Gerard Escudero

Boosting applied to Word Sense Disambiguation

In this paper Schapire and Singers AdaBoost. MH boosting algorithm is applied to the Word Sense Disambiguation (WSD) problem. Initial experiments on a set of 15 selected polysemous words show that the boosting approach surpasses Naive Bayes and Exemplar-based approaches, which represent state-of-the-art accuracy on supervised WSD. In order to make boosting practical for a real learning domain of thousands of words, several ways of accelerating the algorithm by reducing the feature space are studied. The best variant, which we call LazyBoosting, is tested on the largest sense-tagged corpus available containing 192, 800 examples of the 191 most frequent and ambiguous English words. Again, boosting compares favourably to the other benchmark algorithms.

An Empirical Study of the Domain Dependence of Supervised Word Disambiguation Systems

This paper describes a set of experiments carried out to explore the domain dependence of alternative supervised Word Sense Disambiguation algorithms. The aim of the work is threefold: studying the performance of these algorithms when tested on a different corpus from that they were trained on; exploring their ability to tune to new domains, and demonstrating empirically that the Lazy-Boosting algorithm outperforms state-of-the-art supervised WSD algorithms in both previous situations.

A comparison between supervised learning algorithms for word sense disambiguation

This paper describes a set of comparative experiments, including cross-corpus evaluation, between five alternative algorithms for supervised Word Sense Disambiguation (WSD), namely Naive Bayes, Exemplar-based learning, SNoW, Decision Lists, and Boosting. Two main conclusions can be drawn: 1) The LazyBoosting algorithm outperforms the other four state-of-the-art algorithms in terms of accuracy and ability to tune to new domains; 2) The domain dependence of WSD systems seems very strong and suggests that some kind of adaptation or tuning is required for cross-corpus application.

Simultaneous fault diagnosis in chemical plants using support Vector Machines

Abstract One of the main limitations of the current plant supervisory control systems is the correct management of multiple simultaneous faults, which is crucial for supporting plant operators decision-making. In this work, Support Vector Machines (SVM) are used because of its proved efficiency dealing with multiclass problems in other technical areas. A Fault Diagnosis System has been developed implementing a multilabel approach using SVM and has been tested addressing a difficult diagnosis problem widely studied in the literature, the Tennessee Eastman process. Successful results have been obtained when diagnosing up to four simultaneous faults. These very first results are very promising since they have been achieved without any data processing or parameter tuning. Furthermore, they have been obtained just using training sets consisting of single faults, thus proving the achievement of a very powerful learning capacity.

Dynamic kriging based fault detection and diagnosis approach for nonlinear noisy dynamic processes

This paper presents a hybrid approach to improve data-based Fault Detection and Diagnosis (FDD). It is applicable to nonlinear dynamic noisy processes, operated under time-varying inputs. The method is based on the combination of kriging models and Pattern Recognition Techniques. A set of Multivariate Dynamic Kriging-based predictors (MDKs) is built and used to estimate the process dynamic behavior, while static kriging models are used to smooth the eventually noisy process outputs. The estimated and the actual smoothed outputs are compared, taking advantage of the higher capacity of the residual patterns generated in this way to characterize the process state. The performance of the method is illustrated through its application to a well-known benchmark case study, for which the FDD performance has been significantly improved. This improvement is consistently maintained in different dynamic operating conditions and faulty situations, including scenarios with modified fault severities and fault styles.

Fault diagnosis based on support vector machines and systematic comparison to existing approaches

Abstract An innovative data based fault diagnosis system (FDS) using Support Vector Machines (SVM) is applied on a standard chemical process. Besides its simpler design and implementation, this technique allows dealing better with complex and large data sets. For that reason, it was expected to improve usual pattern classifiers performance reported in chemical engineering literature, such as artificial neural networks or PCA modeling techniques. In order to compare results with previously reported works, a standard case study such as the Tennessee Eastman (TE) process benchmark was considered and SVM achieved consistent and promising results. Besides, the difficulties encountered when comparing the results reported are discussed and a FDS comparison methodology is proposed based on reliability and accuracy of each FDS. In that sense, this study establishes a reference framework for future comparisons.

Optimal Features Selection for Designing a Fault Diagnosis System

Abstract Fault diagnosis (FD) using data-driven methods is essential for monitoring complex process systems, but its performance is severely affected by the quality of the used information. Additionally, processing huge amounts of data recorded by modern monitoring systems may be complex and time consuming if no data mining and/or pre-processing methods are employed. Thus, features selection for FD is advisable in order to determine the optimal subset of features/variables for conducting statistical analyses or building a machine-learning model. In this work, features selection are formulated as an optimization problem. Several relevancy indices, such as Maximum Relevance (MR), Value Difference Metric (VDM), and Fit Criterion (FC), and redundancy indices such as Minimum Redundancy (mR), Redundancy VDM (RVDM), and Redundancy Fit Criterion (RFC) are combined to determine the optimal subset of features. Another approach of features selection is based on the optimal performance of the classifier, which is achieved by a classifier wrapped with genetic algorithm. Efficiency of this strategy is explored considering different classifiers, namely Support Vector Machine (SVM), Decision Tree (DT), K-Nearest Neighbours (KNN) Classifier and Gaussian Naive Bayes (GNB). A Genetic algorithm (GA), as a Derivative Free Optimization (DFO) technique, has been used due to the robustness to deal with different kinds of problems. The optimal subset of obtained features has been tested with SVM, DT, KNN, and GNB for the Tennessee-Eastman process benchmark with 19 classes. Results show that, when the performance of the classifier is used as the objective function the wrapper method obtains the best features set.

Modeling and Simulation of Complex Nonlinear Dynamic Processes Using Data Based Models: Application to Photo-Fenton Process

This paper investigates data based modelling of complex nonlinear processes, for which a first principle model useful for process monitoring and control is not available. These empirical models may be used as soft sensors in order to monitor a reaction’s progress, so reducing expensive offline sampling and analysis. Three different data modelling techniques are used, namely Ordinary Kriging, Artificial Neural Networks and Support Vector Regression. A simple case is first used to illustrate the problem, assess and validate the modelling approach, and compare the modelling techniques. Next, the methodology is applied to a photo–Fenton pilot plant to model and predict the reaction progress. The results show promising accuracy even when few training points are available, which results in huge savings of time and cost of the experimental work.

Incremental Learning Fault Detection Algorithm Based on Hyperplane-Distance

Traditional methods for Fault Detection and Diagnosis (FDD) usually, consider that processes operate under a single steady condition, but because of several reasons (e.g.: equipment aging), operation conditions of industrial processes change continuously in practice. Under these new circumstances, the use of the originally tuned FDD system would cause false alarms and will reduce the fault classification performance. In this study, the Hyperplane-Distance Support Vector Machine (HD-SVM) method is exploited for process FDD to maintain FDD performance when it decays because of the ageing. Its effectiveness is shown through simulation studies on a CSTR reactor, for which an aging term is simulated by progressively decreasing the heat transfer coefficient (5%). This aging will lead to reduce the classification performance accordingly. Next, performance of HD-SVM, Traditional Incremental Learning (TIL) and Non-Incremental Learning (NIL) (using all data) are compared. The HD-SVM incremental learning is shown to reduce the training time of the classifier, while increasing the accuracy of the classifier. Therefore, HD-SVM is shown to cover the weaknesses of Traditional incremental learning algorithms to lose possible information and to improve classification performance in process FDD.

A promising OPC-based computer system applied to fault diagnosis

Abstract Fault detection and diagnosis is a challenging problem for plant economics and safety. In this context, a promising OPC-based modular architecture for a Fault Diagnosis System (FDS) is designed and implemented. This FDS has been validated by performing on-line real-time diagnosis on a simulated process. The modular architecture allows openly connecting a simulator or a real process via OPC. The Tennessee Eastman Process (TEP) is used as data generator and case study, so that several abnormal operation conditions can be diagnosed by the system. The proposed architecture is discussed regarding the integration of future modules for the timely adoption of appropriate corrective actions.