optimization of rebar production process in metallurgy. Uses data analysis tools like Artificial Neural Network (ANN), Clustering, Fuzzy logic, Multiple Regression to arrive at the best solution to manufacturing of rebar steels.
2. • Explore the feasibility of a predictive model for property estimation in NBM.
Current status:
Theoretical
model based on
heat transfer
calculations
which is
highly
inaccurate.
The operator
uses his
experience.
YS distribution
After Implementation:
A machine learning model trained with all possible data can be used for online
prediction of controlling parameters.
4. Approach
WB1flow rate
WB2 flow rate
CS=16
mm
500D
UTS/YS>1.15
Billet Temperature
YS
UTS
Mill Speed
CS=12
mm
Carbon
CS=10
mm
Manganese
5. YS vs FRT
CE was found to be insignificant in
predicting YS.
Multiple Regression did not give any
satisfactory results.
Problem with YS: Taken from any
random point of rebar.
500D,16mm
6. The operator operates at
atleast 2 different ranges of
water flow rates.
Towards Clustering
• Fuzzy clustering
• K-means clustering
7. K means clustering:
• 3 clusters created by taking into
consideration WB1, WB2, FRT
• All the clusters had similar range of FRT
and Billet Temperature but differed in
WB1 and WB2 ranges.
Multiple Linear
Regression
•
•
Satisfactory relation was obtained only
for cluster which contained low values of
water box flow rate implying negligible
leakage.
The results of Cluster Analysis varied for
different data periods.
9. •
•
BLT
FRT
•
WB1
•
•
•
WB2
WB1 Membership Fcn
1
2
WB1 values
3
Takes into account the
error in taking readings.
Inputs and outputs are
assigned membership
functions.
Membership functions
create fuzzy sets.
Results:
RMSE in FRT ~ 11.6C
Predicted values are
concentrated around the
mean
10. ANN Architecture: ANN with 1 hidden layer having 10 neurons.
WB1
WB2
Billet Temp
Speed
FRT
Linear transfer function
Training Data=70%
Validation Data=15%
Test Data=15%
Backpropagation Learning Algorithm:
Works by minimizing error with respect
to weights.
12. • ANN model predicted extremely well (r2>0.84) for unseen test data on which it was
trained.
• Predicted well (r2>0.60) for untrained data which followed a recognizable pattern.
• Failed to predict data which didn’t follow any trained pattern.
• This showed that data varied considerably with time.
• Nozzle wear out increases with time.
13. Cooling in Nozzles
Nozzle
Bore
Diameter
Rebar
Diameter
Sufficient water pressure keeps the
rebar floating in the water channel and
do uniform circumferential cooling
Nozzle wear out
Insufficient pressure built up
could not hold the bar in place
and thus uniform
circumferential cooling is not
possible
14. • Efficiency of water quenching decreases with time due to nozzle wear-out.
A dynamic fuzzy based ANN model with sufficient training can be developed
for online prediction of product properties within a range of +-5 MPa.
Plant Recommendations :
• Leakage minimization.
• Identification of sample taken for tensile testing to correctly
map YS with corresponding FRT.
• Bringing uniformity among operators in different shifts to
reduce data variation with time.
Editor's Notes
A model was developed based upon theortetical calculations but failed in the complex industrial process environment. It is only the experience of the operator. By analysing the data we can develop an automated model
Layout of the entire nbm with billets entering and finishedrebars being delivered from here. The billet is passed through a reheating furnace to heat it to temperatures about 1040C. Passes through a series of horizontal and vertical rollers then through ……splits into line A and B. ……water boxes(maximum of 6000)……YS is obtained, data collection: the point corresponding to the YS testing is located. We go back in time and map it with the corresponding wbs,frt and blt .20mdistance b/w frt and wb2.
Our final goal is to get a higher u/y ratio. Nbm produces rebars of 3 diff css. Both of them vary in a short range
Variation of YS vs CE is Statistically insignificant because Varied within a very short range from 0.33 to 0.37R square of 0.13.Reason : we gather YS from a random point of each sample. We cannot correlate it with FRT, WB,BLT data as we don’t know the point it is taken from. So we’ll shift our attention to finding b/w FRT and WB1,WB2 BLT
This cluster is for the ideal case. ie no leakages and 100% efficiency
Fuzzy approach was a failure
Prediction with ann if trained for the dayTransfer fcn