Skip to Main content Skip to Navigation

Détection d’Anomalies Multiples par Apprentissage Automatique de Règles dans les Séries Temporelles

Inès Ben Kraiem 1
1 IRIT-SIG - Systèmes d’Informations Généralisées
IRIT - Institut de recherche en informatique de Toulouse
Abstract : Supervision and monitoring tools are commonly used in the industry to analyze datafrom different sensors. These data are often affected by unusual events or temporary changes and tend to contain irregularities and outliers that require business knowledge and human intervention to be detected. In such situations, anomaly detection can be a crucial way to identify abnormal events and detect unusual behavior, allowing experts to act quickly and mitigate the effects of an undesirable situation. In this thesis, we focused on the use of automatic learning techniques in order to automate and consolidate the process of detecting anomalies in sensor network data. These data come from sensors and are presented in the form of time series. To do this, we have defined two main objectives: the detection of multiple anomalies and the generation of interpretable rules by humans for the detection of anomalies. The first objective is to detect different types of anomalies in the sensor data. In the existing research, there is extensive work on anomaly detection. However, most techniques look for individual objects that are different from normal objects or a sequence of data but do not take into consideration the detection of multiple anomalies. To solve this problem and reach our first issue, we have created a configurable multiple anomaly detection system that is based on patterns to detect anomalies in time series. The algorithm we propose, Composition of Remarquable Point (CoRP), is based on the principle of pattern search. This algorithm applies a set of patterns to annotate remarkable points in a uni-varied time series, then detects anomalies by pattern composition. Annotation patterns and pattern compositions are defined with the help of the subject matter expert. Our method has the advantage of locating and categorizing the different types of anomalies detected. The second objective of the thesis is the generation of rules that can be interpreted and understood by experts for the detection of anomalies. For this, we have proposedan algorithm, Composition based Decision Tree (CDT), which automatically produces rules that can be adjusted and modified by experts. To do this, we have designed variable modeling of the detection patterns of remarkable points to label the timeseries. Based on the labeled time series, a decision tree is constructed by considering the nodes as compositions of patterns. Finally, the tree is converted into a set of decision rules, understandable by experts. We have also defined a quality measure for the rules produced. We tested the performance of CoRP and CDT with competitors, on real data and data from the literature (benchmarks). Both methods are effective in detecting multiple anomalies. The results have good precision offering a high detection rate with a low false-positive rate. This PhD was supported by the Management and Exploitation Service (SGE) ofthe Rangueil campus attached to the Rectorate of Toulouse and the researchis madein the context of the neOCampus project (Paul Sabatier University,Toulouse).
Complete list of metadata
Contributor : Ines Ben Kraiem <>
Submitted on : Wednesday, February 10, 2021 - 11:28:37 AM
Last modification on : Thursday, March 18, 2021 - 2:18:34 PM
Long-term archiving on: : Tuesday, May 11, 2021 - 6:28:08 PM


Files produced by the author(s)


  • HAL Id : tel-03137163, version 1


Inès Ben Kraiem. Détection d’Anomalies Multiples par Apprentissage Automatique de Règles dans les Séries Temporelles. Intelligence artificielle [cs.AI]. Université de Toulouse-Jean Jaurès, 2021. Français. ⟨tel-03137163⟩



Record views


Files downloads