More updates to Taguchi optimisation docs.

这个提交包含在:
Craig Warren
2016-05-04 10:10:11 +01:00
父节点 7539510b43
当前提交 c80b7ab441
共有 7 个文件被更改,包括 104 次插入4 次删除

查看文件

之前

宽度:  |  高度:  |  大小: 443 KiB

之后

宽度:  |  高度:  |  大小: 443 KiB

查看文件

之前

宽度:  |  高度:  |  大小: 1.1 MiB

之后

宽度:  |  高度:  |  大小: 1.1 MiB

二进制文件未显示。

之后

宽度:  |  高度:  |  大小: 45 KiB

查看文件

之前

宽度:  |  高度:  |  大小: 92 KiB

之后

宽度:  |  高度:  |  大小: 92 KiB

查看文件

@@ -22,7 +22,7 @@ Information
AustinMan and AustinWoman (http://bit.ly/AustinMan) are open source electromagnetic voxel models of the human body, which are developed by the Computational Electromagnetics Group (http://www.ece.utexas.edu/research/areas/electromagnetics-acoustics) at The University of Texas at Austin (http://www.utexas.edu). The models are based on data from the National Library of Medicine’s Visible Human Project (https://www.nlm.nih.gov/research/visible/visible_human.html).
.. figure:: images/AustinMan_head.png
.. figure:: images/user_libs/AustinMan_head.png
:width: 600 px
FDTD geometry mesh showing the head of the AustinMan model (2x2x2mm^3).
@@ -68,7 +68,7 @@ To insert a 2x2x2mm^3 AustinMan with the lower left corner 40mm from the origin
For further information on the `#geometry_objects_file` see the section on object contruction commands in the :ref:`Input commands section <commands>`.
.. figure:: images/AustinMan.png
.. figure:: images/user_libs/AustinMan.png
:width: 300 px
FDTD geometry mesh showing the AustinMan body model (2x2x2mm^3).

查看文件

@@ -22,7 +22,7 @@ Information
#
# Please use the attribution at http://dx.doi.org/10.1190/1.3548506
The package features an optimisation technique based on Taguchi's method. It allows the user to define parameters in an input file and optimise their values based on a fitness function.
The package features an optimisation technique based on Taguchi's method. It allows users to define parameters in an input file and optimise their values based on a fitness function, for example it can be used to optimise material properties or geometry in a simulation.
Taguchi's method
@@ -44,11 +44,13 @@ Package overview
.. code-block:: none
antenna_bowtie_opt.in
OA_9_4_3_2.npy
OA_18_7_3_2.npy
optimisation_taguchi_fitness.py
optimisation_taguchi_plot.py
* ``antenna_bowtie_opt.in`` is a example model of a bowtie antenna where values of loading resistors are optimised.
* ``OA_9_4_3_2.npy`` and ``OA_18_7_3_2.npy`` are NumPy archives containing pre-built OAs from http://neilsloane.com/oadir/
* ``optimisation_taguchi_fitness.py`` is a module containing fitness functions. There are some pre-built ones but users should add their own here.
* ``optimisation_taguchi_plot.py`` is a module for plotting the results, such as parameter values and convergence history, from an optimisation process when it has completed.
@@ -58,7 +60,7 @@ Implementation
The process by which Taguchi's method optimises parameters is illustrated in the following figure.
.. figure:: images/taguchi_process.png
.. figure:: images/user_libs/taguchi_process.png
:width: 300 px
Process associated with Taguchi's method.
@@ -93,3 +95,8 @@ Example
The following example demonstrates using the Taguchi optimisation process to optimise values of loading resistors used in a bowtie antenna. The bowtie design features 3 slots in each arm of the bowtie where loading resistors are placed, and a substrate with a perimittivity of 4.8 is used. The antenna is modelled in free space, and an output point (the electric field value) is specified at a distance of 60 mm from the feed of the bowtie.
.. figure:: images/user_libs/antenna_bowtie_opt.png
:width: 600 px
FDTD geometry mesh showing bowtie antenna with slots and loading resistors.

查看文件

@@ -0,0 +1,93 @@
#taguchi:
optparams['resinner'] = [0.1, 5000]
optparams['resmiddle'] = [0.1, 5000]
optparams['resouter'] = [0.1, 5000]
fitness = {'name': 'compactness', 'stop': 30, 'args': {'outputs': 'Ex60mm'}}
maxiterations = 5
#end_taguchi:
#python:
import numpy as np
from gprMax.input_cmd_funcs import *
title = 'antenna_bowtie_opt'
print('#title: {}'.format(title))
domain = domain(0.180, 0.120, 0.160)
dxdydz = dx_dy_dz(0.001, 0.001, 0.001)
timewindow = time_window(5e-9)
fr4_dims = (0.120, 0.060, 0.002)
bowtie_dims = (0.050, 0.040) # Length, height
flare_angle = np.arctan((bowtie_dims[1]/2) / bowtie_dims[0])
tx_pos = (domain[0]/2, domain[1]/2, 0.050)
# Vertical slot positions, relative to feed position, i.e. txpos[0]
vcut_pos = (0.014, 0.027, 0.038)
# Loading resistor values
res = np.array([optparams['resinner'], optparams['resmiddle'], optparams['resouter']])
rescond = ((1 / res) * (dxdydz[1] / (dxdydz[0] * dxdydz[2]))) / 2 # Divide by number of parallel edges per resistor
# Materials
material(4.8, 0, 1, 0, 'fr4')
for i in range(len(res)):
material(1, rescond[i], 1, 0, 'res' + str(i + 1))
# Source excitation and type
print('#waveform: gaussian 1 2e9 mypulse')
print('#transmission_line: x {:g} {:g} {:g} 50 mypulse'.format(tx_pos[0], tx_pos[1], tx_pos[2]))
# Output point - distance from tx_pos in z direction
print('#rx: {:g} {:g} {:g} Ex60mm Ex'.format(tx_pos[0], tx_pos[1], tx_pos[2] + 0.060))
# Bowtie - upper x half
triangle(tx_pos[0], tx_pos[1], tx_pos[2], tx_pos[0] + bowtie_dims[0], tx_pos[1] - bowtie_dims[1]/2, tx_pos[2], tx_pos[0] + bowtie_dims[0], tx_pos[1] + bowtie_dims[1]/2, tx_pos[2], 0, 'pec')
# Bowtie - upper x half - vertical cuts
for i in range(len(vcut_pos)):
for j in range(int(bowtie_dims[1] / dxdydz[2])):
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] - bowtie_dims[1]/2 + j * dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] - bowtie_dims[1]/2 + j * dxdydz[1], tx_pos[2], 'free_space')
# Bowtie - upper x half - vertical cuts - loading
for i in range(len(vcut_pos)):
gap = ((vcut_pos[i] * np.tan(flare_angle) * 2) - 4*dxdydz[1]) / 5
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] - (1.5 * gap) - dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] - (1.5 * gap) - dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] - (1.5 * gap) - 2*dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] - (1.5 * gap) - 2*dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] - (0.5 * gap), tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] - (0.5 * gap), tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] - (0.5 * gap) - dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] - (0.5 * gap) - dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] + (0.5 * gap), tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] + (0.5 * gap), tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] + (0.5 * gap) + dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] + (0.5 * gap) + dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] + (1.5 * gap) + dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] + (1.5 * gap) + dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] + vcut_pos[i], tx_pos[1] + (1.5 * gap) + 2*dxdydz[1], tx_pos[2], tx_pos[0] + vcut_pos[i] + dxdydz[0], tx_pos[1] + (1.5 * gap) + 2*dxdydz[1], tx_pos[2], 'res' + str(i + 1))
# Bowtie - lower x half
triangle(tx_pos[0] + dxdydz[0], tx_pos[1], tx_pos[2], tx_pos[0] - bowtie_dims[0], tx_pos[1] - bowtie_dims[1]/2, tx_pos[2], tx_pos[0] - bowtie_dims[0], tx_pos[1] + bowtie_dims[1]/2, tx_pos[2], 0, 'pec')
# Bowtie - lower x half - cuts for loading
for i in range(len(vcut_pos)):
for j in range(int(bowtie_dims[1] / dxdydz[2])):
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] - bowtie_dims[1]/2 + j * dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] - bowtie_dims[1]/2 + j * dxdydz[1], tx_pos[2], 'free_space')
# Bowtie - lower x half - vertical cuts - loading
for i in range(len(vcut_pos)):
gap = ((vcut_pos[i] * np.tan(flare_angle) * 2) - 4*dxdydz[1]) / 5
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] - (1.5 * gap) - dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] - (1.5 * gap) - dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] - (1.5 * gap) - 2*dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] - (1.5 * gap) - 2*dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] - (0.5 * gap), tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] - (0.5 * gap), tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] - (0.5 * gap) - dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] - (0.5 * gap) - dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] + (0.5 * gap), tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] + (0.5 * gap), tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] + (0.5 * gap) + dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] + (0.5 * gap) + dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] + (1.5 * gap) + dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] + (1.5 * gap) + dxdydz[1], tx_pos[2], 'res' + str(i + 1))
edge(tx_pos[0] - vcut_pos[i] - dxdydz[0], tx_pos[1] + (1.5 * gap) + 2*dxdydz[1], tx_pos[2], tx_pos[0] - vcut_pos[i], tx_pos[1] + (1.5 * gap) + 2*dxdydz[1], tx_pos[2], 'res' + str(i + 1))
# PCB
box(tx_pos[0] - fr4_dims[0]/2, tx_pos[1] - fr4_dims[1]/2, tx_pos[2] - fr4_dims[2], tx_pos[0] + fr4_dims[0]/2, tx_pos[1] + fr4_dims[1]/2, tx_pos[2], 'fr4')
# Detailed geometry view of PCB and bowtie
#geometry_view(tx_pos[0] - fr4_dims[0]/2, tx_pos[1] - fr4_dims[1]/2, tx_pos[2] - fr4_dims[2], tx_pos[0] + fr4_dims[0]/2, tx_pos[1] + fr4_dims[1]/2, tx_pos[2] + dxdydz[2], dxdydz[0], dxdydz[1], dxdydz[2], title + '_tx', type='f')
# Geometry view of entire domain
#geometry_view(0, 0, 0, domain[0], domain[1], domain[2], dxdydz[0], dxdydz[1], dxdydz[2], title)
#end_python: