An expert systems approach to model based signal processing of shock phenomena

Abstract

Expert systems have been proposed as a means of storing and applying a human expert's knowledge and problem solving skills. This would be a valuable facility in the area of signal processing as analysts already rely on computers for numerical manipulation of data. This research describes the successful conception and realising of software to assist a technically competent person in directing, applying and interpreting of signal processing algorithms with particular reference to the interpretation of measured responses to shock excitation. If a structure has been excited by some driving force, a measured response contains information about parts of the structure involved in the movement. An analyst given the measurement and asked to understand and explain the process that caused it will interpret the signal by building a model. The model is initially an empty shell but is made specific to the data by extracting information from the signal or prior knowledge. Information in the signal is stored in two ways; the numeric data values that represent the measured variable and the patterns visible to the eye when the data values are plotted against time. The information from the numeric values is extracted using algorithms which emphasise the previously hidden information. The design of the expert system has a model of the vibration process at its heart and aims to make it specific to the data just as the analyst does. There are three sections to the program, choice of a model, defining the components of the model and finally producing a report of the analysis. The first and third sections use rule based inference but the middle section is founded on a new architecture tuned for model building. It is a blackboard control structure organised to represent a linear system model. Knowledge sources are attached to each component and are scheduled by the user. A graphical interface is provided through which the user can view any part of the model in signal or symbolic format. Software is provided that creates the signals that form a view of the model and keeps the linear system causal. There are two main areas of knowledge application that give the program unique powers. One is in characterising features in the measurement emphasised by some algorithm as parameters of one of the components. The other is in finding a comparison whereby the accuracy of the parameter, in terms of how well it lets the model mimic the measurement, is established. Both these procedures are dominated by exploiting the user's ability to find patterns in noisy signals. This is particularly true when the model can generate different views of a signal. One statistical method of assessing a parameter value is explored and that is an adaptation of the maximum likelihood function used to find the confidence of epoch locations.

More information

Identifiers

Catalogue record

Export record