Sylviane Gentil

Trends extraction and analysis for complex system monitoring and decision support

This paper presents an effective trend extraction procedure, based on a simple, yet powerful, representation. Its usefulness for complex system monitoring and decision support is illustrated by three examples. The method extracts semi-qualitative temporal episodes on-line, from any univariate time series. Three primitives are used to describe the episodes: {Increasing, Decreasing, Steady}. The method uses a segmentation algorithm, a classification of the segments into seven temporal shapes and a temporal aggregation of episodes. It acts on noisy data, without prefiltering. The first illustration is devoted to decision support in intensive care units. The signals contain information and noise at very different frequencies, and smoothing must not mask some interesting high-frequency data features. The second illustration is dedicated to a food industry process. On-line trends of key variables represent a very useful monitoring tool to control the end product quality despite high variations of raw materials at the input and a long delay. The last example concerns operator support and predictive maintenance. The results issued from a diagnostic module are complemented by the extrapolation of the key variable trends, which gives an idea of the time left to repair or reconfigure the process.

Brief Causal fault detection and isolation based on a set-membership approach

This paper presents a diagnostic methodology relying on a set-membership approach for fault detection and on a causal model for fault isolation. Set-membership methods are a promising approach to fault detection because they take into account a priori knowledge of model uncertainties and measurement errors. Every uncertain model parameter and/or measurement is represented by a bounded variable. In this paper, detection consists of verifying the membership of measurements to an interval. First order discrete time models are used and their output is explicitly computed with interval arithmetic. Fault isolation relies on a causal analysis and the exoneration principle, which allows focusing the consistency tests on simple local models. The isolation strategy consists of two steps: performing minimal tests found with the causal graph and determining on line additional relevant tests that reduce the final diagnosis. An application for a nuclear process is used in order to illustrate the methods efficiency.

Combining FDI and AI approaches within causal-model-based diagnosis

This paper presents a model-based diagnostic method designed in the context of process supervision. It has been inspired by both artificial intelligence and control theory. AI contributes tools for qualitative modeling, including causal modeling, whose aim is to split a complex process into elementary submodels. Control theory, within the framework of fault detection and isolation (FDI), provides numerical models for generating and testing residuals, and for taking into account inaccuracies in the model, unknown disturbances and noise. Consistency-based reasoning provides a logical foundation for diagnostic reasoning and clarifies fundamental assumptions, such as single fault and exoneration. The diagnostic method presented in the paper benefits from the advantages of all these approaches. Causal modeling enables the method to focus on sufficient relations for fault isolation, which avoids combinatorial explosion. Moreover, it allows the model to be modified easily without changing any aspect of the diagnostic algorithm. The numerical submodels that are used to detect inconsistency benefit from the precise quantitative analysis of the FDI approach. The FDI models are studied in order to link this method with DX component-oriented reasoning. The recursive on-line use of this algorithm is explained and the concept of local exoneration is introduced.

Recurrent neuro-fuzzy system for fault detection and isolation in nuclear reactors

This paper presents an application of recurrent neuro-fuzzy systems to fault detection and isolation in nuclear reactors. A general framework is adopted, in which a fuzzification module is linked to an inference module that is actually a neural network adapted to the recognition of the dynamic evolution of process variables and related faults. Process data is fuzzified in order to reason rather on qualitative than on quantitative values. The fuzzified attributes feed the neural network. Two different network topologies are tested over data simulated by a commissioned simulator of a nuclear reactor: a feed-forward topology and a recurrent topology, where the additional network inputs are considered as delayed activation of output units. The later approach shows better generalization performance for the detection and isolation of a number of security related faults. A graphic interface presents a qualitative representation of symptoms and diagnostic results by colored shades that evolve with time allowing a friendly and efficient communication with operators in charge of the process security.

Model-based causal reasoning for process supervision

Abstract This paper presents the activity of complex process supervision in terms of reasoning: to observe, to validate, to decide, to act, such are the human activities in control rooms. In order to build a system which helps operators in making their decisions, their different tasks are first analyzed, decomposed and shown to rise from causal reasoning. A model-based approach for causal reasoning is then emphasized; the specifications for the model are detailed. An implementation based on a signed directed graph representing causality, and on simple transfer functions describing the propagation of information through the graph, is finally proposed. The capacities of this modelling method to simulate, to explain and to give action advice are demonstrated on an industrial example.

Reformulation of parameter identification with unknown-but-bounded errors

In this paper a new formulation of the problem of identification of discrete-time linear models in the case of Unknown-But-Bounded errors is proposed. The bounds of the error at each sampling time are specified over a measurement noise rather than over an equation error, which is mainly motivated by experimental considerations. The method is particularly suitable for ARMAX models, as it accounts for the presence of uncertainties in the autoregressive terms.

Dynamic causal model diagnostic reasoning for online technical process supervision

Model-based diagnosis is founded on the construction of fault indicators. The methods proposed for this purpose generally represent the process by means of an extremely inflexible formalism that limits the scope of applications. Moreover, it is usually difficult and costly to develop precise mathematical models of complex plants. New and more flexible techniques intended notably to explain the observed behavior open new perspectives for fault detection and diagnosis. The diagnostic procedures for such plants are generally integrated into a supervisory system, and must therefore be provided with explanatory features that are essential interpretation and decision-making supports. Techniques based on causal graphs constitute a promising approach for this purpose. A causal graph represents the process at a high level of abstraction, and may be adapted to a variety of modeling knowledge corresponding to different degrees of precision in the underlying mathematical models. When the process is dynamic the causal structure must allow temporal reasoning. Lastly, because reasoning on real numbers is often used by human beings, fuzzy logic is introduced as a numeric-symbolic interface between the quantitative fault indicators and the symbolic diagnostic reasoning on them; it also provides an effective decision-making tool in imprecise or uncertain environments. An industrial application in the nuclear fuel reprocessing industry is presented.

Fuzzy reasoning in co-operative supervision systems

Abstract This paper considers a decision support system dedicated to fault detection and isolation from a human–machine co-operation point of view. Detection and isolation are based on different models of the process (non-linear and linear causal local models). Reasoning using real numbers is often used by human beings; fuzzy logic is introduced as a numerical-symbolic interface between the quantitative fault indicators and the symbolic diagnostic reasoning on them; it also provides an effective decision-making tool in imprecise or uncertain environments while managing model uncertainty, sensor imprecision and vague normal behavior limits. Fuzzy rules are modelled geometrically; fuzzy sets are represented as points in a description space. A prototype graphical interface with structural, causal and historical views gives complete information to the human operator. In such an interface, fuzziness is displayed as a colour palette evolving with time.

Hierarchical representation of complex systems for supporting human decision making

The work presented in this paper is devoted to intelligent on-line supervision tools. In the proposed approach, the human operator remains in the decision loop, at the highest level, and acts on the process. To help operators make decisions, process knowledge is represented with a model whose complexity can be adapted on line to the operation needs at the request of the operator. The model thus helps to focus only on the phenomena that are relevant at a given time. To give the model explanatory capacity, it is represented as a causal directed graph, and allows the representation of temporal phenomena, which is fundamental for dynamic monitoring. A hierarchical representation of the functional properties of the process is proposed. The conception of a hierarchy of causal models with a top-down analysis is discussed. Path algebra is used to construct a higher-level graph on-line at the request of the operator from the most detailed graph, while conserving the semantics of the latter. No intermediate level is defined a priori; only the highest and lowest level graphs are fixed: the others are constructed dynamically. Finally, a study of how graphs can convey information on the dynamics of the process for approximate temporal reasoning that is largely sufficient for supervision purposes is analyzed. An example of a causal graph hierarchy for a nuclear process illustrates the method. As a final point, the use of such causal graphs in advanced industrial supervision tools is considered.

SEXI: an expert identification package

Abstract This paper deals with an identification package. In the first part, the various classical off-line parameter estimation methods implemented are described. A graphic editor allows one to prepare the input-output data easily, which ensures a correct estimation. However, good methods and robust algorithms are not sufficient to guarantee a successful use of identification in industry. Another necessary component is the know-how of the person carrying it out. In the second part, the expert system SEXI (“Systeme EXpert en Identification”) is considered, which uses the software mentioned above so as to determine the model structure. SEXI behaves as a supervisor: it selects and runs the appropriate numerical module to obtain the quantitative results necessary for its reasoning and it iterates this process until a relevant model is found.