Brought to you by EarthWeb
IT Library Logo

Click Here!
Click Here!

Search the site:
 
EXPERT SEARCH -----
Programming Languages
Databases
Security
Web Services
Network Services
Middleware
Components
Operating Systems
User Interfaces
Groupware & Collaboration
Content Management
Productivity Applications
Hardware
Fun & Games

EarthWeb Direct EarthWeb Direct Fatbrain Auctions Support Source Answers

EarthWeb sites
Crossnodes
Datamation
Developer.com
DICE
EarthWeb.com
EarthWeb Direct
ERP Hub
Gamelan
GoCertify.com
HTMLGoodies
Intranet Journal
IT Knowledge
IT Library
JavaGoodies
JARS
JavaScripts.com
open source IT
RoadCoders
Y2K Info

Previous Table of Contents Next


4.3. HYBRID SYSTEMS

One example of a hybrid system is that developed by MacIntyre et al. for National Power's Blyth Power Station, U.K. (MacIntyre et al., 1995). As part of an ongoing project at Blyth, several software systems have been developed for the analysis of vibration spectra from the Primary Air Fan bearings. However, the most successful application has been a hybrid system which uses an expert system rule base, along with numerical techniques, to pre-process vibration data in order to produce a reduced input vector for presentation to a neural network.


No. Agent Function
1 Telemetry agent Receives telemetry and converts it into a standard format
 
2 Information agent (IA) Contains information about the physical network
 
3 Weather watcher agent Knows about the time and location of lightning strikes
 
4 High voltage diagnosis agent Uses telemetry to diagnose the location, time and type of fault
 
5 Low voltage diagnosis Takes into account customer information, lightning information and telemetry to diagnose faults
 
6 Switch planning agent Allows the user to plan maintenance work
 
7 Advisor agent Acts as the interface between the system and the user

FIGURE 7 The seven agents in CIDIM.

The neural networks used within the hybrid system are based on the Multi-Layer Perceptron (MLP) architecture, using the back propagation training algorithm. The use of back propagation allows the adjustment of weights in the neural connections in multiple layers; this is critical if a network is to solve non-linearly separable problems. The MLP architecture using back propagation relies on a technique referred to as "supervised training," in which an input vector (that is, a numerical representation of the input pattern, in vector form) is presented to the neural network along with a target output vector (a numerical representation of the desired output for the given input, again in vector form). The importance of producing the best input vector is paramount; it is for this reason that expert systems are used along with numerical techniques to ensure that this occurs. Similarly, expert systems are used to interpret the output of the system.

The actual output is compared to the target output for each input vector, and the root mean squared (RMS) error is calculated. This error is then propagated backwards through the neural connections, and the process is repeated until the RMS error is within an acceptable threshold.

The heart of the system is the Neural Bearing Analyzer (NBA) model. This model was developed initially taking only certain areas of the vibration spectrum as the input vector, selected by the use of an expert system. The network had a relatively large output set, with separate classes for each of the bearing components, and different levels of defect severity. This version of the network demonstrated the difficulties of back propagation in achieving convergence to within an acceptable RMS error level. To achieve convergence it was necessary to alter the learning parameters in order to avoid the algorithm sticking in local minima. Convergence was eventually achieved, and in testing the network produced 93% agreement with a consultant diagnostic engineer's classifications.

A different approach involved presentation of the full vibration spectrum (400 datu points) to the network, and a simplified output set, with condition estimates for the bearing as a whole. In both cases, training data sets were constructed from real data collected from Station machinery, where the target output could be confidently generated as a result of known bearing condition due to inspection or replacement of bearings. Testing of the models was performed both with real data acquired from the Primary Air Fans, some of which had known conditions from inspection, and also with artificially-generated data to examine the network's performance in identifying particular defect types. The network's classifications were compared with known condition data, and also with the classifications of the consultant engineer. This model proved much more difficult to train effectively. Convergence was only achieved after raising the RMS threshold to 0.01, and much "tweaking" of the learning parameters. Running on a 33Mhz 486 PC, the network took almost 1.5 hours to train, using over 100,000 training iterations. In training, this version of the network performed less well than the first, with only 81% agreement in classifications.


FIGURE 8 The hybrid system.

These models have been further developed and refined, in the areas of selection of the input parameters, and the formulation of the output classes. The use of all 400 points in a vibration spectrum presents difficulties due to file sizes and can be slow; furthermore, it is apparent that a substantial part of this information is redundant in terms of performing the classification task. In order to reduce the dimensionality of the problem, the numerical technique of Principal Components Analysis (PCA) has been used to determine which, if any, variables should be added to those selected by the expert system approach as the necessary inputs to the network. A diagram of the system is shown in Figure 8.

5. SUMMARY

This chapter has presented an overview of the applications of expert systems and related technology in the power industry. There is clearly a lot of activity in this area. The authors believe that this work will continue, and that, in particular, the field of hybrid systems will receive increased attention in the future.

The major research issues to be addressed in the future will focus around how hybrid technology can best be utilized to more closely model power systems. The systems in use in power generation are extremely complex, so any improvement in modeling capability will represent a great advance. Future hybrid systems will integrate AI, neural, and simulation technology. It is also vital to find a way to fully validate the complex software systems that will result.


Previous Table of Contents Next

footer nav
Use of this site is subject certain Terms & Conditions.
Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details.