Thales (United Kingdom)

Automation and rigorous safety procedures have enabled commercial aviation organizations to operate with relatively low incident rates. Even so, the few incidents reported in 2024 continued to cite crew decision-making during the landing phase of flight as a contributory factor. Crew decisionmaking was also highlighted as a contributory factor in the Asiana 214 accident in 2013 that culminated in a crash following an unsafe and destabilized approach. This suggests that more needs to be done to support pilots in the landing phase of flight. The current work uses the Event Analysis of Systemic Teamwork (EAST) framework to understand why pilots present during the Asiana 214 incident continued to land the aircraft in an unsafe and destabilized configuration. Interrogation of the information networks enabled the identification of design recommendations for a novel contextdependent energy support system. Such systems tailor the presentation of information based on the current operational environment and the state of the vehicle. The potential utility of this novel decision support system was then examined by integrating the design recommendations within the information network to explore its potential impact on the system.

Abstract In this work, we investigate channel decoding techniques based on machine learning, and more specifically, on Support Vector Machines (SVM). Existing SVM-based decoders suffer from a scalability problem, characterized by the exponential growth of both the number of classifiers required to decode and the size of the training dataset as a function of the code dimension. This phenomenon, often referred to as the curse of dimensionality, renders existing SVM approaches impractical for larger codes. To address this limitation, we introduce a novel bit-wise SVM decoding strategy coupled with a noiseless optimization framework, which significantly reduces the computational burden. Specifically, for a code of length $n$ and dimension $k$, our approach decreases the number of SVM classifiers from an exponential dependence to a linear in $k$, while the required training dataset is minimized to a single noiseless codeword per class. We formulate the resulting optimization problem and derive its analytical solution. Moreover, we demonstrate that, under specific conditions, the solution of the optimization process produced by the proposed framework is equivalent to the optimal Maximum Likelihood (ML) decoding rule when applied to transmissions over Additive White Gaussian Noise (AWGN) channels.

One of the key challenges of manufacturing a precision guided ammunition is to achieve the dynamical model of the projectile. It is so far the first and one of the deepest concerns of researchers and engineers as control design and analysis mostly depend on the accuracy of the flying projectile model. When it comes to a high-speed flying object, the flight dynamic gets even more complicated as the dynamical model of the airframe varies with the change of operating conditions. These non-linear systems are very often modeled as a series of linear systems which vary among them with some so-called parameters defined for each operating point. As one of the key manufacturers of ammunition in Europe and in the world, KNDS focuses on precision guided aero stabilized ammunition system where an accurate approximation of dynamic modeling of the airframe is the first step to further development. This paper assesses a non-linear flight dynamic model of an aero-stabilized projectile, a corresponding MIMO LPV model as well as its application to stability and maneuverability analysis.

Autonomous systems (AS) often use Deep Neural Network (DNN) classifiers to allow them to operate in complex, high-dimensional, non-linear, and dynamically changing environments. Due to the complexity of these environments, DNN classifiers may output misclassifications during operation when they face domains not identified during development. Removing a system from operation for retraining becomes impractical as the number of such AS increases. To increase AS reliability and overcome this limitation, DNN classifiers need to have the ability to adapt during operation when faced with different operational domains using a few samples (e.g. 2 to 100 samples). However, retraining DNNs on a few samples is known to cause catastrophic forgetting and poor generalisation. In this paper, we introduce Dynamic Incremental Regularised Adaptation (DIRA), an approach for dynamic operational domain adaption of DNNs using regularisation techniques. We show that DIRA improves on the problem of forgetting and achieves strong gains in performance when retraining using a few samples from the target domain. Our approach shows improvements on different image classification benchmarks aimed at evaluating robustness to distribution shifts (e.g.CIFAR-10C/100C, ImageNet-C), and produces state-of-the-art performance in comparison with other methods from the literature.

Recent work has highlighted the potential benefits of exploiting ellipsoidal uncertainty-set-based robust Capon beamformer (RCB) techniques in passive sonar. Regrettably, the computational complexity required to form RCB weights is cubic in the number of adaptive degrees of freedom, which is often prohibitive in practice. For this reason, several low-complexity techniques for computing RCB weights, or equivalent worst case robust adaptive beamformer weights, have recently been developed. These techniques, whose complexities are only quadratic in the number of adaptive degrees of freedom, use gradient-based, reduced-dimension Krylov-subspace or Kalman-filtering methods. In this work, we review these techniques for passive sonar, analyzing their complexities and evaluating them initially on simulated data. The best performing methods are then evaluated on two in-water recorded passive sonar data sets. One set, containing a strong controlled acoustic source, demonstrates the ability of the algorithms to protect against signal cancellation when pointing at the source, and their ability to reject the source when pointing away from it. The other data set, recorded during a period when the boat was accelerating, demonstrates the ability of the algorithms to operate in the presence of speed-induced noises.