International audienceAnomaly detection in time series has attracted considerable attention due to its importance in many real-world applications including intrusion detection, energy management and finance. Most approaches for detecting outliers rely on either manually set thresholds or assumptions on the distribution of data according to Chandola, Banerjee and Kumar. Here, we propose a new approach to detect outliers in streaming univariate time series based on Extreme Value Theory that does not require to hand-set thresholds and makes no assumption on the distribution: the main parameter is only the risk, controlling the number of false posi-tives. Our approach can be used for outlier detection, but more generally for automatically setting thresholds, making it useful in wide number of situations. We also experiment our algorithms on various real-world datasets which confirm its soundness and efficiency

Nutrient requirements of sows during lactation are related mainly to their milk yield and feed intake, and vary greatly among individuals. In practice, nutrient requirements are generally determined at the population level based on average performance. The objective of the present modeling approach was to explore the variability in nutrient requirements among sows by combining current knowledge about nutrient use with on-farm data available on sows at farrowing [parity, BW, backfat thickness (BT)] and their individual performance (litter size, litter average daily gain, daily sow feed intake) to estimate nutrient requirements. The approach was tested on a database of 1,450 lactations from 2 farms. The effects of farm (A, B), week of lactation (W1: week 1, W2: week 2, W3+: week 3 and beyond), and parity (P1: 1, P2: 2, P3+: 3 and beyond) on sow performance and their nutrient requirements were evaluated. The mean daily ME requirement was strongly correlated with litter growth (R2 = 0.95; P < 0.001) and varied slightly according to sow BW, which influenced the maintenance cost. The mean daily standardized ileal digestible (SID) lysine requirement was influenced by farm, week of lactation, and parity. Variability in SID lysine requirement per kg feed was related mainly to feed intake (R2 = 0.51; P < 0.001) and, to a smaller extent, litter growth (R2 = 0.27; P < 0.001). It was lowest in W1 (7.0 g/kg), greatest in W2 (7.9 g/kg), and intermediate in W3+ (7.5 g/kg; P < 0.001) because milk production increased faster than feed intake capacity did. It was lower for P3+ (6.7 g/kg) and P2 sows (7.3 g/kg) than P1 sows (8.3 g/kg) due to the greater feed intake of multiparous sows. The SID lysine requirement per kg of feed was met for 80% of sows when supplies were 112 and 120% of the mean population requirement on farm A and B, respectively, indicating higher variability in requirements on farm B. Other amino acid and mineral requirements were influenced in the same way as SID lysine. The present modeling approach allows to capture individual variability in the performance of sows and litters according to farm, stage of lactation, and parity. It is an initial step in the development of new types of models able to process historical farm data (e.g., for ex post assessment of nutrient requirements) and real-time data (e.g., to control precision feeding).

Abstract. This paper deals with the incremental off-line computation of diagnosis of discrete-event systems. Traditionally, the diagnosis is computed from the global automaton describing the observations emitted by the system on a whole time period. The idea of this paper is to slice this global automaton according to temporal windows and to compute local diagnoses for each of these windows. It is shown that, under some conditions, the global diagnosis can be computed from the local diagnosis. This paper presents the formalization used to compute an incremental diagnosis, relying on the new concept of automata chain. It is then shown that it is possible to take into account the diagnosis obtained for the previous temporal windows to incrementally compute the current diagnosis more efficiently. This work is a first and necessary step before considering the on-line diagnosis computation. The main difficulty is then to ensure the correct slicing of the observation automaton and to determine the appropriate temporal windows.

Understanding data distributions is one of the most fundamental research topic in data analysis. The literature provides a great deal of powerful statistical learning algorithms to gain knowledge on the underlying distribution given multivariate observations. We are likely to find out a dependence between features, the appearance of clusters or the presence of outliers. Before such deep investigations, we propose the folding test of unimodality. As a simple statistical description, it allows to detect whether data are gathered or not (unimodal or multimodal). To the best of our knowledge, this is the first multivariate and purely statistical unimodality test. It makes no distribution assumption and relies only on a straightforward p−value. Through real world data experiments, we show its relevance and how it could be useful for clustering.