Scientific poster VCO summer school 2012
-
Upload
alvaro-gil -
Category
Education
-
view
244 -
download
0
Transcript of Scientific poster VCO summer school 2012
MotivationThis project aims to investigate new manufacturing strategies for sawmills in order to produce application-based hardwood lumber, instead of the traditional methods based on NHLA classi�cation, so as to increase the rate of lumber recovery and hence, maximize the pro�t.
Agent-Based Approach
Classi�cation trees are nonparametric statistical learning methods that incorporate feature selection and interactions to predict the membership of each observation into a categorical variable. Each branch of a classi�cation tree can be interpreted as a rule, so it can be considered as a set of rules (see example for cabinets below)
Overview
Typical Transformation of Hardwood
1. Floor 2. Wardrobe 3. Staircase 4. Panel 5. Cabinet 6. Molding 7. Pallete
1. Sorting 2. Sawing
There are 5 big steps from the harvesting to the �nal users
At the sawmill level there are four main process
Several research projects have been done to study sawing, drying and planing
This research focuses in the sorting process in order to improve log allocation to the next pro-cesses by considering the needs of the second transformation
It’s necessary to �rst identify the logs attributes that increase the yield of each type of transfor-mation.
Data BaseThe analysis starts from a data
base with several logs mea-sures (inputs) and it’s procure-ment cost ($/Pmpm) and yield
(m3/mpm) for each type of transformation (outputs).
In order to identify the best logs for each
category, it is neces-sary to identify an es-
timation method. However, the predic-
Classi�cation Trees
These binary matrices are used to built the classi�cation trees (one for each category), which are represented as equations that can be transformed into classi�cation diagrams
Inpu
tO
utpu
t
Cost distribution Percentiles Cabinet sample Threshold classi�cation
tion accuracy of traditional methodologies is very low (<50%).The data-mining �eld o�er some interesting approaches for
solving this issue, like classi�cation trees
This procedure allows to construct a set of binary matrices for each percentile threshold, where each row classi�es the membership of each log to each
threshold and category (F, W, S, P, C, M and T)
3. Drying
4. PlaningFinal productready to second transformation
Classi�cation Diagram for Cabinets
They can also be used to build a joint classi�cation grid by using only 5 of the set of the original 13 attributes:
By using the classi�cation grid, it is possible to transform the reception process into a demand-driven approach.
Log Handling and Sorting Implementation
Classi�cation Grid
Preliminary Results
Classi�cation accuracy: 81.5% Procurement cost reduction: 44%
This methodology implies however an extra work force at the logs reception as well as more loaders and space for storage, which means an increase in the handling costs
In order to evaluate this, we use agent-based simulation to test several manufacturing control strate-gies in order to �nd the best way to implement the solution.
Before After
All categories pile
Category 1 pile Category 2 pile
Category 3 pile Category 4 pile
The actual phase of the project is the construction of a simulation platform which combine discrete event with agent-based simulation.
The discrete event approach simulates the push process : Sawing, Drying and PlanningThe agent-based approach simulates the decision process of the sorting and allocation of logs by using the classi�cation grid for the sawing according to an external and random demand (pull).
External demand in terms of secondary
transformation
Planning AgentRecieve the exeternal
demand and the internal processing information
Secondary transformation
market
Final storage ready to be sold
Sawing / Planin / Drying (PUSH)
Loader AgentTakes decisions about logs
sorting and allocation
Reception process
Sorting process(piles by category)
A preliminary discret-event simulation evaluation leads to an average procurement cost re-duction of 44%. In the next phase of the research project, we are building an agent-based simulation in order to evaluate the impact of various log handling and sorting techniques to implement these classi�cation trees.
Saw
mill
Optimization of Sorting and Allocation Activities in Sawmills with Data-Mining Techniques and Agent-Based Modeling
Alvaro GilM.Sc. Student École Polytechnique de MontréalSupervisor: Jean-Marc Frayret, Ph.D.