B.H. Chen

Application of wavelets and neural networks to diagnostic system development, 1, feature extraction

An integrated framework for process monitoring and diagnosis is presented which combines wavelets for feature extraction from dynamic transient signals and an unsupervised neural network for identification of operational states. Multiscale wavelet analysis is used to determine the singularities of transient signals which represent the features characterising the transients. This simultaneously reduces the dimensionality of the data and removes noise components. A modified version of the adaptive resonance theory is developed, which is designated ARTnet and uses wavelet feature extraction as the substitute of the data pre-processing unit. ARTnet is proved to be more effective in dealing with noise contained in the transient signals while retains being an unsupervised and recursive clustering approach. The work is reported in two parts. The first part is focused on feature extraction using wavelets. The second part describes ARTnet and its application to a case study of a refinery fluid catalytic cracking process.

Neural network based fault diagnosis using unmeasurable inputs

Abstract Much of the earlier work presented in the area of on-line fault diagnosis focuses on knowledge based and qualitatively reasoning principles and attempts to present possible root causes and consequences in terms of various measured data. However, there are many unmeasurable operating variables in chemical processes that define the state of the system. Such variables essentially characterise the efficiency and really need to be known in order to diagnose possible malfunction and provide a basis for deciding on appropriate action to be taken by operators. This paper is concerned with developing a soft sensor to assist in on-line fault diagnosis by providing information on the critical variable that is not directly accessible. The features of dynamic trends of the process are extracted using a wavelet transform and a qualitative interpretation, and then are used as inputs in the neural network based fault diagnosis model. The procedure is illustrated by reference to a refinery fluid catalytic cracking reactor.

Application of wavelets and neural networks to diagnostic system development, 2, an integrated framework and its application

Abstract A method for feature extraction from process dynamic transient signals using wavelet multiscale analysis was introduced in part 1 of this paper. In part 2 we describe an integrated framework combining wavelet feature extraction and an unsupervised neural network for identification of operational states. Application of the system to a refinery residual fluid catalytic cracking process is also presented.

An efficient and stable Ru–Ni/C nano-bimetallic catalyst with a comparatively low Ru loading for benzene hydrogenation under mild reaction conditions

A prepared 0.024%Ru–1.00%Ni/C nano-bimetallic catalyst reported in this work is a highly efficient and stable catalyst for the hydrogenation of benzene to cyclohexane, with an unprecedented TOF up to 7905 h−1 under mild reaction conditions (20 °C, 40 psi H2) without adding any solvent. The method for the preparation of catalyst is very simple and low-cost. Therefore, this cyclohexane synthetic approach is potentially more environmentally friendly and economical than existing technology.

Reaction modelling and simulation to assess the integrated use of transketolase and ω-transaminase for the synthesis of an aminotriol

The most attractive, as well as challenging, multistep organic syntheses would preferably be carried out in a single reactor, as a one-pot synthesis. For biocatalytic syntheses, multistep reactions in one-pot mode bring a number of advantages, while at the same time raising unique challenges such as the compatibility of different biocatalysts. In this paper, we have developed a transketolase–transaminase (TK-TAm) two-step one-pot aminotriol synthesis reaction model, which integrates reaction kinetic models with process characterization (consisting of component degradation as a function of pH and concentration, aldehyde toxicity towards the enzyme, and ketol donor and acceptor side-reactions with TAm). Based on the analysis of the effect of the TAm/TK activity ratio on product yield, simulations provided guidance for further process and biocatalyst development.

Fuzzy rule generation from data for process operational decision support

Neural networks and expert systems are now beginning to realise their potential in developing intelligent operational decision support systems. While neural networks are able to learn complex nonlinear functional relations between multiple inputs and outputs, there remains the important limitation that knowledge embedded in the neural network is opaque. This image of a black-box technology is a major factor influencing the acceptability of this approach because it does not improve the heuristic understanding of the domain problem. On the other hand, expert systems make use of logic rules to carry out heuristic reasoning. Knowledge used to reach a conclusion is transparent and can be displayed through HOW and WHY explanation facilities which are an integral part of well-founded expert systems. However, a critical problem with this approach is knowledge acquisition. This contribution introduces a fuzzy neural network to extract fuzzy rules automatically from numerical data. By changing fuzzy membership functions, three types of rules can be extracted, i.e. rules with and without fuzzy membership values and neuro-expert systems which overcome the problems described above.

Data Mining for Failure Diagnosis of Process Units by Learning Probabilistic Networks

Recently, there has been a growing interest in developing and applying knowledgebased technologies to aid hazard identification methods such as Hazop (Hazard and Operability Studies), fault tree analysis and check-lists which have traditionally been carried out manually.A critical factor is the knowledge which is used. Previous experience and cases of failure provide an important source of information which can be used to update knowledge. However, the volume of data is normally too great to carry out manual analysis. Moreover, the data is complex in structure and of diverse types, as well as being noisy and having missing elements. This results in the databases being mainly used for archive and retrieval. This paper reports the application of probabilistic networks and how they can be used for learning about failure diagnosis of process units by extracting knowledge from the databases in the form of rules, which can be used either by experts or in building expert systems.

Neural nets, fuzzy sets and digraphs in safety and operability studies of refinery reaction processes

Neural nets, fuzzy sets and graph theories have been used individually to the detection and diagnosis of faults during process operation, as well as assessment of potential hazards and operability problems in design, but all three approaches have their limitations. This study builds on the individual strengths and seeks to blend the three approaches to compensate for inadequacies by adopting the algorithm from each of the method into a composite procedure for capturing process knowledge. A fuzzy qualitative approach is proposed for interpreting dynamic data including patterns of change in process variables, meeting process constraints and closed-loop dynamic behaviour so that a neural network for fault diagnosis can deal with dynamic data effectively. The neural network learning algorithm is adapted to train a fuzzy relation matrix between two fuzzy sets which provides a covering model for fault diagnosis. Back Propagation Neural Network (BPNN), single layer perceptron (PCT) and Fuzzy Set Covering (FSC) methods are examined with respect to the fault diagnosis of a fluid catalytic cracking process. While the three methods have similar capabilities in identifying significant disturbances and faults, BPNN is better than PCT and FSC in isolating small disturbances. These concepts enable a new fuzzy-SDG approach to be developed which uses fuzzy membership functions to replace the sigmoid function in the nodes with the effect weights of the graph being obtained by training. With this technique it is possible to extract qualitative knowledge from plant records to model disturbance propagation in the process.

Wavelet shrinkage data processing for neural networks in bioprocess modeling

Abstract The modeling of biological systems has now become an essential prerequisite for effective bioprocess design, optimization and analysis. The difficulties present in using conventional techniques to model such a complex system make the application of artificial neural networks (ANNs) to these problems particularly attractive because of their capability for nonlinear mapping and lack of necessity for detailed mechanistic knowledge. However, building a reliable ANN model requires sufficient training data, which may be difficult when data are collected from litre-scale experiments. In this work, a bioconversion (with only limited experimental data) was firstly modeled by a radial basis function (RBF) neural network. Although the model provided a very low variance between experiment and simulation, it tended to result in oscillatory behaviour, which clearly does not reflect the accurate profile of the reaction. In order to overcome this drawback, wavelet shrinkage with biorthogonal filters was used to generate a reconstructed function using the RBF model as a base. The synthesis of N -acetyl- d -neuraminic acid by the enzymatic condensation of pyruvate with N -acetyl- d -mannosamine was used as a case study to show the effectiveness of the approach. The effects of alternative filters and border distortion are also discussed.

Clustering of infrared spectra of lubricating base oils using adaptive resonance theory

Adaptive resonance theory (ART2) is applied to the classification of lubricating base oils on the basis of their infrared spectra. Fifty-nine data samples are used which were collected from 12 refineries representing eight crude oil origins. The ART2 neural network is an unsupervised machine learning approach which is different from supervised learning, such as back-propagation neural networks, in that it determines both the number of classes and the assignments of data samples. The ART2 groups the 59 data samples into seven classes. Five of the seven classes are found to perfectly match the crude oil origins. Two of the seven classes (eight samples in total) are combined to form a single class. The class assignment of one data sample (sample 35) does not match the crude oil origin but is consistent with the prior observation of its spectrum. The work demonstrates that ART2 represents a useful alternative tool for infrared spectra interpretation.