60







 06
 06










 20

 10
 36

 0020




 500 1000 1500 2000 2500 3000


Figure5-7 Data set: (a) Flow; (b) Pressure.

 We then test the same algorithms and settings as before on this data set, keeping the

network dimensions and embedding the same, but fine tuning the learning, competition

and performance parameters to achieve segmentation of the data: in the end, the

following results were achieve with the same parameters as for the first data set:


Yp, =0.02 and M=5.


5.3.2 RBF Network Models

 Figures 5-8 /5-9 show the result for the RBFs' k-means clustering; Figures 5-10/5-

11 show the results for the OLS clustering. In both cases, all regime changes are detected,

but the latter is less accurate, detects the changes later and is more susceptible to spurious

switches than the former. Also on the gate graph, one can see that the competition is not

as hard with OLS clustering, which means the models are less specialized.