Study on the Application of a New Prediction Model Based on Penalty
Constraint to Flotation Process
Zhang Yong
, Liu Xuqiang
Electronic and Information Engineering, University of Science and Technology Liaoning, Anshan 114051,China
E-mail: zy9091@163.com
Abstract: The flotation process is a complicated multi-input and multi-output process with the characteristic of strong
non-linearity, heavy coupling and large delay. Due to the difficulty of measuring the concentrate grade and tailing grade
index online, and its dynamics varying with the process conditions, such a control objective by far is difficult to achieve
by the existing control methods to control the product quality indices into their technical targeted ranges and even cause
fault work-condition. This paper presents a clustering algorithm based on punishing constraint of swarm intelligence
(PCSI). Directed by the nature of PSO, PCSI could randomly search the centers of clusters and obtain the number of
clusters. The process prior knowledge and PCA method are used to reduce dimension of the input data and select
auxiliary variables. And then a new hybrid recursive algorithm of RBFNN based on simplified rival penalized
competitive learning method (SRPCL) to make an adaptive clustering is developed. The method proposed has
successfully been applied to two production lines of a mineral processing plant of Anshan Iron and Steel Group
Corporation, and its effectiveness is proved evidently.
Keywords: flotation process; adaptive penalty constraint; principal component analysis (PCA); particle swarm
optimization algorithm (PSO); RBF neural network
This work is supported by National Nature Science Foundation of China(61473054)
1
INTRODUCTION
Flotation is a widely used mineral processing technology, it
is a method [1] of mineral separation with useful mineral
(mainly
34
Fe O
) and gangue minerals (mainly
2
iSO
) in
the pulp fully dispersed under condition, according to the
different surface properties of various minerals, from the
pulp with bubble buoyancy. A typical plant usually has more
than two such production lines, with an annual output of 4
million tons of concentrate, billions of output value, the
typical production line has high concentrate screening, high
efficiency mixing, flotation slot, scavenging tank, magnetic
separation system, dosing system, tailings return system and
other multi-channel production processes. The main
features are: 1) The procedure is very much in production
process, long process, complex process; 2) Production line
is a multi slot series, multi capacity object of intercoupling;
3㸧Anti flotation process is a typical process of continuous
production, normal and stable operation of the production
process involves the economic and technical indicators of
iron ore output, the quality of their products related to the
enterprise profit realization, these factors are decided by the
integrated effect of multiple variables of the process. In the
flotation process, the recovery rate of ore concentrate only
on the basis of the off-line analysis value adjustment,
hysteresis is very great, accuracy is not high, it is difficult to
achieve closed-loop control, so researching inference
estimation technique of the concentrate grade and recovery
rate of flotation with flotation in urgent need[2]. Based on
the above reasons, this paper puts forward an inference
estimation strategy about the parameters of the flotation
economy, which is based on the combination of adaptive
penalty particle swarm and the new RBF network. First of
all having a primary election for process variables according
to the process mechanism and experience knowledge, then
having the selection and dimension reduction of auxiliary
variables for network high dimension input vector in many
related flotation process detection variables with the method
of principal component analysis. This method has faster
learning speed and better dynamic tracking ability with
making full use of the current measurement data, finally
obtaining good estimation performance with the algorithm
is applied to estimation of economic indicators of anti
flotation in mineral processing plant.
2
SWARM INTELLIGENCE CLUSTERING
BASED on ADAPTIVE CONSTRAINT
PENALTY
The most important task is finding the clustering number
and clustering center of each category, K-means clustering
algorithm to exhaustive search the whole training set, each
class of repeated calculation of the distance between the
center and all samples, constantly updating the class center,
to minimize the sum of the distance between all samples and
their class centers, the higher the computational complexity
with the higher the number of samples or the higher the
feature vector dimension of the sample.
The objective function of the particle swarm optimization is
the mean square error of all the particle positions in the
population of the particle and the center of the class, and the
clustering process minimizes the mean square error[3].
2
2
11
11
min ( ) exp , (1)
2
MJ
D
ij j ij
ij
fz x z x R
N
λ
==
ªº
§·
=−− ∈
¨¸
«»
©¹
¬¼
¦¦
1
978-1-5090-4657-7/17/$31.00
c
2017 IEEE