Brigitte Chebel-Morello

A mix method of knowledge capitalization in maintenance

This paper deals with knowledge capitalization in maintenance especially in diagnosis and repair of industrial equipments. The goal is to propose a method of knowledge capitalization in order to develop a decision support system for maintenance operators. The knowledge capitalization cycle was adopted as the underlying principle. It consists of four principal steps: detect, preserve, capitalize and actualize the strategic knowledge. Different knowledge management tools and methods that can be used in the cycle are reviewed. We propose a mix method of knowledge capitalization in maintenance. This method integrates a representation and a reasoning model both completing each other and suitable to represent and manipulate the domain knowledge. The knowledge representation model using unified modelling language (UML) diagram proposes different domain models based on maintenance analysis to guide the domain expertise. The reasoning model uses the case-based reasoning which allows the manipulation of represented domain knowledge. Finally, the method is implemented on the pallet transfer system Sormel in the context of Proteus e-maintenance platform.

Direct Remaining Useful Life Estimation Based on Support Vector Regression

Prognostics is a major activity in the field of prognostics and health management. It aims at increasing the reliability and safety of systems while reducing the maintenance cost by providing an estimate of the current health status and remaining useful life (RUL). Classical RUL estimation techniques are usually composed of different steps: estimations of a health indicator, degradation states, a failure threshold, and finally the RUL. In this work, a procedure that is able to estimate the RUL of equipment directly from sensor values without the need for estimating degradation states or a failure threshold is developed. A direct relation between sensor values or health indicators is modeled using a support vector regression. Using this procedure, the RUL can be estimated at any time instant of the degradation process. In addition, an offline wrapper variable selection is applied before training the prediction model. This step has a positive impact on the accuracy of the prediction while reducing the computational time compared to existing indirect RUL prediction methods. To assess the performance of the proposed approach, the Turbofan dataset, widely considered in the literature, is used. Experimental results show that the performance of the proposed method is competitive with other existing approaches.

PETRA: process evolution using a TRAce-based system on a maintenance platform

To meet increasing needs in the field of maintenance, we studied the dynamic aspect of process and services on a maintenance platform, a major challenge in process mining and knowledge engineering. Hence, we propose a dynamic experience feedback approach to exploit maintenance process behaviors in real execution of the maintenance platform. An active learning process exploiting event log is introduced by taking into account the dynamic aspect of knowledge using trace engineering. Our proposal makes explicit the underlying knowledge of platform users by means of a trace-based system called “PETRA”. The goal of this system is to extract new knowledge rules about transitions and activities in maintenance processes from previous platform executions as well as its user (i.e. maintenance operators) interactions. While following a Knowledge Traces Discovery process and handling the maintenance ontology IMAMO, “PETRA” is composed of three main subsystems: tracking, learning and knowledge capitalization. The capitalized rules are shared in the platform knowledge base in order to be reused in future process executions. The feasibility of this method is proven through concrete use cases involving four maintenance processes and their simulation.

Reutilization of diagnostic cases by adaptation of knowledge models

This paper deals with design of knowledge oriented diagnostic system. Two challenges are addressed. The first one concerns the elicitation of expert practice and the proposition of a methodology for developing four knowledge containers of case based reasoning system. The second one concerns the proposition of a general adaptation phase to reuse case solving diagnostic problems in a different context. In most cases, adaptation methods are application-specific and the challenge in this work is to make a general adaptation method for the field of industrial diagnostics applications. This paper is a contribution to fill this gap in the field of fault diagnostic and repair assistance of equipment. The proposed adaptation algorithm relies on hierarchy descriptors, an implied context model and dependencies between problems and solutions of the source cases. In addition, one can note that the first retrieved case is not necessarily the most adaptable case, and to take into account this report, an adaptation-guided retrieval step based on a similarity measure associated with an adaptation measure is realized on the diagnostic problem. These two measures allow selecting the most adaptable case among the retrieved cases. The two retrieval and adaptation phases are applied on real industrial system called Supervised industrial system of Transfer of pallets (SISTRE).

RUL prediction based on a new similarity-instance based approach

Prognostics is a major activity of Condition-Based Maintenance (CBM) in many industrial domains where safety, reliability and cost reduction are of high importance. The main objective of prognostics is to provide an estimation of the Remaining Useful Life (RUL) of a degrading component/ system, i.e. to predict the time after which a component/system will no longer be able to meet its operating requirements. This RUL prediction is a challenging task that requires special attention when modeling the prognostics approach. In this paper, we proposes a RUL prediction approach based on Instance Based Learning (IBL) with an emphasis on the retrieval step of the latter. The method is divided into two steps: an offline and an online step. The purpose of the offline phase is to learn a model that represents the degradation behavior of a critical component using a history of run-to-failure data. This modeling step enables us to construct a library of health indicators (HIs) from run-to-failure data which are then used online to estimate the RUL of components at an early stage of life, by comparing their HIs to the ones of the library built in the offline phase. Our approach makes use of a new similarity measure between HIs. The proposed approach was tested on real turbofan data set and showed good performance compared to other existing approaches.

A formal ontology for industrial maintenance

The rapid advancement of information and communication technologies has resulted in a variety of maintenance support systems and tools covering all sub-domains of maintenance. Most of these systems are based on different models that are sometimes redundant or incoherent and always heterogeneous. This problem has lead to the development of maintenance platforms integrating all of these support systems. The main problem confronted by these integration platforms is to provide semantic interoperability between different applications within the same environment. In this aim, we have developed an ontology for the field of industrial maintenance, adopting the METHONTOLOGY approach to manage the life cycle development of this ontology, that we have called IMAMO Industrial MAintenance Management Ontology. This ontology can be used not only to ensure semantic interoperability but also to generate new knowledge that supports decision making in the maintenance process. This paper provides and discusses some tests so as to evaluate the ontology and to show how it can ensure semantic interoperability and generate new knowledge within the platform.

An ontology-based model for providing semantic maintenance

Abstract Maintenance is becoming more and more crucial in Asset Lifecycle Management information models. Issues such as collecting, handling and using the data of the asset produced during the assets lifecycle in a lean and efficient manner are on top of todays research. In this work we combine the benefits of two previous developed models and we develop a model for Semantic Maintenance. The first model we are based on is the PROMISE semantic object model which was made for supporting Closed-Loop Product Lifecycle Management. The second model is the semantic model of e-maintenance developed in PROTEUS project. The new model described in this paper is named “SMAC-Model”. Its aim is to provide advanced maintenance services as well as feedback for the Beginning of Life and input for End of Life. The model is generic and may be used in various Asset Lifecycle Management cases. It is developed to facilitate complex physical assets and to work in industrial environment.

Remaining useful life estimation based on discriminating shapelet extraction

In the Prognostics and Health Management domain, estimating the remaining useful life (RUL) of critical machinery is a challenging task. Various research topics including data acquisition, fusion, diagnostics and prognostics are involved in this domain. This paper presents an approach, based on shapelet extraction, to estimate the RUL of equipment. This approach extracts, in an offline step, discriminative rul-shapelets from an history of run-to-failure data. These rul-shapelets are patterns that are selected for their correlation with the remaining useful life of the equipment. In other words, every selected rul-shapelet conveys its own information about the RUL of the equipment. In an online step, these rul-shapelets are compared to testing units and the ones that match these units are used to estimate their RULs. Therefore, RUL estimation is based on patterns that have been selected for their high correlation with the RUL. This approach is different from classical similarity-based approaches that attempt to match complete testing units (or only late instants of testing units) with training ones to estimate the RUL. The performance of our approach is evaluated on a case study on the remaining useful life estimation of turbofan engines and performance is compared with other similarity-based approaches.

Feature selection for fault detection systems: Application to the Tennessee Eastman Process

A fault detection system based on data mining techniques is developed in this work. A novel concept of feature selection based on the k-way correlation is introduced and used to detect redundant measures relevant features (strong and weak relevant) and/or redundant ones is introduced. The authors propose to apply STRASS, a contextual filter algorithm to identify the relevant features on simulated data collected from the Tennessee Eastman chemical plant simulator. In effect the TEP process has been studied in many articles and three specific faults are not discriminated with a myopic filter algorithm. The results obtained by STRASS are compared to those obtained with reference feature selection algorithms. The features selected by STRASS reduced the data correlation and the overall misclassification for the testing set using K-nearest-neighbor decreased further to 0.8%.

A new contextual based feature selection

The pre processing phase is essential in knowledge data discovery process. We study too particularly the data filtering in supervised context, and more precisely the feature selection. Our objective is to permit a better use of the data set. Most of filtering algorithm use myopic measures, and give bad results in the case of the features correlated part by part. Consequently in the first time, we build two new contextual criteria. In the second part we introduce those criteria in an algorithm similar to the greedy algorithm. The algorithm is tested on a set of benchmarks and the results were compared with five reference algorithms: Relief, CFS, Wrapper (C4.5), consistancySubsetEval and GainRatio. Our experiments have shown its ability to detect the semi-correlated features. We conduct extensive experiments by using our algorithm like pre processing data for decision tree, nearest neighbours and naive Bays classifiers.