Search

International audience; System diagnosis has been of a great interest for all aspects of industrial processes more precisely to gain in quality. It is based essentially on the analysis of the links between the variables of a system and more precisely on the changes of the relations between these variables, which testify to the presence of faults or anomalies. For that purpose, data modeling is the process of finding a mathematical expression that provides a good fit between given finite sample values of the independent variables and the associated values of the dependent variables of the process.The aim of this paper is to detect and, above all, localize faults affecting a system with nonlinear behavior, when its model is not known a priori. An important part of the presentation is dedicated to the construction of fault indicators capable of locating faults, ie, recognizing the input or output of a system affected by a fault. The first part of this paper is devoted to how to predict the output of a nonlinear behavior system. The second part proposes a way for the detection and isolation of measurement faults based on the proposed prediction model. The relevance of the proposed technique, for modeling and system diagnosis, is illustrated on a simulated example in the context of SIMO and MIMO systems.

International audience; This paper suggests a method for damage detection and localization in pipelines. It was proposed to address the problem of the effects of environmental and operational conditions on the monitoring signals. It consists of establishing a model on demand of the reference state of the structure whenever a new measured signal is presented. This model is constructed by calculating sparse estimation of the actual signal. In this case, damage detection is accomplished by using the estimation error as a damage index. To localize the damage, the idea is to calculate the estimation error only on a sliding window over the signal from the damaged pipeline. The proposed method is evaluated on a pipeline segment placed in a facility where the temperature is the only environmental factor that varies during the monitoring period. Damage was created at the end of the monitoring period by removing material from the inside of the pipeline in order to simulate corrosion. Results have shown that the damage was detected successfully and localized with a minor error.

International audience; Artificial intelligence using machine learning techniques has become omnipresent in different field of studies. Structural health monitoring is an emerging field which should take advantage of these techniques in order to be efficient and more reliable. It aims mainly at detecting and localizing damage in structures automatically without the need of the interpretation of data (i.e. human analysis) as it is done in non-destructive techniques. In this work, some cases of study of detection and localization of damage in pipeline during a monitoring period will be presented. Two types of learning environment: stationary and non-stationary will be considered. In the latter, the variation of environmental and operational conditions could have an impact on the reference signals and hence, as these signals are generally compared to the actual signals, this will give rise to false alarm. To deal with this issue, we propose to build a model of reference data whenever a new measured signal is available. In the former, intelligent feature extraction and selection methods are considered to ensure a better damage detectability and to avoid overfitting. In the case of learning in stationary environment, reference data are supposed to be constant. These data was gathered on a tube where an artificial damage was placed on the surface. However, in non-stationary environment, data was collected on pipe segment where temperature fluctuates due to the weather changes. A corrosion-like damage was machined in this case in several steps in order to increase its size. In both cases, damage was detected successfully using the above mentioned machine learning techniques. Besides, a procedure for damage localization was developed in the second case. Further technical details will be provided in the oral presentation.

International audience; This paper addresses the problem of multiple fault isolation based on kernel principal component analysis and proposes a sparse fault estimation method to evaluate the reconstruction-based contribution. The fault magnitude estima- tion is here formulated as an optimization problem under non- negativity and sum-to-one constraints. A multiplicative iterative scheme and its initialization procedure are proposed to solve it. The effectiveness of the proposed method is demonstrated on the simulated continuous stirred tank reactor (CSTR) process.

International audience; This article proposes an analytical method for decomposing a dynamic nonlinear system into a multiple model form in order to reduce its complexity and to study more easily identification, stability analysis and controller design problems. The majority of existing methods are order reduction based techniques, which come with an information loss of the initial system, whereas the method proposed here avoids this particular loss. The multiple model constitutes an efficient tool to represent nonlinear systems. These are decomposed into several linear time invariant systems (LTI) which are weighted and aggregated that allows to benefit from important analysis tools. This method is applied to a simplified activated sludge reactor model.

International audience

International audience; This article deals with identification design for sensor fault diagnosis of a wastewater treatment station. First, we established a physical model of this station by simplifying and linearizing the model ASM1 (Henze et al., 1987). Then, we study the measurement values in order to find other redundancy equations. Multiple linear regression allowed the determination of a static black box type model, coherent with the physical model. The addition of temporal lags between the signals makes possible to improve the performances of this model. We show finally the use of this model for detection-localization of faults measurements.

International audience; Diagnosis method based on Principal Component Analysis (PCA) has been widely developed. However, this method deals only with data which are described by single-valued variables. The purpose of the present paper is to generalize the diagnosis method to interval PCA. The fault detection is performed using the new indicator [SPE]. To identify the faulty variables, this work proposes a new method based on the reconstruction principle to this indicator. It aims to solve an interval linear system to obtain the reconstructed variables. The analysis of the reconstruction conditions permits to determine the useful directions. Then, the residuals structuring according to these directions allows to identify the set of faulty variables. This new diagnosis method based on interval PCA model is validated by a simulation example.

International audience; In this paper, we propose a method to model a fleet of identical machines. These latest may work in the same or in different operating conditions. Firstly, the method consists in estimating a linear model, using PCA, based on the data collected on the machine itself. Secondly, the common parts of the models of the different machines are identified. New models for the machines are then generated taking into account the identified common parts. An academic example is finally presented to illustrate the results of the method.

International audience; Multi-Task Learning (MTL) has become an active research topic in recent years. While most machine learning methods focus on the learning of tasks independently, multi-task learning aims to improve the generalization performance by training multiple related tasks simultaneously. This paper presents a new approach to multi-task learning based on one-class Support Vector Machine (one-class SVM). In the proposed approach, we first make the assumption that the model parameter values of different tasks are close to a certain mean value. Then, a number of one-class SVMs, one for each task, are learned simultaneously. Our multi-task approach is easy to implement since it only requires a simple modification of the optimization problem in the single one-class SVM. Experimental results demonstrate the effectiveness of the proposed approach.

International audience

International audience; Supervision of nonlinear and complex processes is of great importance to industries as a means of achieving improved productivity and stable product quality. The advanced model-based condition monitoring methodologies, can contribute significantly to the achievement of these objectives. In this paper, a fuzzy evidential model based fault detection and diagnosis method is presented. The multi-model based symptom generation procedure is used to detect changes of the current process behavior. The diagnosis task is accomplished by an evidential approximate reasoning scheme to handle different kinds of uncertainty that are inherently present in many real word processes. The validity of the method is illustrated on the well-known benchmark of three tanks and different faults can be detected and isolated continuously, over all ranges of operation.

International audience; Recently, fault detection and process monitoring using principal component analysis (PCA) were studied intensively and largely applied to industrial process. PCA is the optimal linear transformation with respect to minimizing the mean squared prediction error. If the data have nonlinear dependencies, an important issue is to develop a technique which can take into account this kind of dependencies. Recognizing the shortcomings of PCA, a nonlinear extension of PCA is developed. This paper proposes an application for sensor failure detection and isolation (FDI) to an air quality monitoring network via nonlinear principal component analysis (NLPCA). The NLPCA model is obtained by using two cascade three layer RBF-Networks. For training these two networks separately, the outputs of the first network are estimated using principal curve algorithm [7] and the problem is transformed as two nonlinear regression problems.

International audience; The paper aims to describe the problem of modelling non-linear dynamic systems using a multiple model representation, also known in fuzzy set literature as Takagi-Sugeno model. The basic principle of this approach is to represent the system as an interpolation of locally valid models. Proposed identification algorithms concern the parameter estimation and the structure determination (local model orders, number of local models). Three applications of this approach to environmental system management are reported: modelling of a wastewater treatment plant, ozone concentration prediction is an urban area and finally modelling of rainfall-runoff relationship for sensor fault diagnosis in urban sewage network.

International audience; A multi-model provides a powerful tool fornon-linear dynamic systems identification. Theoreticalstudies have shown that any continuous function on acompact domain can be approximated arbitrarily well by afuzzy model. However, due to the curse of dimensionality,building a multi-model is a difficult task in practicewithout some precautions. Hence, only the simpleststructure fitting the training data must be chosen in orderto keep good generalisation capabilities of the model. Inthis paper, parsimony is obtained by removing redundantand/or useless rules. The proposed approach has beensuccessfully illustrated by simulation examples from literature.

International audience; In this paper, the model of ozone concentration in an urban area is considered. The mechanisms of the phenomenon are non-linear and time-varying. To solve the problem, we adopt a multiple model approach which is based on a decomposition principle. The main difficulty related to the identification of this structure concerns the partition of the feature space. The decomposition of the feature space along a decision tree is applied. The application to ozone concentration modelling is reported as well as the expected outcome.

International audience; For process control improvement, coherency of information supplied by sensors must first be ensured. Because of the presence of random and possibly gross errors, the model equations of the process are not generally satisfied. The problem of data reconciliation in order to satisfy the model constraints is considered in this article. The simultaneous presence of errors in process input and output measurements poses serious problem in the rectification of data. The proposed procedure to solve this problem involves the use of a special filter to estimate both the parameters, the states and the inputs of a process; smoothing of both estimations of the input and the output is increased by adding in the filter a variance term. Application is proposed for rainfall data validation in order to improve urban sewer network control

International audience; The aim of this paper is to compare the behaviour of the robust state estimator to that of the extended Kalman filter whose principle is to estimate simultaneously the state of the system and its unknown parameters through simulated numerical examples. These parameters are considered as random variables. Some results on robust state estimation for stochastic parameters system have been developed which take into account of the effects of the uncertain parameters. First, the two solutions on a problem of state estimation are compared. Second, the residuals derived from these filters are applied to sensor failure detection. Numerical examples are given to show the behaviour of the robust filters in presence of sensor failure via a number of simulated systems.

International audience; In this paper, practical aspects of non-linear black-box modeling applied to hydrological processes are investigated. The proposed structure consists in a set of local linear dynamic sub-models. Local models are combined by smooth interpolation into a global model. This approach is then used for modeling a rainfall-runoff on a real urban watershed which permits real-time outlet flow estimation.

International audience; Considering the great importance of sensor data in the study of environmental systems, efforts are focused on the validation problem. Although off-line sensor data validation is necessary for providing validated database to the end user, on-line failure detection and isolation (FDI) is essential for reliable system monitoring and management. Thus paper thus investigates the possibility of implementing on-line FDI techniques for monitoring an urban rain gauge network. A quantitative model-based approach is presented and the basis of a FDI scheme is described. Analytical middles between sensors are designed and involved in the generation of failure sensitive signals. These signals are the evaluated in order to detect failure occurrences, isolate the faulty sensors and generate alarms for monitoring operators. Results obtained with actual homogeneous rainfall data are provided as illustrations.

The objective of this work is to describe a numerical technique to identify parameters of a fuzzy model. When this model and the membership functions of the input variables are continuously differentiable, we show that the estimation of parameters can be performed with a two-levels hierarchical algorithm, comprising the estimation of the model parameters and the estimation of the membership functions parameters. The proposed algorithm is then applied to an example describing a non-linear system

International audience; Unsupervised learning in diagnosis of technological system is viewed as a clustering problem and a survey on fuzzy clustering is presented. Limitations of these techniques are discussed along with a review on present approaches to overcome these limitations. The paper concludes with directions for future works.