An adaptive artificial neural network to model a cu pb zn flotation circuit
Download
1 / 40

An Adaptive Artificial Neural Network to Model a Cu/Pb/Zn Flotation Circuit - PowerPoint PPT Presentation


  • 111 Views
  • Uploaded on

An Adaptive Artificial Neural Network to Model a Cu/Pb/Zn Flotation Circuit. Saiedeh Forouzi and John A. Meech University of British Columbia Department of Mining and Mineral Process Engineering. Introduction.

loader
I am the owner, or an agent authorized to act on behalf of the owner, of the copyrighted work described.
capcha
Download Presentation

PowerPoint Slideshow about ' An Adaptive Artificial Neural Network to Model a Cu/Pb/Zn Flotation Circuit' - imelda-byers


An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript
An adaptive artificial neural network to model a cu pb zn flotation circuit

An Adaptive Artificial Neural Network to Model a Cu/Pb/Zn Flotation Circuit

Saiedeh Forouzi and John A. Meech

University of British Columbia

Department of Mining and Mineral Process Engineering


Introduction
Introduction

  • Massive sulfide deposit of Zn, Pb, Cu and Ag located close to Bathurst in New Brunswick

  • Ore production began in 1964 with a milling rate of 4500 tpd

  • More than 80 million tonnes have been processed with about 55 million tonnes remaining

  • The mill capacity has increased to 10,500 tpd


Introduction continued
Introduction - continued

9.2% Zn

Zn Conc. (52.5%)

Pb Conc. (40%)

3.6% Pb

Process

0.36% Cu

105 g/t Ag

Cu Conc. (23.5%)

Bulk Conc.

(39% Zn + 19% Pb)

Tailing


Concentrator flow sheet
Concentrator Flow Sheet

Crushing Plant

Lines1&2 Fine Ore Bins

Lines3 Fine Ore Bins

Line 1&2 Grinding

Line 3 Grinding

L3 CuPb

L1&2 CuPb

L3 Zn

L1&2 Zn

Cu Sep.

Pb Upgrd

Bulk Float

Tailings

Pb Conc.

Cu Conc.

Bulk Conc.

Zn Conc.


Project background
Project Background

  • process inputs include tonnage rate, water addition, particle size distribution, reagents, flotation cell levels

  • process outputs are Cu/Pb/Zn grades of concentrates and tailing

  • process setpoints for final & intermediate variables are not fixed because of

    - variable head grades

    - changable smelter contracts


Project background continued
Project Background (continued)

  • instrumentation

    - XRF on-stream analyser - particle size monitor

    - tonnage weigh scale - pulp flowmeters

    - pulp density gauges - reagent flowmeters

    - pH meters - thermocouples

    - flotation cell levels - air flowmeters

  • control systems

    - “grind” control - tonnage control

    - cell level control - reagent control


Project objectives scope
Project Objectives & Scope

Long Term

  • use smelter contracts and headgrades to find best achievable product grades and establish setpoints for control variables

    Short Term

  • establish an Artificial Neural Network Model to predict product assays for Line 3 Cu/Pb circuit

    to provide proof-of-concept of the model


General concentrator flow sheet
General Concentrator Flow Sheet

Crushing Plant

Lines1&2 Fine Ore Bins

Lines3 Fine Ore Bins

Line 1&2 Grinding

Line 3 Grinding

L3 CuPb

L1&2 CuPb

L3 Zn

L1&2 Zn

Cu Sep.

Pb Upgrd

Bulk Float

Tailings

Pb Conc.

Cu Conc.

Bulk Conc.

Zn Conc.


Cu pb flotation circuit line 3
Cu/Pb Flotation Circuit - Line 3

Aeration

2nd Clnr

1st Clnr

Roughers

From Grinding

Cu-Pb

To Zn

Conc.

Circuit

Regrind

Mill


Methodology
Methodology

  • Develop an ANN model to represent the relationships between the I/O variables

  • Update the model as required to reflect relationship changes and maintain accuracy and robustness

  • Develop a fuzzy algorithm to decide when to adapt the model


Tools
Tools

  • G2, from Gensym, a software tool for intelligent real-time system development

  • GDA and NeurOnline are related tools

  • provide direct access to real-time data

  • object-oriented software and user friendly

  • available for use at Brunswick

  • graphical output and configuration features


Artificial neural networks
Artificial Neural Networks

  • Inspired from neuronal structure of the human brain

  • Learning and Recall processes in neuron cell connections

  • Network consists of several layers of processing element

X1

n

Wj1

Sj =  WjiXi

X2

Wj2

Yj

.

i=1

Sj

f(Ij)

.

.

Yj = f(Sj) =f( WjiXi)

Wjn

Xn


Ann activation function
ANN - activation function

Sigmoid function

- most popular method

- ouput signal scaled

between 0 and 1

- a convenient

differentiable form

f '(Sj) = ej (1- ej )

-Sj

f(Sj)= 1/ (1+ e )


Ann model architecture
ANN Model Architecture

  • Input, output and hidden layers

Wji

Wkj

X1

n

Hj = f (Ij) = f (  WjiXi)

i=1

X2

Y1

1j  m

.

.

.

.

.

.

.

.

.

m

k=1...p

i=1...n

j=1...m

Yk = f (Ik) = f (  WkjHj)

Xn

Yp

j=1

1k  p

Bias

_

p

Input

Hidden

Output

E = 1/p  ( Yk - Yk)

k =1


Ann models features
ANN Models - features

Advantages:

  • generates knowledge by learning from actual data

  • able to perform massive parallel processing

  • able to model very complex nonlinear problems

    Drawbacks:

  • lots of data may be required

  • data collection plays a vital role

  • Learning can be slow for large networks


Ann models issues
ANN Models - issues

  • quality of data and reliability of instrumentation

  • separation of data into training and testing sets

  • lots of data may be required

  • data preprocessing (filtration, synchronization) is vital

  • scaling of all data is essential

  • Learning can be slow for large networks

  • stability of process relationships

  • design issues (bias, hidden nodes, algorithm, learning rate)

  • separate networks for predictive models

  • inclusive networks for classification models


Ann models at bms
ANN Models at BMS

  • Number of networks (12 separate models)

  • Variables: actual values and changes in value

Inputs (58)

Output

Values of control and load variables

Model

Change in Assay

Changes in control and load variables

Actual Assay


Data pre processing for ann model
Data Pre-processing for ANN Model

  • Phase lags determined by mass flow and

    process capacity

Input

Output

Time


Data pre processing for ann model1
Data Pre-processing for ANN Model

  • Phase lags as a function of mass flow

  • All data are scaled between 0 and 1


Ann model at bms
ANN Model at BMS

  • Phase lags as a function of mass flow

  • All data are scaled between 0 and 1

  • Use sigmoid function in all layers

  • Data set size (~1300 records)

  • Random data separation for training and testing

  • Error calculation on both training and testing data


Adaptive ann model
Adaptive ANN Model

  • process relationships are never fixed

  • How often is retraining required?

  • Which data to use for retraining?

  • establish an intelligent algorithm to answer these questions


Adaptive model at bms

Process

Adaptive Model at BMS

New Model

Predicted Output

ANN Model

-

Error

Inputs

+

Actual Output

Retraining Algorithm

Data file Updating

Data Validation


Adaptive model at bms1
Adaptive Model at BMS

Current Data at t(adjusted phase lags)

Actual Inputs at t

Changes in Inputs from t-1

Actual Assay at t-1

Predicted Change in the Assay at t

ANN Model

Predicted Assay at t

Actual Assay at t-1

Actual Assay at t

Predefined

Threshold

>

Error &

Cumulative Error

Actual Change in Assay at t

Predicted Change in Assay at t


Adaptive model at bms2
Adaptive Model at BMS

Case One - New Data Set

Output

Inputs

Current Data

Max

Old Data

Min


Adaptive model at bms3
Adaptive Model at BMS

Case One - New Data Set

Output

Inputs

Max

Data set for retraining

Min


Adaptive model at bms4
Adaptive Model at BMS

Case Two - Similar Data Set

Output

Inputs

Current Data

Max

Old Data

Min


Adaptive model at bms5
Adaptive Model at BMS

Case Two - Similar Data Set

Output

Inputs

Max

Data set for retraining

Min


Adaptive model at bms6
Adaptive Model at BMS

Case Three - Similar input/different output

Output

Inputs

Current Data

Max

Old Data

Min


Adaptive model at bms7
Adaptive Model at BMS

Case Three - Similar input/different output

Output

Inputs

Max

Data set for retraining

Min


Adaptive model at bms8
Adaptive Model at BMS

Case Four - New data set but data file is full

Output

Inputs

Current Data

Max

Old Data

Min


Adaptive model at bms9
Adaptive Model at BMS

Case Four - New data set but data set is full

Output

Inputs

Max

Data set for retraining

Min


When to retrain
When to Retrain?

  • The error in the model

  • The amount of new data

High Error

Low Error

No retraining

High

Low

Retraining

% of New Data


Compensating feed forward model based control

Process

Compensating Feed Forward Model-Based Control

New Setpoints

Predicted Output

Knowledge Base

ANN Model

Inputs

Actual Output


Compensating feed forward model based control1
Compensating Feed Forward Model-Based Control

  • Knowledge about the control variables and their relationships

  • Using the knowledge of experts

  • Setting rules to change setpoints of control variables to obtain desired output

  • Using the model to test established rules


Implementation schedule
Implementation Schedule

Description

Oct

Dec

Nov

Organizing Data Filtering and Phase lags

Model Training and Testing ANN Model

Setting Up On-Line Data Collection

Creating and Testing On-Line Adaptive System


Advantages
Advantages

  • A self-adaptive model which represents the current process

  • Having a model to predict process outputs under any condition

  • Accounting for economics in establishing setpoint to achieve higher efficiencies

  • Better process control by increased flexibility


Results
Results

  • Original model gave the following errors:

    RMSE = 0.107 for Training Data

    RMSE = 0.181 for Testing Data

  • Retraining began after less than 8 hours

  • New model gave

    RMSE = 0.103 for Training Data

    RMSE = 0.443 for Testing Data

  • After 1 month, RMSE settled into the range of

    0.1-0.2 and retraining periods were ~ 7 days


Potential problems
Potential Problems

  • Data should be derived from testwork

  • Once supervisory control is implemented, the process relationships will be masked

  • Instrumentation reliability issues exist

  • Real-time issues with 12 ANN models running in parallel

  • Identification of "optimum" set points


Conclusion
Conclusion

  • ANN modeling of plant data with G2 is feasible and practical

  • Adaptation of the model can occur in real-time

  • Model-based supervisory control can now be tested and implemented

  • ANN can be used to identify the important control variables in the process


Recommendations
Recommendations

  • Application of the model to the remaining circuit assays should proceed

  • Development of the supervisory control knowledge base should proceed

  • Database updating should be done using forced variations in the process to ensure the discovered relationships are valid


ad