Upload
others
View
17
Download
0
Embed Size (px)
Citation preview
P. Tawdross and A. Königslide:1
Particle Swarm Optimization for Synthesis and Dynamic Reconfiguration of Sensor Electronics
Peter Tawdross and Andreas König
Contents:Introduction Dynamic Reconfiguration ApproachTarget HardwareExperiments and ResultsConclusions
Institute of Integrated Sensor Systems
Dept. of Electrical Engineering and Information Technology
P. Tawdross and A. Königslide:2
Sensor 1
Sensor 2
Sensor 4
Sensor 3
Sensor 5 SignalCondit.
ADC1
ADCn
Actor 1
Actor 2
DSP/Embedded
system
DAC
Introduction
P. Tawdross and A. Königslide:3
Genetic algorithm or genetic programming design topologies and dimension them
To many operators, too many methods for each operator
Building arbitrary topologies with unpredictable behaviour
Starting from scratch after each environmental change as the current arbitrary topologies is not guaranteed to operate well at the new environment
The evaluation in the intrinsic level is the relation between the input and the output (no explicit spec. optimization)
Hard to find industrial acceptance
Dynamic Reconfiguration Approach State of the Art
P. Tawdross and A. Königslide:4
≤∀
>∀−
=
ii
iii
ii
i
fspec
fspecspec
fspecE
0
≥∀
<∀−
=
ii
iii
ii
i
fspec
fspecspec
specfE
0
Min:Max:
The environment with multi-objective in the device level
Multi-objective
TsSR
Avo
offset
CMRR
CMRPSRR
PSO
ngspice/ Real chip
∑ ×= ii EHF
Ro
Pc
BWol BW3db
Swing
ϕ
PD
.....
Dynamic Reconfiguration ApproachThe Design Environment
P. Tawdross and A. Königslide:5
The extrinsic environment
Dynamic Reconfiguration Approach The Extrinsic Design Environment
Optimization Library
Simulator SimulationLibrary
RequiredSpec.
P. Tawdross and A. Königslide:6
RequiredSpec.
The extrinsic environment
Dynamic Reconfiguration Approach The Extrinsic Design Environment
Optimization Library
Simulator SimulationLibrary
optimization library
PSO
HPSO
CPSOMSPSO
Other optimizationmethods
optimizationselector
Fitness functionSelector
P. Tawdross and A. Königslide:7
The extrinsic environment
Dynamic Reconfiguration Approach The Extrinsic Design Environment
Optimization Library
Simulator SimulationLibrarysimulation
library
netlist generator
postprocessingobjectives
costaccumulator
RequiredSpec.
P. Tawdross and A. Königslide:8
The extrinsic environment
Dynamic Reconfiguration Approach The Extrinsic Design Environment
Optimization Library
Simulator SimulationLibraryngspice
netlist files
output files
RequiredSpec.
P. Tawdross and A. Königslide:9
The extrinsic environment
Dynamic Reconfiguration Approach The Extrinsic Design Environment
RequiredSpec.
Optimization Library
Simulator SimulationLibrary
SIM: Block Type, Specification, and weights
Opt.: Optimization method, and parameters
P. Tawdross and A. Königslide:10
The intrinsic environment
Dynamic Reconfiguration Approach The Intrinsic Design Environment
RequiredSpec.
Optimization Library
Hardware IntrinsicLibraryIntrinsic
library
Bit-stream Gene.
postprocessingobjectives
costaccumulator
P. Tawdross and A. Königslide:11
Hardware
RequiredSpec.
Optimization Library
IntrinsicLibrary
The intrinsic environment
Dynamic Reconfiguration Approach The Intrinsic Design Environment
Measurement circuit
selector
Chip
P. Tawdross and A. Königslide:12
The search space consists of n dimensions, each dimension represent a component (e.g., transistor, capacitor,...)
In the extrinsic approach switches are omitted and immediate aspectratio change by width modification is carried out.
In the intrinsic approach switches are implicitly includedEach transistor is programmable with integer value within the range of 1 to 257, which is the width of the transistorEach resistor or capacitor is programmable with integer value between 1 to 255
M1 M2 M3 M4 M5 M6........
Dynamic Reconfiguration ApproachRepresentation of the Optimization Problem
P. Tawdross and A. Königslide:13
( ) ( )idt
idt
idt
idt
gdt
idt
idt
idt
idt
vxxxprandCxprandCvwv
11
211 ()()
++
+
+=
−××+−××+×=
For all the particles:
Sensing elements
Sensing elements are used to detect environmental changes
An action is taken after any environmental change detection
• e.g. Re-evaluate all the bests
• Re-initialize a part of the population, and set the current position as the best to the rest
Dynamic Reconfiguration Approach PSO Dynamic Approach
P. Tawdross and A. Königslide:14
In addition to the standard technique advanced PSO methods are available, which we will study for their aptness in this application
There is more than one swarm in the optimisation
If two swarms are closed to each other, reinitialise one of them
Some particles are charged, charged particles dispel each other
Complexity f(N²)
( ) ( )
∑=
+
++
−×−
=
+−××+−××+×=N
j
jt
itj
tit
jiidt
it
it
gt
it
it
it
it
xxxx
QQa
axprCxprCvwv
131
1211
)(
()() rr
Dynamic Reconfiguration Approach Multi-swarm Particle Swarm Optimisation (MSPSO)
P. Tawdross and A. Königslide:15
The particle formation is a hierarchical tree
Each particle fly to its best and to the best of the particle above it
If a particle best fitness better than the particle above it, swap them
3 21
7654
Dynamic Reconfiguration Approach Hierarchical Particle Swarm Optimisation (HPSO)
The Global best Particle
Worse than All the particle
above
P. Tawdross and A. Königslide:16
The aspired generic sensor electronic front-end consists of basic reconfigurableblocks connected to each other with programmable switches
Target Hardware Basic Principle
Programmableswitches
Vout
-+Vin+
Vin-
Vout
-+Vin+
Vin-
P. Tawdross and A. Königslide:17
Case Study: Fixed Topology, Miller-OpAmp as an example for our approach designed by S.K. LakshmananA Miller- structure has been investigated first in Austriamicrosystems 0.35 µm CMOS technology [S.K. Lakshmanan].
Shift-Register for simple interface in first implementation
8
3
2
1
w=4
w=1
w=2
w=128
Shift
-Reg
iste
r
Target HardwareReconfigurable Miller - OpAmp
M5
Gnd Gnd
RL
Vdd
Vin + Vin -
Vout
Bias
CC
M1 M2
M3 M4
M6M7
P. Tawdross and A. Königslide:18
Experiments and ResultsExperimental Setup
DAQCard
CalibrationCircuits
P. Tawdross and A. Königslide:19
Vout
-
+Vin+
Vin-
1010-4Ts↑1010-4Ts↓101.5CMR0.1
103103
Weights Hi
0.1m
104104
Speci
offset
SR ↓SR↑
The chip is optimized at room temperature for 20 iterations, then, heating of is 150°C appliedThe global best position is used as a sensing elementC1=2, C2=2, and w=1 [j. Kennedy 97]
Experiments and ResultsDevice Deviations
P. Tawdross and A. Königslide:20
In the current setup, the temperature of the chip during the optimization is not exactly known Results are not repeatable
The temperature controlled oven should be used to heat all the chip to a known temperature
Experiments and ResultsIdeal Experiment
P. Tawdross and A. Königslide:21
Experiments and ResultsExperimental Results
Measurement input Signal for one particle
Good Particle
P. Tawdross and A. Königslide:22
0 20 40 60 80 100 120 140 160 180 200-4
-3
-2
-1
0
1
2
3
4
MSPSOHPSO
Converged fas ter at s tarting
More s table in dynamic environment
More rebels in dynamic environment
HPSO converged faster in static environment
MSPSO more stable in dynamic environment
Heating gradually to 150 °C
Average of 5 runs20 particle / population
Experiments and ResultsExperimental Results
P. Tawdross and A. Königslide:23
Intrinsic evolution is applied to optimize operational amplifier according to an industrial specificationPSO is applied to optimize a reconfigurable operational amplifier in a dynamic environment
Reconfigurable analog electronics allows rapid prototyping and scalabilityInherent dynamic fault-tolerance and self-healing (-x) capability
The complete list of specification values of the components will be included in the intrinsic evolution in futureOur approach will be extended to different amplifier types and there application for signal conditioning in a generic sensor electronicsfront-end
Conclusions
P. Tawdross and A. Königslide:24
Thanks