it will be confronted with the difficulty of not being accepted by human controllers. To address this dilemma, this paper proposes a new thinking about applying AI methods, i.e., an AI method should be developed in such a way of assisting human controllers, but never in the way of competing and replacing human controllers. This new thinking is called human-machine coevolutionary intelligence (HMCEI). A methodological framework of HMCEI is further developed for decision-making support of ATM, in order to demonstrate the concept of HMCEI is practicably possible.

Abstract

International audience; This paper addresses the problem of estimating aircraft on-board parameters using ground surveillance available parameters. The proposed methodology consists in training supervised Neural Networks with Flight Data Records to estimate target parameters. This paper investigates the learning process upon three case study parameters: the fuel flow rate, the flap configuration, and the landing gear position. Particular attention is directed to the generalization to different aircraft types and airport approaches. From the Air Traffic Management point of view, these additional parameters enable a better understanding and awareness of aircraft behaviors. These estimations can be used to evaluate and enhance the air traffic management system performance in terms of safety and efficiency.

Abstract

International audience; This article describes a contribution to improving the usual safety analysis methods used in the certification of railway transport systems. The methodology is based on the complementary and simultaneous use of knowledge acquisition and machine learning. We used the ACASYA software environment to support the safety analysis aid methodology. ACASYA aims to provide experts with suggestions of potential failures which have not been considered by the manufacturer and which are capable of jeopardizing the safety of a new rail transport system. ACASYA consists of two main modules: CLASCA and EVALSCA, respectively dedicated to the classification and evaluation of accident scenarios. CLASCA is an inductive, incremental and interactive learning system. EVALSCA, built around a learning system called CHARADE, aims to provide experts with suggestions of potential failures which have not been considered by the manufacturer and which are capable of jeopardizing the safety of a new rail transport system.

Abstract

International audience; In addition to identifying safety deficiencies evidenced by accidents, the independent investigative body makes recommendations to eliminate or reduce these deficiencies. Its main purpose is to advance transportation safety by conducting investigations of accidents in rail and other modes of transportation. He must answer several questions: what happened, why did it happen, and what can be done to reduce the risk of it happening again? In this process, one of the difficulties involved is finding abnormal accident scenarios which are capable of generating a specific hazard. This paper proposes an original method based on machine learning to assist investigators experts in their crucial task of analysis and assessment of the safety for railway transport systems in France. This contribution is based on the use of artificial intelligence techniques and has involved the development of several approaches and tools which assist in the modeling, storage and assessment of knowledge about safety. The proposed approach has two objectives, firstly to record and store experience concerning safety analyses, and secondly to assist those involved in the development and assessment of the systems in the demanding task of evaluating safety studies.

Abstract

International audience; The most important part of the train's energy is consumed by the traction system. The tractive energy depends mainly on the driving behaviour. Improving driving strategies has great potential to enhance the energy efficiency. This paper presents a speed profile optimization approach based on a genetic algorithm. The objective of the genetic algorithm is to find, for each interstation, the best speed profile which minimizes the energy consumption. The optimized profile takes into account both the physical and the operational constraints such as the maximum allowed travel time, the speed limitations per section and the maximum allowed acceleration and jerk. The fitness function is based on a Random Forest model which is built using on-board measurements. The aim of the model is to estimate accurately the energy consumption corresponding to each speed profile. The initial population of genetic algorithm is mainly composed of the best realistic speed profiles extracted from the collected data. Document type: Conference object

Abstract

International audience; The Machine Learning (ML) world is in constant evolution, as the amount of different algorithms in this context is evolving quickly. Until now, it is the responsibility of data scientists to create ad-hoc ML pipelines for each situation they encounter, gaining knowledge about the adequacy between their context and the chosen pipeline. Considering that it is not possible at a human scale to analyze the exponential number of potential pipelines, picking the right pipeline that combines the proper preprocessing and algorithms is a hard task that requires knowledge and experience. In front of the complexity of building a right ML pipeline, algorithm portfolios aim to drive algorithm selection, learning from the past in a continuous process. However, building a portfolio requires that (i) data scientists develop and test pipelines and (ii) portfolio maintainers ensure the quality of the portfolio and enrich it. The firsts are the developers, while the seconds are the operators. In this paper, we present a set of criteria to be respected, and propose a pipeline-based meta-model, to support a DevOps approach in the context of Machine Learning Pipelines. The exploitation of this meta-model, both as a graph and as a logical expression, serves to ensure continuity between Dev and Ops. We depict our proposition through the simplified study of two primary use cases, one with developer's point-of-view, the other with ops'.

Abstract

International audience; When two or more airplanes find themselves less than a minimum distance apart on their trajectory, it is called a conflict situation. To solve a conflict, air traffic controllers use various types of information and decide on actions pilots have to apply on the fly. With the increase of the air traffic, the controllers’ workload increases; making quick and accurate decisions is more and more complex for humans. Our research work aims at reducing the controllers’ workload and help them in making the most appropriate decisions. More specifically, our PhD goal is to develop a model that learns the best possible action(s) to solve aircraft conflicts based on past decisions or examples. As the first steps in this work, we present a Conflict Resolution Deep Neural Network (CR-DNN) model as well as the evaluation framework we will follow to evaluate our model and a data set we developed for evaluation.