Rule I IF (INCREASE Argon flowrate) AND (INCREASE Power) THEN (INCREASE Ar-species).
Rule 38 IF (DECREASE Power and Nitrogen flowrate) AND
X. IF (Decrease Methane and Hydrogen flow rates AND Pressure) and (Increase
6.1 Relevance to OES Problem Domain
There were three specific areas covered in this research work. They were 1. Data classification via ANN recognition.
2. Nonlinear predictive modeling to relate multi-input variables to multi-output data m a continuously-valued problem domain providing a viable rule extraction technique.
3. Neuro-fuzzy approach to generate accurate rules.
The results obtained from these three areas answered two essential questions that formed the basis of the entire research project. From an overall view point of this project, here are the two essential questions and the responses attained.
• How well does the MLP network characterise OES patterns to identify species intelligently?
The MLP's performance on characterising species was very good. A fully connected leedforward BP MLP network was trained on a sufficient number of different OES spectral pattern types. The trained network had excellent generalising capabilities in distinguishing between spectral patterns to identify different chemical species.
Hie OES data set used was relatively small, yet once the MLP model had been successfully trained on approximately 80% of the patterns, it was able to generalise well without overfitting the data set, in order to classify seven different species. A performance level of 100% was achieved for the plasma system CH4/H2/Ar/N2 in identifying all seven species accurately within an independent test set. The network classifier determined the presence of Ar, H, H2, N2, N2+, CH and CH+ accurately which demonstrated its ability to distinguish between a mixture of species. Prerequisite normalisation of the spectral lines was implemented for network training
and testing. The simple rule-based system was used as a verification tool of ,he network's performance.
Can the MLP network identify a relationship between the controllable proces; parameters and the size of the spectral lines of species in order to generate rule; that aie useful for monitoring species quantities?
This second and more important question was of primary concern in terms of its direct applicability to the practical problem domain from which the OES data used here was obtained.
To answer it, three stages of work were implemented: Stage 1 . the creation of the ANN species models;
Stage 2 : the extraction of rules from the trained species models;
Stage 3 : a fuzzy rule extraction system to ensure accuracy or confidence in rules.
As a result of these three stages presented and discussed throughout the thesis, the following goals were achieved:
Predictive MLP species models were created.
These ANN models were successful in modeling the complex nonlinear relationship between the controllable plasma process variables (flow rates, power and pressure) and the spectral line size of seven different chemical species, without overfitting the data.
The excellent generalisation capabilities of the ANN models and their small network topology, provided the premise for extracting embedded knowledge from the trained networks in the form of comprehensible if-then rules.
The new rule extraction technique was based on a combination of sensitivity analysis and a backtrackingprocedure on the trained models.
38 specific rules were generated which from which the OES data was obtained.
were applicable to the ten plasma systems
The rules suggested which process variables could alter the si spectral lines thus providing a very useful process control tool.
size oi individual
The rules were tested empirically on independent data in the validation dataset.
Extracting rules by monitoring the sensitivity of the trained network models i pedagogical rule extraction technique.
The rule confirmation procedure using the backtracking method relies
summed weightedinput contributions to significantly identified hidden
on the
units.
The weight contributions of the significant hidden units to the summed weight of the outputs provided a direct relationship between input variables and output units.
The set of rules generated from backtracking matched those from the sensitivity analysis.
The backtrackingprocedure also introduced a set of new rules
I he rules obtained are categorised as globalrules.
The novelty of the rule extraction technique is its ability to efficiently cope with continuous-valued multi-output data.
I he rules extracted directly from the trained MLP networks tested very well on the validation data set. The method did not, however, generate all the possible rules for
,he SPeClral PIOblem d0mai"- No measure of ,„e r„,es was provided ertlrer. The „ u,o-fi,„ y approach has provided an adaptive.f e y rote extraction sysren, which has competent,,, addressed these two issues. The goals achieved are as follows
The adapted fttzzy rule extraction system used the MLP architecture.
Prerequisite discretisation of the input-output data fuzzified the adapti
process. ve learning
A large set of all the possible rules for a system based on the six controllable input process variables and one spectral line size (the predominant spectral line, for Ar750) were generated.
Each fuzzy rule generated had an accuracy value (CL).
There was a large set ol 635 highly accurate rules (out of the 2187) which determined the process conditions for obtaining a Small. Medium or Large Ar750 spectral line size.
Selective rule sets were found to subsume rules with fewer antecedent data.
The rule sets performed very well on the separate test data in the validation set.
In summary, this system can be deployed to predict the size of spectral lines for seven particular species for the purposes of generating useful process rules. The rule extraction procedure can be applied to problem domains consisting of multi-
dimensional continuously-valued input and output data. The representation language
of the rule extraction technique constitutes If-Then linguistic rules.
he fuzzy rule extraction system extended the precise extraction method (in chapter 4) by Pr°viding the set of all possible rules for the problem domain. More importantly,
the fuzzy system attached an accuracy value to the generated rule instantly. This provided the quantitative measure for determining the relevance of inputs to outputs The process was also transferable to ntulti-di,data.
These results have effectively answered the second query addressed i
which was: in this thesis
"Can the trained MLP network relate the controllable process parameters to the size of the spectes spectral lines in order to generate useful rules for monitoring species quantities?"
The answer is yes - the techniques developed have been very successful in implementing MLP networks that can model the complex nonlinear relationslup between the controllable process parameters and the size of the spectral lines of species. The extraction method generates useful rules to suggest to the process controller for monitoring species quantities via the size of the spectral lines.