International audience

Lithium-ion batteries are used in most battery powered devices. Today’s research on Lithium-ion batteries mainly focuses on better energy management strategies and predictive maintenance. In this paper, a new approach based on auto-encoders and long short-term memory neural networks applied to usage data (voltage, current, temperature) is used to make a State of Health prediction. Encouraging results are obtained when conducting tests on various battery ageing datasets published by Sandia National Laboratories, the Massachusetts Institute of Technology and NASA’s Prognostics Center of Excellence.

Lithium Ion batteries are a core component of many lately designed devices. It is of crucial importance to be able to fully master the behaviour of batteries in order to meet the requirements in terms of safety, and performances. Predicting the Remaining Useful Life of batteries can help preventing a failure before it occurs, with an increased safety for the user and reduced costs linked to maintenance.The work described in this paper is an attempt to accurately predict the Remaining Useful Life of Li-Ion batteries using machine learning regression methods applied to a new set of ageing data published by the department of chemical engineering of the Massachusetts Institute of Technology. By changing the usual approach applied to data and feature management, very encouraging results were obtained. Compared with previous approaches in the literature using linear regression or Convolutional Neural Networks on the same dataset, our work on how to build a more efficient representation of ageing data combined with Artificial Neural Networks leads to more accurate predicting performances.

In this paper, we propose to improve an unsupervised segmentation algorithm based on the graph diffusion and regularization model described by Ta in [1] by using Haralick texture features. With this framework, segmentation is performed by diffusing an indicator function over a graph representing an image. The benefit of our approach is to combine two non-local processing techniques: at pixel level with texture features and through the use of a graph structure, which allows to efficiently express relations between non-adjacent pixels. Applied on ultrasound images, and compared to a vector-valued Chan & Vese active contour, our method shows an improvement of the quality of segmentation.

In this paper we propose a new supervised active contour model evolving with Haralick texture features. This model is divided in two stages. First, we use a supervised step where the user defines an ideal segmentation on a learning image. A linear programming model, modeling the behavior of the active contour, is then used to determine the weights of the Haralick features leading to the optimal segmentation. In a second step, a texture-oriented active contour based on the Chan-Vese model is launched on several test images with the learned weights and the closest segmentations to the one defined on the learning image is determined. Results of our method are presented on medical echographic images.

We describe a 3D+T active surface model for segmentation and tracking of the left ventricular endocardium within 3D MRI sequences of the cardiac cycle. In order to perform segmentation and tracking simultaneously, the surface structure is divided through both time and space, and is therefore handled as an array of planar active contours, interconnected between adjacent slices and frames, providing spatial and temporal consistency. In a given frame, the stacking of slice contours constitute a 3D triangular mesh with a cylindrical topology. Extraction of ventricle border is performed by means of energy minimization, using a combination of boundary-based term and a new computationally efficient region-based term.

This paper presents a new approach to the mesh segmentation based on watershed transformation. We propose an original method to compute markers from the topological information of the mesh. The skeleton of the mesh allows the interpretation of the meaningful parts and the watershed transformation builds the boundaries of theses parts. Our method, which combines the patch-type and the part-type segmentation approaches, is particularly well adapted to the problematic of meaningful part segmentation.

We introduce a narrow band region approach in explicit de-formable models for 2D and 3D image segmentation. Embedding a region term into the evolution process, we derive a general formulation which is applied both on a 2D parametric contour and a 3D triangular mesh. Evolution of deformable models is performed by means of energy minimization using the computationally efficient greedy algorithm. The use of a region energy related to the vicinity of the evolving surface overcomes limitations of edge-based active models while remaining time effective. Experiments with segmentation quality assessment are carried out on medical images.

Deformable models, such as the discrete active contour and surface, imply the use of iterative evolution methods to perform 2D and 3D image segmentation. Among the several existing evolution methods, we focus on the greedy algorithm, which minimizes an energy functional, and the physics-based method, which applies forces in order to solve a dynamic differential equation. In this paper, we compare the greedy and physics-based approaches applied on 2D and 3D models, as regards overall speed and segmentation quality, quantified with an evaluating function mainly based on the mean distance between the model and the desired shape.

When long-term dependencies are present in a time series, the approximation capabilities of recurrent neural networks are difficult to exploit by gradient descent algorithms. It is easier for such algorithms to find good solutions if one includes connections with time delays in the recurrent networks. One can choose the locations and delays for these connections by the heuristic presented. As shown on two benchmark problems, this heuristic produces very good results while keeping the total number of connections in the recurrent network to a minimum.

We extend the Bayesian learning framework to the modelling of multivariate time series with feedforward neural networks. The extension presented here concerns both regression and classification problems. Finally, we present preliminary results regarding the choice of appropriate priors for building such sequential models.