Search

With the increase of network virtualization and the disparity of vendors, the continuous monitoring and detection of anomalies cannot rely on static rules. An advanced analytical methodology is needed to discriminate between ordinary events and unusual anomalies. In this paper, we focus on log data (textual data), which is a crucial source of information for network performance. Then, we introduce an algorithm used as a pipeline to help with the pretreatment of such data, group it into patterns, and dynamically label each pattern as an anomaly or not. Such tools will provide users and experts with continuous real-time logs monitoring capability to detect anomalies and failures in the underlying system that can affect performance. An application of real-world data illustrates the algorithm.

International audience; Microelectronics production failure analysis is an important step in improving product quality and development. Indeed, the understanding of the failure mechanisms and therefore the implementation of corrective actions on the cause of the failure depend on the results of these analysis. These analysis are saved under textual features format. Then such data need first to be pre-processed and vectorized (converted to numeric). Second, to overcome the curse of dimensionality caused by the vectorisation process, a dimension reduction is applied. We are first interested in studying the potential of using an unsupervised variable selection technique to identify the variables that best demonstrate discrimination in the separation and compactness of groups of textual data. Variable selection has been approached by several variable or feature selection methods. Some of them have not been adapted for use in large data sets or are difficult to tune, and others require additional information. This work investigates the potential of using a genetic algorithm to find, in an unsupervised way, the variables allowing the best discrimination of the classes, to select variables correlated to particular textual groups. The proosed genetic algorithm uses a combination of the K-means clustering and validity indices as a fitness function for optimization. Such a function improves both compactness and class separation. Experiments on textual datasets demonstrate the effectiveness of the proposed method of variable selection which allows better discrimination of textual classes compared to the use of K-means clustering on all data variables.

Unsupervised clustering algorithms are a very important source of information for how a dataset may be classified into subgroups of homogeneous sets. For these algorithms we present a full analysis of their quality by introducing a clustering confidentness metric. Based on this metric, a statistical test and a correction of cluster probabilities are introduced to improve the performance of the different algorithms. These results are illustrated by simulation analysis and by an application on a real world data set.

International audience; In the semiconductor industry, Failure Analysis (FA) is an investigation to determine the root causes of a failure. It also involves an intermediate analysis to build the steps of the failure analysis in order to mitigate future failures and to facilitate the future FA. In the framework of the FA 4.0 project, the reporting system records three items of information using natural language: the failure analysis request description (input space) and analysis steps (paths), as well as generic categories of root cause conclusion (output space). The main objective of this article is to develop a system which is able to automatically help industries carry out fault analysis diagnoses with Artificial intelligence (AI). This article extends and validates the adapted methodology proposed by [1] to transform text data into numeric data based on Natural Language Processing (NLP). It transforms the text data from the input space and output space. Different deep learning algorithms based on a Variational AutoEncoder (VAE) are applied to the output space to reduce the dimension of the numeric data, and the performance of each VAE is evaluated with different metrics. The Generalized-Controllable VAE (GCVAE) is the one best suited to our case. A Gaussian Mixture Model (GMM) is then used to perform clustering in the latent space generated by the GCVAE. A centroid analysis is also conducted to verify the similarity of each cluster.

Poster; International audience; Failure analysis (FA) is key to a reliable semiconductor industry. Fault analysis, physical analysis, sample preparation and package construction analysis are arguably the most used analysis activity for determining the root-cause of a failure in semiconductor industry 4.0. As a result, intelligent automation of this analysis decision process using artificial intelligence is the objective of the Industry 4.0 consortium. The research presents natural language processing (NLP) techniques to find a coherent representation of the expert decisions during fault analysis using β-variational autoencoder (β-VAE) for space disentanglement or class discrimination and Gaussian Mixture Model for clustering of the latent space for class identification.

International audience; The growing trend towards free access to satellite images raises the economic valuation question with a particular acuity. Valuing the satellite images is not a simple process. Asking "how valuable is a satellite image" should be followed by the question, "how valuable to whom". A distinction must be made between the impacts of use of the satellite images, and measuring quantitatively the value of the images. Recently, Laxminaryan & Macauley (2012) states that the contingent valuation method, which refers to the consumers' willingness to pay (WTP), can be the most important method for the assessment of the benefits derived from Earth observation data. The valuation study comes to assess the importance of satellite imagery as a support the territorial planning and development economics. It provides elements allowing to establish on a next level, pricing scenarios in order to sustain the GEOSUD business model in the long run.

In a time series, an observation is called a record at time «t» if its value is greater than all previous values. As «t» increases, consider the sequence of records and the sequence of indices of occurrence of the records.The stochastic properties of sequences of record values have been much studied in the case where the observations are independent and identically distributed (iid) random variables. It turns out that many of these properties are universal, i.e. they hold for any cumulative distribution function for the underlying observations. In particular, records have a tendency to become further separated in time as «t» increases. However, this is not what is observed in many real data sets. This has lead to the development of more comprehensive models to provide better prediction.One of the simplest and popular model for a series of records extracted from independent but not identically distributed observations is the linear drift model (LDM). This model has been studied by many authors and found to be in agreement with some data sets where the iid assumption does not hold. However, for its uses in practical situations, the LDM requires the specification of the drift parameter of the model and this brings the problem into the realm of statistics.There are similarities between records and censored data in e.g. survival analysis. In particular, all observations that fall between two consecutive records and beyond the last record, can be seen as censored, by the last observed record. To highlight these similarities, consider the sequence of record indicators which are 1 if the observation is a record and 0 otherwise.Another popular model is the Yang-Nevzorov model. This model is interesting because it has the structure of a proportional hazard model, which have been shown to provide good fit to many data sets in survival analysis. However, to the best of our knowledge, statistical inference for the Yang- Nevzorov model has been little developed.The goal of this work is to introduce some estimators of the parameters in LDM and Yang’s model respectively and derive their statistical properties. It is shown that the censoring mechanism is informative for certain parameters. This justifies investigating the usefulness of estimators that can be extracted from record indicators. We give some exact and asymptotic properties of these estimators. It turns out that in a Yang’s model, the behavior of these estimators is distribution-free, i.e. does not involve the underlying CDF. Note that our estimators can be used even when the exact value of the records are themselves unavailable or of poor quality and only the indicators of their occurrence are available or trustworthy. Also, it is shown that distribution-free goodness-of-fit tests for Yang’s model can be derived from these indicators. These tests even have some diagnostic capabilities that can help in suggesting corrections to the model.Still in the context of a Yang’s model, we study the stochastic behavior of the inter-record time and give its asymptotic distribution regardless of the choice of the underlying distribution. In addition, we apply our theoretical results to a previously analyzed data set.Finally, we turn to the use of all available data (record values and indices/indicators) in order to calculate, by several methods, estimators of parameters in LDM and Yang-Nevzorov’s model. In addition, we introduce statistical tests that help us to check the conformity of the choice of the underlying distribution and to choose between LDM and Yang.; Dans une série chronologique, une observation est appelée record au temps «t» si sa valeur est supérieure à toutes les valeurs précédentes. Suivant l’augmentation de «t», considérons la suite des valeurs des records et la suite des indices d’occurrence des records. Les propriétés stochastiques des suites de valeurs des records ont été largement étudiées dans le cas où les observations sont des variables aléatoires indépendantes et identiquement distribuées (iid). Il se trouve que beaucoup de ces propriétés sont universelles, c’est-à-dire tiennent pour n’importe quelle loi de probabilité commune des observations. En particulier, les records ont tendance à devenir plus espacés dans le temps à mesure que «t» augmente. Cependant, ce n’est pas ce que l’on observe dans de nombreux jeux de données réelles. Ceci a conduit à l’élaboration de modèles plus complexes pour fournir une meilleure prédiction.Le modèle, peut-être le plus simple mais en tout cas le plus populaire, pour une série de records issus d’observations indépendantes mais non identiquement distribuées est le modèle à dérive linéaire (LDM). Ce modèle a été étudié par de nombreux auteurs et trouvé en accord avec certains types de données où l’hypothèse iid ne tient pas. Cependant, dans des situations pratiques, l’utilisation du LDM nécessite la détermination du paramètre de dérive du modèle et cela pousse le problème dans le domaine de la statistique.Il existe une similitude entre les records et le traitement de données censurées en analyse de survie. En particulier, toutes les observations qui tombent entre deux records consécutifs et au-delà du dernier record peuvent être considérées comme des observations censurées par le dernier record observé. Pour mettre en évidence cette similitude, on introduit la suite des indicatrices de records qui prennent la valeur 1 si l’observation est un record et 0 sinon.Un autre modèle populaire est le modèle Yang-Nevzorov. Ce modèle est intéressant car il a la structure d’un modèle à risque proportionnel en analyse de survie, lequel a montré son utilité dans ce domaine pour modéliser de nombreux jeux de données. Cependant, à notre connaissance, l’inférence statistique pour le modèle Yang-Nevzorov a été peu développé.Le but de ce travail est d’introduire certains estimateurs des différents paramètres des modèles LDM et Yang respectivement et d’en obtenir leurs propriétés statistiques. Il est montré que le mécanisme de censure est informatif pour certains paramètres. Cela justifie l’analyse des qualités d’estimateurs qui peuvent être obtenus à partir de ces indicatrices de records. Nous donnons quelques propriétés exactes et asymptotiques de ces estimateurs. Il se trouve que dans le modèle de Yang, le comportement des différents estimateurs est indépendant de la distribution sous-jacente. Notons que nos estimateurs peuvent être utilisés même lorsque les valeurs exactes des records sont elles-mêmes indisponibles ou de mauvaise qualité et les seules indicatrices sont disponibles ou fiables. En outre, il est montré que des tests d’ajustement du modèle de Yang peuvent aussi être dérivés de ces indicatrices. Ces tests ont même des capacités diagnostiques qui peuvent aider à suggérer des corrections au modèle.Toujours dans le contexte d’un modèle de Yang, nous étudions le comportement stochastique du temps inter-records et nous donnons sa loi asymptotique, indépendamment de la loi des va sous-jacentes. De plus, nous appliquons nos résultats théoriques à des données analysées précédemment par Yang.Enfin, nous passons à l’utilisation de la totalité des données disponibles (valeurs et indices/indicatrices de records) afin de calculer, par plusieurs méthodes, des estimateurs des paramètres des modèles LDM et Yang-Nevzorov. De plus, nous introduisons des tests statistiques qui nous aident à vérifier la conformité du choix de la distribution sous-jacente des observations et à choisir entre un modèle LDM et de Yang.