Doctoral Dissertations

Date of Award

8-1994

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Engineering Science

Major Professor

Robert E. Uhrig

Committee Members

Mohan M. Trivedi, J. A. M. Boulet, Belle R. Upadhyaya

Abstract

New methodologies for the automation of vibration monitoring and for information fusion are introduced in this dissertation. The foundation of all the techniques is neural network technology, and most of them also utilize fuzzy logic and genetic algorithms. The combination of these paradigms contributes to high speed and flexibility of the systems developed and allows us to take advantage of the strong points of each paradigm.

For classification of vibration data in case of information gathered by one sensor exclusively, a hybrid neuro-fuzzy system was developed that extensively utilizes the traditional vibration analysis methods. The presence of peaks at characteristic frequencies of faults, their harmonics and their immediate neighborhood is assessed. Then the amplitudes, transformed by membership functions, constitute the input to a Kohonen self-organizing map with categorization. The output of the network is the classification of the fault (or faults) present.

When developing automatic methods for information fusion, three major issues arise. The first is the knowledge about the reliability of the information sources; second - the degree of redundancy/complementarity of the information sources; third - if a hierarchical scheme is needed, the structure of the hierarchy. Since very often this information is not available, there is a need for a general scheme for finding the parameters of the aggregation function. When all three of the above factors are missing, the problem is particularly difficult, and no satisfactory solution has yet been proposed in the literature.

The methodologies devised in this work address all the three issues mentioned above. They can be subdivided into techniques consisting of one fusion step and techniques consisting of multiple fusion steps. The one-level aggregation methods are intended for data that show a similar degree of redundancy/complementarity. The one level fusion techniques developed include a probabilistic neural network-based technique and a neuro-fuzzy-genetic technique. The foundation of the first technique is Bayesian decision theory as implemented by probabilistic neural networks. The second technique uses fuzzy set theoretic aggregation connectives, the parameters of which are established by a genetic algorithm. These connectives are capable of combining information not only by union and intersection used in traditional theories but also by compensatory connectives that mimic the human reasoning process very closely.

Some of the multi-step techniques developed herein are intended for cases when auxiliary knowledge about the fusion process exists, or when some simplifying assumptions can be made. The simplifying assumptions can be, for example, that the reliability of each sensor is the same. Another assumption that can be made when the number of sensors is large, is that at the beginning of the fusion process, i.e. when fusing information from a small number of sensors (2 or 3), the information is complementary, and later as the number of sensors increases, the information becomes more and more redundant. Therefore, using a union-like operation at the beginning of the fusion process and an intersection-like operation as the number of sensors increases, constitutes a simple, yet powerful, scheme.

The distinctive feature of the next method devised is that the optimal parameters of the aggregation connective are automatically found by a genetic algorithm. Therefore assumptions about reliability and redundancy/complementarity of the sensors are not needed. At each step the method performs fusion from two sensors and only the knowledge about the order of sensors for the fusion process is needed.

The most difficult fusion case occurs when no information whatsoever about the reliability, redundancy/complementarity of the sensors or the hierarchy of the process exists. The hierarchical aggregation scheme devised herein is intended for this particularly difficult, but frequently encountered case. The hierarchical scheme is capable of building the sensor's hierarchy without any prior knowledge about its structure (data from which sensors to fuse together), nor does it know the reliability or the degree of redundancy/complementarity of the sensors.

The novel methodologies proposed are applied to the problem of classification of vibration data from laminar flow table rolls in a steel sheet manufacturing mill, rolling element bearings data and some simulated data sets. The techniques developed for aggregation of evidence have a very broad range of applications; they can be used for any problem involving fusion of decisions. The technique is not limited to vibration monitoring; it also applies to recognition of objects in a computer vision system, area surveillance systems for military purposes and managerial decision making systems. Once the decisions to be fused are obtained, the systems are capable of aggregating them in a near optimal manner.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS