LAB 3

1
LAB 3

• AIRBAG DEPLOYMENT
• SENSOR PREDICTION NETWORK

Warning This lab could save someones
life!
2
Two Approaches to Investigate

• Lab 3

Part II Time Series Forecasting Using Tapped
Delay Line Neural Network (TDNN)
Part I Time Series Forecasting Using Madaline
Submit formal report on Lab 3 by 11/2/04 (no
class 10/26)
3
Part I Time Series Forecasting Using Madaline

• Aim is to predict airbag sensor output at time t
  1 based on prior outputs at t, t-1, t-2, t-3
  . using Madaline
• Initially 3 delay elements will be applied to
  inputs
• Select and organize appropriate data file AIRBAGx
  for Training and Test of Madaline where xlast
  digit of your ss
• Select Network Architecture with 1 hidden layer
  of nodes and one output
• Select suitable of hidden layer nodes for
  problem giving reasons for your choice

4

• Construct Madaline Network
• Train Test. (Choose suitable epoch)
• Plot and Comment on Prediction Results
• How accurate is Madaline prediction for
  additional future time interval t2 ?
• Compile a brief report on the results of Madaline
  forecasting

5
Example of Training File

• !Square Rotate Training File (80 samples)
  3/27/04
• !This data collected at 800 samples per second.
• !Shaker table moving at 20 Hz with a square
  wave.
• !Noise added to the sample by having the MEMs
  loosely mounted.
• !MEMS rotated during collection of samples.
• !
• !
• ! t -2 t-1 t t1
• 2.0769 2.0769 2.0769 2.0769
• 2.0769 2.0769 2.0769 2.0134
• 2.0769 2.0769 2.0134 1.6813
• 2.0769 2.0134 1.6813 2.0403
• 2.0134 1.6813 2.0403 1.928
• 1.6813 2.0403 1.928 2.1184
• 2.0403 1.928 2.1184 2.1819
• 1.928 2.1184 2.1819 2.2943
• 2.1184 2.1819 2.2943 2.4139
• 2.1819 2.2943 2.4139 2.4481
• 2.2943 2.4139 2.4481 2.475

6
Results for 3-3-1 Madaline

• Construct, Train and Test Madaline
• Report your test results using Excel chart
• Compute the RMS error and comment on accuracy of
  Madaline prediction

7
Test Results for 3-3-1 Madaline
Actual Predicted
2.382200 1.510400 2.487200 2.518900
2.470100 2.518900 2.487200 2.518900 2.494500 2
.518900 2.489600 2.518900 2.479900 2.518900 2.4
67600 2.518900 2.492100 2.518900 2.501800 2.5189
00 2.484700 2.518900 2.494500 2.518900 2.475000
2.518900 2.470100 2.518900 2.462800 2.518900 2
.352900 2.518900 2.186800 1.510400 2.089100 1.51
0400 1.510400 1.510400 1.935300 2.518900
8
Part II Time Series ForecastingTapped
Delay-Line Neural Network Backprop Learning

• Aim is to predict airbag sensor output at time t
  1 based on prior outputs at t, t-1, t-2, t-3
  . using a TDNN
• Initially 3 delay elements will be applied to
  inputs
• Select and Organize Airbag appropriate data file
  AIRBAGx for Training and Test of TDNN where
  xlast digit of your ss
• Select Network Architecture as Multilayer
  Perceptron Feedfoward with 1 hidden layer of
  nodes and one output
• Select of hidden layer nodes giving reasons for
  your choice

9

• Construct Network
• Train Test. Choose suitable learning
  coefficient, momentum term and epoch
• Record train test parameters as well as RMS
  Error and Classification Rate after experiment
• Plot and Comment on TDNN Prediction Results
• How accurate is your TDNN prediction for
  additional future time interval t2 ?
• Repeat experiment for memory depth of 6 compare
  with results for depth of 3
• Note that in this scheme, delayed inputs have
  same weight as current input creating a Linear
  Trace Memory

10
Non-Linear Trace Memories

• Should more recent inputs have greater influence
  than older inputs?
• Empirically verify your answer by applying kernel
  function to delayed inputs to produce a
  non-linear trace memory
• Refer to Mohan page 140 Use convolution of
  input sequence with kernel function ci
• One example is kernel function in which previous
  input has half the weight of input immediately
  succeeding it. Write MATLAB fn to compute
  transformed inputs
• Use a memory depth of 6

11
My TDNN After Training
12
Test Results File
DESIRED PREDICTED
Linear Trace Memory Memory Depth 3

• 2.382200 2.315152
• 2.487200 2.398239
• 2.470100 2.475807
• 2.487200 2.463754
• 2.494500 2.438213
• 2.489600 2.453228
• 2.479900 2.447508
• 2.467600 2.437722
• 2.492100 2.428249
• 2.501800 2.442653
• 2.484700 2.459934
• 2.494500 2.445343
• 2.475000 2.443128
• 2.470100 2.439142
• 2.462800 2.426756
• 2.352900 2.426402
• 2.186800 2.351290
• 2.089100 2.184385
• 1.510400 2.076567

Train for 20,000 presentations
80 Training Samples, 20 test
RMS Test ERROR 0.0602
13
TDNN Test Results Linear Trace Memory, Depth 3
14
Airbag Deployment with Advance Warning

• Devise a scheme for airbag deployment with
  advance warning using TDNN Predictor
• Assume under crash conditions, airbag sensor
  output voltage is in range 0.25 0.25 or 4.75
  0.25
• Modify your database to introduce a minimum of 2
  crash conditions at least 10 time delays apart
• Does your TDNN predict impending crash
  conditions?
• How much notice does TDNN offer driver (in time
  delays) ?
• If one time delay 1ms, can the average adult
  driver react to the information in time to
  prevent disaster?
• Comment on usefulness of Advance Warning scheme

15
Airbag Sensor Reliability Enhancement

• Automobile airbag sensors sometimes deploy
  accidentally leading to a many unnecessary car
  crashes
• Describe a scheme to enhance airbag sensor
  reliability through redundancy (multiple sensors)