TSAPTSINOS, Dimitris (1992). Crystal filter tuning using machine learning. Doctoral, Sheffield City Polytechnic.
Documents
3145:2653
Archive (ZIP) (Ethos)
Tsaptsinos306227.zip - Accepted Version
Available under License All rights reserved.
Tsaptsinos306227.zip - Accepted Version
Available under License All rights reserved.
Download (21MB)
3145:487407
PDF (Version of Record)
10701103.pdf - Accepted Version
Available under License All rights reserved.
10701103.pdf - Accepted Version
Available under License All rights reserved.
Download (11MB) | Preview
Abstract
Manual tuning of electronic filters represents a time-consuming process which can benefit from some computer assistance. A prototype computer-based system for the tuning of crystal filters after manufacture was developed. This system solved the
problem of crystal filter tuning in a novel way.
The system, called AEK (Applied Expert Knowledge), was developed using crystal filters and is a hybrid system with the following two functions:
(1) Required values of features are extracted from the filter waveform and passed to the expert system which determines the component to adjust and the direction to turn, or the end of the tuning.
(2) Sampled values of the waveform are extracted and passed to a neural network which determines how far to turn the component chosen in (1).
The prominent aspects were:
- Work using the protocol analysis elicitation technique indicated the need to separate the process into two sub-tasks (stopband and passband). Each sub-task was divided into three classification parts which determined (i) the continuation of the tuning process, (ii) the component and direction to turn, and (iii) the distance to turn respectively. Unfortunately, it was not possible to extract rules from the operator.
- Three learning techniques (IID3, Adaptive Combiners, Neural Networks) were used and compared as the means of automated knowledge elicitation. All three techniques used case knowledge in the form of examples. The investigations suggested the use of ID3 for the first two parts of each subtask employing features with linguistic values. The number of linguistic values each feature has, was also derived.
- Neural networks were trained for the third part. It was necessary to have one network for each component/direction combination and to use examples from just one mal-a(busting process.
- Tests of the hybrid system for a number of cases indicated that it performed as well as a skilled operator, and that it can be used by novice operators but situations arose where there was either no knowledge or contradictory knowledge.
The prototype system was developed using one type of crystal filters but the generic construction procedure can be followed to build other systems for other types.
More Information
Statistics
Downloads
Downloads per month over past year
Downloads per month over past year for
"Tsaptsinos306227.zip"
Downloads per month over past year for
"10701103.pdf"
Share
Actions (login required)
![]() |
View Item |