Search

The Oncotype DX (ODX) breast cancer assay is the worldwide most common and used Gene Expression Profiling (GEP) test. This ODX assay has a great impact on Adjuvant ChemoTherapy (ACT) decision. However, many standard approaches have been proposed and suggested to practitioners. The accuracy of such methods never reached the highest level. This paper deals with the Breast Cancer Computer Aided Diagnosis (BC-CAD) based on a Deep Constructive Neural Network used for the Recurrence Score (RS) prediction of the ODX assay. The proposed ConstDeepNet algorithm was tested to build two classifiers. In the first architecture, a ”one against all” structure is used where one Deep Neural Network is built for each class. In the second architecture, one DNN is used for the three classes. The proposed BC-CAD algorithm is tested on a real data-set and exhibits good performance. The study data set contains 92 cases carcinoma mammary luminal B with available Oncotype DX test results from 2012 to 2017 taken from the Georges Francois Leclerc cancer centre and the North Trevenans County Hospital located respectively in Dijon and Belfort in France.

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.

Breast cancer is the second most common cancer (after lung cancer) that affect women both in the developed and less developed countries. Nowadays, using the Computer Aided Diagnosis (CAD) techniques becomes a necessity for several reasons: assisting and improving physicians, speed in data processing, harmonization and aid of diagnosis, better access to advanced online-medicine. Recently, several works about Breast Cancer Computer Aided Diagnosis (BC-CAD) have been published, and Neural Networks techniques, particularly deep architectures represent a significant part of these works. In this paper, we prpose a BC-CAD based on joint variable selection and a Constructive Deep Neural Network “ConstDeepNet”. A feature variable selection method is applied to decrease the number of inputs used to train a Deep Learning Neural Network. Experiments have been conducted on two datasets, the Wisconsin Breast Cancer Dataset (WBCD) and real data from the north hospital of Belfort (France) to predict the recurrence score of the Oncotype DX. Consequently, the use of joint variable algorithm with ConstDeepNet outperforms the use of the Deep Learning arechitecture alone.

Condition monitoring and fault diagnosis of rolling element bearings (REBs) are at present very important to ensure the steadiness of industrial and domestic machinery. According to the non-stationary and non-linear characteristics of REB vibration signals, feature extraction method is based on empirical mode decomposition (EMD) energy entropy in this paper. A mathematical analysis to select the most significant intrinsic mode functions (IMFs) is presented. Therefore, the chosen features are used to train an artificial neural network (ANN) to classify bearings defects. Experimental results indicated that the proposed method based on run-to-failure vibration signals can reliably categorize bearing defects. Using a proposed health index (HI), REB degradations are perfectly detected with different defect types and severities. Experimental results consist in continuously evaluating the condition of the monitored bearing and thereby detect online the severity of the defect successfully. This paper shows potential application of ANN as effective tool for automatic bearing performance degradation assessment without human intervention.

International audience; In fault detection systems, massive amount of data gathered from the life-cycle of equipment is often used to learn models or classifiers that aims at diagnosing different kind of errors or failures. Among this huge quantity of information, some features (or sets of features) are more correlated with the kind of failures than others. The presence of irrelevant features might affect the performance of the classifier. To improve the performance of a detection system, feature selection is hence a key step. We propose in this paper an algorithm named STRASS, that aims at detecting relevant features for classification purposes. In certain cases, when there exists a strong correlation between some features and the associated class, classical feature selection algorithms fail at selecting the most relevant features. In order to cope with this problem, STRASS algorithm makes use of k-way correlation between features and the class to select relevant features. To assess the performance of STRASS, we apply it on simulated data collected from the Tennessee Eastman chemical plant simulator.The Tennessee Eastman process (TEP) has been used in many fault detection studies and three specific faults are not well discriminated with classical algorithms. The results obtained by STRASS are compared to those obtained with reference feature selection algorithms. We show that the features selected by STRASS always improve the performance of a classifier compared to the whole set of original features and that the obtained classification is better than with most of the other feature selection algorithms.

Remaining Useful Life (RUL) prediction accurately is an imperative industrial challenge. In this sense, the monitoring of lithium-ion battery is very significant for planning repair work and minimizing unexpected electricity outage. As the RUL estimation is essentially a problem of pattern recognition, the most valuable feature extraction techniques and more accurate classifier are needed to obtain higher prognostic effectiveness. Consequently, this paper discusses the importance of non linear feature reduction techniques for more adequate prognosis feature data base. For more convenience, the isometric feature mapping technique (ISOMAP) is used to reduce some features extracted from lithium-ion batteries, with different health states, in both modes of charge and discharge. Experimental results show that non linear feature reduction techniques are very promising to provide some trend parameters for industrial prognostic.

International audience; Many models and methodologies in order to predictthe remaining useful life (RUL) of a critical component areinvestigated nowadays. However, estimating remaining useful lifeof multi-component systems is still an under explored area,especially when there are interdependencies among the componentsof these systems. Practically, prognostics can be quitecomplicated when there is absence of prior knowledge aboutthese interactions. To optimize the availability and reliabilityof the system, it is required to embed a dependency modelthat implements component strength and the component’s healthstatus. In this paper, a novel approach is proposed in orderto emphasize the importance of interactions between complexsystem’s components in the RUL’s calculating. The effectivenessof the approach is judged by applying it to numerical studies inorder to estimate system’s remaining useful life.

In this paper, a conceptual model of maintenance process is proposed in the aim of releasing different types of help (decision-help systems) to maintenance actors. The maintenance domain is defined and modelled with the aid of knowledge management. The object-oriented representation was chosen in order to propose real domain model for development of information system or prototype. This system should comply with decision support and allow technico-economical knowledge capitalisation.

International audience; Accurate remaining useful life (RUL) prediction of critical assets is an important challenge in condition based maintenance to improve reliability and decrease machine's breakdown and maintenance's cost. Bearing is one of the most important components in industries which need to be monitored and the user should predict its RUL. The challenge of this study is to propose an original feature able to evaluate the health state of bearings and to estimate their RUL by Prognostics and Health Management (PHM) techniques. In this paper, the proposed method is based on the data-driven prognostic approach. The combination of Simplified Fuzzy Adaptive Resonance Theory Map (SFAM) neural network and Weibull distribution (WD) is explored. WD is used just in the training phase to fit measurement and to avoid areas of fluctuation in the time domain. SFAM training process is based on fitted measurements at present and previous inspection time points as input. However, the SFAM testing process is based on real measurements at present and previous inspections. Thanks to the fuzzy learning process, SFAM has an important ability and a good performance to learn nonlinear time series. As output, seven classes are defined; healthy bearing and six states for bearing degradation. In order to find the optimal RUL prediction, a smoothing phase is proposed in this paper. Experimental results show that the proposed method can reliably predict the RUL of rolling element bearings (REBs) based on vibration signals. The proposed prediction approach can be applied to prognostic other various mechanical assets.

Case base maintenance.Prototyping.Instance reduction.Competence and performance optimization.Structuring and updating a case base. To avoid performance degradation and maintain the quality of results obtained by the case-based reasoning (CBR) systems, maintenance becomes necessary, especially for those systems designed to operate over long periods and which must handle large numbers of cases. CBR systems cannot be preserved without scanning the case base. For this reason, the latter must undergo maintenance operations.The techniques of case base's dimension optimization is the analog of instance reduction size methodology (in the machine learning community). This study links these techniques by presenting case-based maintenance in the framework of instance based reduction, and provides: first an overview of CBM studies, second, a novel method of structuring and updating the case base and finally an application of industrial case is presented.The structuring combines a categorization algorithm with a measure of competence CM based on competence and performance criteria. Since the case base must progress over time through the addition of new cases, an auto-increment algorithm is installed in order to dynamically ensure the structuring and the quality of a case base. The proposed method was evaluated through a case base from an industrial plant. In addition, an experimental study of the competence and the performance was undertaken on reference benchmarks. This study showed that the proposed method gives better results than the best methods currently found in the literature.

International audience; Thanks to ICT, Web emergency and Internet, the achievement of maintenance services and monitoring can be performed automatically, remotely and through various distributed information systems. Hence the emergence of the concept of services offered through maintenance architectures, ranging from autonomic systems to integrated systems where knowledge management, cooperation and collaboration are vital to any operation. Into this context, new services like intelligent maintenance, self maintenance, etc are required. To this end, a new concept called s-maintenance is emerged. This concept defines a new generation of maintenance systems founded on a knowledge based system. While existing systems don't respond to the characteristics of this new generation of systems, we design in this paper an architecture of a maintenance component based system respecting the characteristics of s-maintenance. Each component in the system is defined to respond to one or many characteristics of this concept.

Accurate Remaining Useful Life (RUL) prediction of critical assets is an important challenge in condition based maintenance to improve reliability and to decrease machine's breakdown and maintenance's cost. Bearing is one of the most important components in industries that need to be monitored and the user should predict its RUL. The challenge of this study is to propose a new strategy for RUL feature extraction. The proposed methodology provides better features in term of monotonicity. This specification ensures a better RUL prediction by comparing the test degradation features to the library of instance. Experimental results show that the proposed methodology is very promising for RUL estimation by industry.

Bearings are important assets for most industrial applications. The non-destructive diagnosis of these elements needs an accurate and reliable acquisition of its dynamic vibration signals affected by noise and the other part of system such as gears, shafts, etc. Empirical mode decomposition is an advanced signal processing tool for bearing fault feature extraction. In this paper, empirical mode decomposition is used to decompose non-linear and non-stationary bearing vibration signals into several stationary intrinsic mode functions and the empirical mode decomposition energy entropy is computed for each intrinsic mode function. Moreover, principal component analysis and linear discriminant analysis are used for feature reduction. Based on the Fisher's criterion, experimental results show that linear discriminant analysis features are highlighted compared to principal component analysis features and original empirical mode decomposition features for bearing fault diagnosis as type (inner race, outer race, rolling element) and severity (normal, degraded, faulting).

International audience; In the Prognostics and Health Management domain, estimating the remaining useful life (RUL) of critical machinery is a challenging task. Various research topics as data acquisition and processing, fusion, diagnostics, prognostivs and decision are involved in this domain. This paper presents an approach for estimating the Remaining Useful Life (RUL) of equipments based on shapelet extraction and characterization. This approach makes use in a first step of an history of run-to-failure data to extract discriminative rul-shapelets, i.e. shapelets that are correlated with the RUL of the considered equipment. A library of rul-shapelets is extracted from this step. Then, in an online step, these rul-shapelets are compared to different test units and the ones that match these units are used to estimate their RULs. This approach is hence different from classical similarity-based approaches that matches the test units with training ones. Here, discriminative patterns from the training set are first extracted and then matched to test units. The performance of our approach is assessed on a data set coming from a previous PHM Challenge. We show that this approach is efficient to estimate the RUL compared to other approaches.

In the most of industrial and domestic applications bearings present important assets. The diagnostic of these elements needs accurate and reliable acquisition of its dynamic vibration signals affected by noise and other part of system such as gears, bars... Empirical Mode Decomposition (EMD) is a new signal processing method used to decompose non-stationary and non-linear vibration bearing signals into several stationary empirical mode components called Intrinsic Mode Functions (IMF). For each IMF, the energy entropy mean is computed. This technique is compared to the most used statistical features (RMS, Kurtosis) using a characterization degree. Experimental results show that time domain feature extraction is effective for bearing fault feature extraction as type (inner race, outer race, rolling element) and severity (normal, degraded, faulting). The choice of the most significant IMFs is also discussed in this paper.

International audience; In case of a radiological emergency situation involving accidental human exposure, a dosimetry evaluation must be established as soon as possible. In most cases, this evaluation is based on numerical representations and models of victims. Unfortunately, personalised and realistic human representations are often unavailable for the exposed subjects. However, accuracy of treatment depends on the similarity of the phantom to the victim. The EquiVox platform (Research of Equivalent Voxel phantom) developed in this study uses Case-Based Reasoning (CBR) principles to retrieve and adapt, from among a set of existing phantoms, the one to represent the victim. This paper introduces the EquiVox platform and the Artificial Neural Network (ANN) developed to interpolate the victim's 3D lung contours. The results obtained for the choice and construction of the contours are presented and discussed.

International audience; Prognostics and Health Management platforms are founded on Condition Based Maintenance process. Major works on this topic are focused on the diagnostic and prognostic modules and neglect the Decision Support Module which must be investigated to give more efficiency to PHM platforms. To improve this module with intelligence and rapidity we propose in this work to integrate a Trace Based System (TBS) into the decision support module. This TBS provides three main services (Traceability, self-learning and self-management) which are developed in this work.

International audience; In case of a radiological emergency situation involving accidental human exposure, a dosimetry evaluation must be established as soon as possible. In most cases, this evaluation is based on numerical representations and models of subjects. Unfortunately, personalised and realistic human representations are often unavailable for the exposed subjects. However, accuracy of treatment depends on the similarity of the phantom to the subject. The EquiVox platform (Research of Equivalent Voxel phantom) developed in this study uses Case-Based Reasoning principles to retrieve and adapt, from among a set of existing phantoms, the one to represent the subject. This paper introduces the EquiVox platform and Artificial Neural Networks developed to interpolate the subject’s 3D lung contours. The results obtained for the choice and construction of the contours are presented and discussed.