(302g) Enhancing Fault Diagnosis by Incorporation of Intelligent Filtering Knowledge

The importance and difficulties of incipient fault detection and isolation is well known in the process industry. The process operators face a difficult task when they have to interpret the existence of a fault in the plant by examining time sequenced data, among the hundreds of variables of modern chemical processes. It is even more complicated when they have to find out the root cause of the detected anomalies.

The difficulties encountered to develop accurate process models, the highly coupled nature of chemical processes signals and the big amount of stored data, have made the process history based methods one of the most successfully applied solution approaches in fault diagnosis.

Different data based methodologies have been developed to identify patterns from process data in presence of faults. These pattern recognition techniques, as artificial neural networks (ANN) or fuzzy logic systems (FLS) are suited process history methods to automatically interpret the variables values coming from plant. As they are based on historical data, it is of crucial importance to achieve the most homogeneous and representative data set for each possible state of the plant, to ensure that the diagnosis system learns the most simple pattern, which should result in an improved on-line resolution of the fault diagnosis system (FDS). Nevertheless, the presence of random errors in process data causes that the contribution of each variable to a faulty state is not always as clear as it would be desired, resulting in data sets that are not sufficiently consistent and representative of each state.

On the other hand, the usefulness of data filtering has been widely described in the literature as a monitoring data pre-processing and gross error detection technique. However the influence of such data pre-processing methodologies on a later data-based FDS has not been studied. Such data de-noising could cause two opposite effects on the FDS performance that should be clarified:

1) A greater homogeneity of variables after data filtering would help to identify more clearly the actual state while 2) a too deep filtering could remove essential information for a pattern recognizer.

As a data based FDS, an automatically designed rules based FLS is implemented (E. Musulin et al., 2006). Then, a carefully study of filtering techniques applied to increase the diagnosis effectiveness is carried out in order to find out if the FDS performance is damaged or improved. Thus, in this work, data filtering techniques are not just used to improve detection as it has been pointed out in most of literature (Bakshi, B.R., 1998, Wang, D. & Romagnoli, J.A., 2005), but also to get better diagnosis.

With that purpose, different kinds of filtering techniques including popular filters used in industry, as Exponential Weighted Moving Average (EWMA) or Simple Moving Average (SMA) and more recent wavelet based approaches (Tona et al. (2005)) were tested on the previously commented data based FDS.

Firstly, data coming from plant were treated by any of the mentioned filtering techniques. Then, these filtered data were introduced in the FDS to check how such treatments affect the diagnosis performance.

In order to compare the above mentioned filtering techniques, two FDS performance measures, accuracy and reliability, are proposed as two quantitative comparative tools. The methodology was evaluated in the Tennessee Eastman benchmark, considering the first 14 faults proposed in the original paper (Downs and Vogel, 1993) as they gather the main challenging problems from the diagnosis point of view. Each filtering methodology was implemented after proper parameters tuning.

After testing the FDS under the different data pre-treatments, clear accuracy and reliability improvements were obtained in all the cases. In particular the use of SMA increased the FDS reliability in more than 9 % with respect to not treated plant data. Analysis of results also shows that the best overall FDS reliability achieved by any filtering technique does not imply the best diagnosis performance for every individual state. In that sense, although the EWMA treatment obtained lower general reliability than SMA it diagnosed more reliably faults 1st, 2nd or 12th. Therefore, it should be concluded that by incorporating the complementary skills given by the different filters, the global accuracy and reliability could be sensibly increased. That useful knowledge was summarized in form of a table containing for each data pre-treatment analyzed (R1 ?Rm) the reliability obtained diagnosing each fault (R1F1? RmFn).

Thus, an expert engine was developed to take advantage of the complementary effectiveness of the different data pre-treatments. This new diagnosis framework parallely uses two FDS distinguished by the pre-treatment applied in each case. The first FDS uses the best data pre-treatment, in terms of general reliability obtained after the previous filter analysis, whereas the second one uses that filter which best complement the best one, that is, this filter which improves the reliability obtained by the best filter in a bigger number of faults. The eventual diagnosis conflicts arising from the parallel FDS, are solved by the expert engine. It uses a set of rules that allows selecting the most reliable response among both available diagnosis by means of the knowledge table obtained from the filtering analysis stage. The procedure is depicted in the attached figure.

This new framework was tested obtaining promising results. It increased by more than 5 % the reliability obtained by the best isolated pre-treatment (SMA). Over 81 % overall reliability was achieved in this challenging diagnosis problem.

E. Musulin, I. Yélamos & L. Puigjaner. Integration of principal component analysis and fuzzy logic system for comprehensive fault diagnosis. Industrial and Chemistry Research, 45, 1739-1750, 2006

B. R. Bakshi. Multiscale PCA with application to multivariate statistical process monitoring. AIChE Journal, 44(7):1596?1610, July 1998.

D. Wang, J. A. Romagnoli. Robust multi-scale principal components analysis with applications to process monitoring. Journal of Process Control, 15 869-882, 2005.

Tona, R., Espuña, A., Puigjaner, L., ?Improving of Wavelets Filtering Approaches?. European Symposium on Computer Aided Process Engineering (ESCAPE-15) (L. Puigjaner and A. Espuña, Eds.), 1369-1374, ISBN: 0-444-51987-4, 2005.

Downs, J., Vogel, E. A plant-wide industrial process control problem. Comput. Chem. Eng. 1993, 17, 245-255.