Xavier Berjaga

Partial least square/projection to latent structures (PLS) regression to estimate impact localization in structures

This paper presents results from the application of partial least squares/projection to latent structures (PLS) as a regression tool in order to estimate the localization of impacts in an aircraft structure using the strain wave produced by the impact and recorded by sensors attached to the structure. PLS is a technique that maximizes the covariance between the predictor matrix X and the predicted matrix Y for each component of the space. The main objectives of PLS are: to model X and Y, and to predict Y from X.The structure used in this work can be considered as a small scale version of a part of an aircraft wing. A total of 574 experiments were performed impacting the wing over its surface and receiving vibration signals from nine sensors. The data set (time history signal) is organized in a matrix to be used as predictors, while the predicted matrix is given by the real localization of the impact (x, y coordinates).Experiments are divided into four groups depending on their localization and probability of occurrence. A PLS model is built using three of these groups (X and Y) and tested using the remaining group. Results are presented, discussed and compared with other methods.

A framework for Case-Based Diagnosis of batch processes in the principal components space

This paper presents a framework for fault detection and diagnosis of batch processes based on the information directly gathered from sensors. First, a statistical model of the process is build using multiway principal component analysis (MPCA) for dimensionality reduction and fault detection tasks. Afterwards, a case-based reasoning (CBR) approach is used for fault diagnosis and for false alarm and missed detection reduction. This framework has been tested in two completely different fields: power quality monitoring for relative location of voltage sags and injection moulding processes for faulty sensor detection and diagnosis. Results obtained show that this framework presents a good performance and is general enough to be applied to any field, if the appropriate preprocess of the data is carried.

Classification of sags according to their origin based on the waveform similarity

A statistical method for classification of sags their origin downstream or upstream from the recording point is proposed in this work. The goal is to obtain a statistical model using the sag waveforms useful to characterise one type of sags and to discriminate them from the other type. This model is built on the basis of multi-way principal component analysis an later used to project the available registers in a new space with lower dimension. Thus, a case base of diagnosed sags is built in the projection space. Finally classification is done by comparing new sags against the existing in the case base. Similarity is defined in the projection space using a combination of distances to recover the nearest neighbours to the new sag. Finally the method assigns the origin of the new sag according to the origin of their neighbours.

Comparison of multivariate analysis techniques in plastic injection moulding process

In this paper, we present a comparison between several statistical discriminant analysis techniques applied to a plastic injection moulding process for monitoring quality of injected moulded parts. Comparison among different ways of training the system can provide useful conclusions about the behaviour of the different models in poor conditions. The goal of this paper is to establish a baseline for comparing the performance between different algorithms. A wide variety of research objectives throughout the literature makes it difficult to provide a feasible comparison between results. The evaluation is intended to provide detailed, empirical information on the effectiveness and impact of different model parameters on the performance of the different approaches. The pros and cons of the approaches used are discussed. In order to predict the quality of a plastic part, we extract a set of salient features that characterise an injection cycle and then match these features against a database of stored examples of predefined classes by using supervised classification. The database was created from 199 real plastic injections without any overlap between training and testing datasets.

Granularity determination of activated sludge through on-line profiles by means of case-based reasoning

Aerobic granulation from floccular sludge is difficult to detect in first stages with the naked eye. This work proposes a combination of multi-way principal components and case-based reasoning to predict the granulation state of a sequencing batch reactor, based solely on the on-line registered profiles of common sensors (i.e. pH, dissolved oxygen and oxidation-reduction potential). The methodology is able to discriminate between two active sludge granularities (floccular and granular). Two different scenarios are presented: one in which both granularities are present, and another scenario for which the granular state is not initially available. Analysis reported pH as the key variable in the transition between both states according to its variation, and that, in general, the granularity of the process can be correctly predicted at the end of the anaerobic phase. This methodology improves process monitoring capabilities during granulation and is an on-line alternative to a microscope analysis before the batch release.

Multiway partial least square (MPLS) to estimate impact localization in structures

This paper presents results from the application of Multiway Partial Least Square (MPLS) as a regressor tool in order to estimate the localization of impacts in an aircraft structure. MPLS is a technique that maximizes the covariance between the predictor matrix X and the predicted matrix Y for each component of the space. The structure can be considered as a small scale version of part of a wing aircraft. 574 experiments were performed impacting the wing over its surface and receiving vibration signals from nine sensors. Experiments are divided in four groups depending on their localization and probability of occurrence. A PLS model is build using three of these groups and tested using the remaining group. Results are presented, discussed and compared with results of other methods.

Statistical fault detection and reconstruction of sensors of the Ariane engine

A statistical multivariate model with minimum variance of the reconstruction error (VRE) has been tested with data captured in several tests of the Arianes engine. Only reconstructible sensors according to the VRE criteria are included in the statistical model and the same criterion is used to determine the number of principal components to retain when creating the model with data acquired during normal operating conditions. The resulting model is used for fault detection and reconstruction by projecting new acquired data in the space defined by that model and their statistical limits. Results show that real faulty situations can be correctly reconstructed when fault directions (sensors) are known.

A Case-Based Centred Approach for Rapid Manufacturing: Definitions

In this paper the conceptual definition of a case-based reasoning methodology to work in the rapid manufacturing of customised medical implants is explained. The aim is to define accurately the case representation and the core concepts that will be used in the decision support and control tasks of the process.