Page 1084 - The Mechatronics Handbook

FIGURE 39.1 Architecture of a typical multilayer feedforward neural network.

The knowledge base contains highly specialized knowledge on the problem area as provided by the expert
in the form of statistical analysis, empirical or semi-empirical rules, theoretic and computer simulation
studies, and experimental testing. It includes problem facts, rules, concepts, and relationships.
Expert systems have obvious knowledge representation forms that make knowledge easy to manage,
have the capability to explain their behavior, and can diagnose new faults using their knowledge bases.
At the same time, self-learning is still a problem and computation time can be quite lengthy for difficult
tasks.
A neural network is a highly nonlinear system with adaptation and generalization capabilities. There
are many different architectures of neural networks; however, the multilayer feedforward neural network
(refer to Fig. 39.1) is one of the most popular ones. This is because of the simplicity, availability of efficient
learning methods, generalization capabilities, and noise tolerance of these networks. This network is a
collection of simple, interconnected nodes, also known as neurons, which operate in parallel and store
knowledge on the strength of connections between the individual nodes. Such a parallel computing
network, inspired by the computational architecture of the human brain, has been successfully applied
to intelligent tasks such as learning, speech synthesis, and pattern recognition. The input vector feeds
into each of the first layer neurons, the outputs of this layer feed into each of the second layer neurons,
and so on. The last layer, which generates output to the external world, is called the output layer. The
hidden layers are not connected to the external world. Often the neurons are fully connected between
the layers, i.e., every neuron in layer l is connected to every neuron in layer l + 1.
Training a neural network consists of the process of finding the set of interconnection weights (there
is an interconnection weight associated with each neuron which modifies the input signal to that neuron
in a specific manner), which results in a network output that satisfies a predefined criterion. Feedforward
neural networks are trained using the backpropagation algorithm. This is a supervized training method.
This means that the network will be presented with sample inputs and correct responses, called a training
pattern. The network is then trained to reproduce the correct responses.
Neural networks have capabilities of association, memorization, error tolerance, self-adaptation, and
multiple complex pattern processing. However, they cannot explain their own reasoning behavior and
cannot diagnose new faults (those not already made available previously in training the network).

39.5 Problems in Intelligent Fault Detection

The fault detection scheme should be capable of detecting and identifying the failures correctly and
promptly with minimum delays. This requires a reconfigurable robust controller. That is, the controller
should distinguish between failures, uncertainties/inaccuracies in the model of the system, and distur-
bances such as sensor noise; and reduce the effect of measurement error and noise, uncertainties in the
system model, and disturbances (even component failure) on the system output.

©2002 CRC Press LLC