Grid e-services for Multi-Layer SOM Neural Network Simulation

(1)

Universidade Lusíada

V.N.Famalicão

Grid e-Services for Multi-Layer

SOM Neural Network Simulation

Grid e-Services for Multi-Layer

SOM Neural Network Simulation

Rui Lima

, Rui Silva

Faculdade de Engenharia

Universidade Lus

í

ada

4760

-

108 V. N. Famalicão, Portugal

{

{rml,rsilva

rml,rsilva}

}@fam.ulusiada.pt

@fam.ulusiada.pt

2007

(2)

Universidade Lusíada V.N.Famalicão

Outline

•

• Overview

Overview

•

• Multi

Multi

-

Layer SOM Background

•

• Architecture and Design

Architecture and Design

•

• Local Implementation

Local Implementation

•

• Grid Proposal Environment

Grid Proposal Environment

-

GRIDSOM

•

(3)

Overview

ILID

–

Instituto Lus

í

ada de Investiga

ç

ão e

Desenvolvimento

(

Lus

í

ada

R&D Institute)

1. Energy e Gestão Industrial (CLEGI)

2. Ciências Sociais e Humanas

3. Ciências Jurídicas

4. Arquitectura

5. Centro Lusíada de Matemática

6. Centro Lusíada de Recursos Humanos (CIDRHU)

7. Centro Lusíada de Turismo, Inovação e Serviços (TIS)

8. Centro Lusíada de Economia e Gestão

Intelligent Tool

Wear Monitoring

http://www.fam.ulusiada.pt/ilid/MIFC/index.htm

Vila Nova de Famalicão

8 - R&D Centers

(4)

V.N.Famalicão

Intelligent Tool Wear Monitoring

Wear monitoring

Cutting Process

(5)

Our Problem

Experiments were

conducted on a CNC

Non intrusive sensors

Modeling technique

based on artificial

neural networks

(6)

V.N.Famalicão

Multi-Layer SOM Background

Division of the output map into several sub

-

maps

1.

1. Randomly initiate weights;

Randomly initiate weights;

2.

2. Present input vector;

Present input vector;

3.

3. Determine the distance

Determine the distance

between input feature vector

and each output node;

4.

4. Adjust weights according to

Adjust weights according to

winning frequency;

5.

5. Select neuron with

Select neuron with

minimum distance to input

feature vector;

6.

6. Update weights;

Update weights;

7.

7. Back to step

Back to step

2.

2 .

Clustering Method

(7)

V.N.Famalicão

Multi-Layer SOM Results

The sub

-

maps

are treated as

jobs in the

GRID

computing

environment

0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00

Contour plot - tool with VBB=0.05 mm Contour plot - tool with VBB=0.17 mm

Contour plot - tool with VB_B=0.21mm _{Contour plot - tool with VB}

(8)

V.N.Famalicão

Our Aim

A

lightweight

architecture to support the execution of

neural networks simulation tasks across the Grid

infrastructure.

• Security

• Authentication

• Encryption

• Model Design

• Build Simulation

• Monitor

• Results

(9)

V.N.Famalicão

Architecture and Design

LCG

Job Manager

gLite

Grid middleware

SL (

SL (Linux

Linux

)

Operating system

P

IV

2GHz

Architecture (processor type)

1 TB

Cluster Storage

80 GB

Disk per node

512 MB

Memory per Node

15

15 Number of Nodes

Local Resources:

(10)

V.N.Famalicão

Architecture and Design

Layer Model

(11)

V.N.Famalicão

Architecture and Design

(12)

V.N.Famalicão

Neural Network Agent

• Manage and control tasks.

• Distribution of work load.

• Control neural network learning and

recognition.

• Interface the communication between neural

network layers.

(13)

V.N.Famalicão

Monitor Agent

• Submit jobs that are waiting in the

Database.

• Querying the Logging and Bookkeeping.

• Collect those that have already been

successfully finished.

• Re-submit unfinished jobs.

• Update the database with the jobs states.

Local Service

Local Database

(14)

Hybrid Approach

•

• Network Agent

Network Agent

_{Simulation Algorithm}

Database Store Results

_{Database Store Results}

•

• Monitor

Monitor

_{Looks for new Jobs in Database}

_{Launch Jobs to the GRID}

(15)

GRIDSOM

Prototype

http

://

gridsom.fam.ulusiada.pt

WEB

Registration

(16)

GRIDSOM

Start Simulation

(17)

V.N.Famalicão

Generate .JDL File

(18)

Tasks in Database

Current Tasks

–

In the Outbox

•

• The Neural Network Agent will queue the

The Neural Network Agent will queue the

initial job of this task

(19)

V.N.Famalicão

Monitor Launch Jobs

(20)

V.N.Famalicão

Queued Jobs in Database

Job status stored in (

(21)

V.N.Famalicão

Monitor Getting Job Status

Job re

-

submit !

Job Done !

(22)

V.N.Famalicão

Monitor Getting Results

Neural Network Agent

–

Generate a new iteration

Job Done !

Monitor

–

Submit this new iteration:

Until the

end of

(23)

V.N.Famalicão

End Simulation

Neural Network Agent

(24)

V.N.Famalicão

User Collect Results

Download de Final Data

(25)

Data Visualization

Data Analysis

(outside GRIDSOM)

(26)

V.N.Famalicão

Conclusions and Future Developments

• Prototype successfully tested

• Hybrid approach

• SSL layer

• Migrate to New Scheduling System

•

• Virtual Organization

Virtual Organization

•

(27)

V.N.Famalicão

Many Thanks