Overview of Minicon Project

1
Overview of Minicon Project

• Condition monitoring and diagnostics for
  elevators
• Dale Addison
• CENTRE FOR ADAPTIVE SYSTEMS
• University of Sunderland
• School of Computing Technology

2
Project overview

• MINICON
• Minimum Cost Maximum Benefit Condition Monitoring
• Framework 5, Competitive and Sustainable Growth,
  project value 3 million)
• Two aspects
• Condition monitoring of elevators
• Condition monitoring of high speed machine tools
  (gt15000 rpm)

3
Project partners

• Kone (4th largest supplier of elevators in the
  world) (Finland)
• VTT (Finland)
• Goratu (Bilbao, Spain
• Tekniker (Bilbao, Spain)
• Rockwell Manufacturing (Belgium Czechoslovakia)
• Technical University of Talinn (Estonia)
• IB Krates (Estonia)
• Monitran ltd (UK)
• University of Sunderland, (UK)
• Entek (UK)
• Truth (Athens, Greece)
• ICCS/NTUA (Athens, Greece)

4
Application software and database with second
level Intelligence. Prior Knowledge Intelligence
System
Maintenance Management System With third level of
intelligence, human, providing Decision Support
Plant or Machine to be monitored e.g. Elevator
or machine tool
Service Engineer with Hand held PC, Email or
paging device etc.
5
(No Transcript)
6
Neural networks

• Adaptive technology device based upon neurons
  found in the human brain
• Neurons are connected together and send signals
  to each other. (networks)
• Signals are summed and when they exceed a certain
  limit, the neuron fires (sends signal to other
  neurons)
• Networks can be trained using algorithms which
  respond to the data.

7
A Neural network

• Multi-layer perceptron
• Each neuron performs a biased weighted sum of
  their inputs.
• This activation level is passed through a
  transfer function (usually sigmoidal) to produce
  its output,
• Neurons are arranged in a layered feed forward
  topology.

8
Multi layer perceptrons

• Networks weights and thresholds are adjusted by a
  training algorithm which alters the weights
  according to the training data.
• This ensures the smallest possible difference
  between the input data and the outputs

9
Dimensionality reduction techniques

• Principal Components Analysis
• Non-Linear
• Weight regularisation techniques (Weigand method)
• Genetic algorithms

10
Non-Linear principal components analysis (Auto
Associative network)

• Neural network which uses its inputs as outputs
• Has at least one hidden layer, with less neurons
  than the input and output layers, which have the
  same number of neurons
• Data is effectively squeezed through a lower
  dimensionality

11
Non-linear Principal components analysis
12
Auto associative training

• Produce Auto associative training set (Inputs map
  to outputs)
• Create auto associative MLP
• 5 layers
• Middle hidden layer has less units than output
  layers
• Other two hidden layers have a relatively large
  number of neurons, both should have the same
  number
• Train network on data set
• Delete last two layers
• Collect reduced dimensionality input data,
  replace the original input data, retain original
  output variables
• Create a second neural networkand train on the
  reduced data set.

13
Use of Genetic Algorithms

• GAs are an optimisation technique which use
  Darwins concept of survival of the fittest to
  breed successively better strings according to an
  objective function
• In this problem, that function helps to determine
  subsets of inter-related bits (correlated or
  mutually required inputs)

14
Sensitivity analysis

• The sensitivity of a particular variable is
  determined, by running the network on a set of
  test cases, and accumulate the network error.
• The network is then run using the same cases, but
  without certain information used
  earlier,(specific input variable) and the network
  error accumulated.
• The sensitivity error is the ratio of error with
  missing value substitution to the original error.

15
Neural Network weight regularisation

• Promotes low curvature by encouraging small
  weights to model the feature surface
• Adds extra term to the error function which
  penalises gratuitous larger weights
• Also prefers to tolerate a mixture of large and
  small weights, rather than medium sized weights

16
Neural networks used

• Multi Layer Perceptrons (MLP)
• Radial basis function (RBF)
• Self Organisng Feature Maps (Kohonen)
• Experiments were ran on several different data
  sets, recorded over a number of time periods
  (day, 2 days one week)

17
Radial basis function nets

• Feature space is divided up using circles
  (hyperspheres)
• Characterised by centre and radius
• Response surface is a gaussian (bell shaped curve)

18
Radial basis function networks

• RBF consists of
• Hidden Layer of Radial Units modelling a Gaussian
  response surface (only one)
• Training of RBFs
• Centres and Deviations of Radial Units are set
• Linear Output layer is optimised
• Centres assigned to reflect clustering of data

19
Radial basis function networks

• Centre assignment methods
• Sub sampling Random number of training points
  are copied to the radial units
• K-means algorithm Set of points are selected and
  placed at the centre of clusters of training data.

20
Radial basis function networks

• Deviation assignment (Determines how spiky the
  gaussian functions are)
• Explicit (choose yourself)
• Isotropic heuristic method using the number of
  centres and volume of space occupied
• K-NN Units deviation is individually set to the
  mean distance of its K-NN
• Small in tightly packed areas (preserves details)
• Higher in sparse areas of space

21
Artificial Neuron, and the Kohonen SOFM
22
Application to MINICON project

• Picture shows test elevator used at KONE site,
  sensors mounted at various sites and results used
  as input to neural networks
• Self Organising feature maps and Multi-Layer
  perceptrons trained on a variety of elevator
  data.

23
Results
24
Results for Genetic Algorithm
25
Results for Sensitivity Analysis
26
Results of weight regularisation techniques
27
Results for auto association
28
Best performing technique per data set
29
Final Networks

• Two types of neural networks used in final
  product
• Multi-Layer Perceptrons(with weight
  regularisation applied)
• SOFM
• Both networks require different numbers of inputs
  depending on the data set (5-15)

30
Alternative methods

• Use of Statistical techniques
• Mean, Kurtosis, Standard Deviation
• For example the mean of one parameter suggests
  a significant rise for data set number 5. Since
  all the other data sets show a consistent mean
  then this example seems to be highly significant.

31
Use of mean value
32
Conclusions

• Removing input data does not improve
  classification performance
• Statistical techniques not consistent to make
  reliable estimates
• MLPs and SOFMs are best performing NN
  techniques
• MLPs performance can be improved by applying
  weight regularisation techniques.