On the Application of Artificial Intelligence Techniques to the Quality
Improvement
of Industrial Processes
P. Georgilakis N. Hatziargyriou
Schneider Electric National Technical University of Athens
Greece Greece
Contents
1. Quality improvement with Decision Trees (DT)
2. Quality improvement with Hybrid DT-NN Classifiers
3. Industrial applications
4. Discussion of results
5. Conclusions
Quality improvement with Decision Trees
The DT is built on the basis of a learning set (LS).
The LS comprises a number of preclassified (e.g., acceptable or non-acceptable) measurement sets (MS) defined by a list of candidate attributes.
A suitable test is sought to divide a node, by applying the optimal splitting rule.
tATi:
The optimal splitting rule decides what is the best attribute and its threshold value, so that the additional information gained through that test is maximized.
The best attribute and its threshold value are obtained by sequential testing of all attributes and candidate thresholds and comparing their information gain.
In the case that no test can be found with a statistically significant information gain, the node is declared a DEADEND and is not further split.
Quality improvement with Decision Trees
In order to detect if a node is terminal, or sufficiently “class-pure” (i.e. the majority of its MS belong to one of the two classes), the stop splitting rule is used. More specifically:
If the classification entropy of the node is lower than a minimum pre-set value Hmin, then the node is declared LEAF and is not further split.
If not the optimal splitting rule is applied.
Quality improvement with Decision Trees
Quality Rule 1 at node 4
if (ATTR94.3568) and (ATTR11.0862) Acceptable quality.
5 rules in total (equal to the number of terminal nodes).
Avoid leading to node 5 (acceptability index of 34.55%, or only 19 out of 55 measurement sets are acceptable.
Quality improvement with Decision Trees
From DT to Entropy Network
AL14AL25AL37AL48AL49
TL11TL22TL33TL46
Entropy Network (EN) versus the fully connected Multilayer Perceptron (MLP)
The tedious task of network structure optimization (i.e. the finding of the number of hidden layers, and the optimum number of neurons per layer) is avoided.
The training time is significantly reduced (more than a factor of 10), since the EN has fewer connections, and as a result fewer unknown parameters.
If the EN is retrained its classification performance can be improved and to be very close to the performance of the MLP.
The ORing Layer of the EN is replaced by a single output neuron, fully connected to all neurons of the ANDing Layer and the resulted network is called Hybrid DT-NN.
Next, the Hybrid DT-NN is trained again.
Next, the trained Hybrid DT-NN is used to predict the quality of the test set, and, finally, to classify them accordingly, providing the Hybrid DT-NN Classifier.
Quality improvement with Hybrid DT-NN Classifiers
Industrial Application - Design of Wound Cores
10000
12000
14000
16000
18000
20000
0.2000 0.6000 1.0000 1.4000 1.8000 2.2000
Specific Losses (W/Kg)
B (
Gau
ss) DWPKi
DWPKTF
DSFL DWPK DKg ii i i * , " ",...," "11 14
DNLL DWPK DKgTF TF TF *
Transformer Iron Losses
Influenced by both qualitative and quantitative parameters.
There is no simple relationship among the parameters involved in the production process that expresses analytically the transformer iron losses.
Importance of iron loss prediction
Protects the manufacturer from paying loss penalties.
Contributes in economic manufacturing.
Artificial Intelligence
Has the ability to automatically learn relationships between
inputs and outputs
Has been applied to the quality improvement of individual core
and transformer
The quality improvement is achieved through better
classification of iron losses (of core / transformer) prior to the
manufacturing (of core / transformer)
Quality Improvement of Individual Cores
First Industrial Application :
Application on Individual Core - Attributes
Symbol DescriptionATTR1 Annealing final temperatureATTR2 Temperature rising timeATTR3 Furnace opening temperatureATTR4 Duration of constant temperatureATTR5 Position of core in the furnaceATTR6 Protective atmosphereATTR7 Actual over theoretical core weight ratioATTR8 Specific losses of core magnetic material
Application on Individual Core - DT
Avoid node 6
One way is to increase ATTR7 Inrease the actual weight of core by adding more magnetic material so that the ratio of actual over theoretical core weight ratio to become greater than 0.98
Lead the production to node 4 or 6
Application on Individual Core - Entropy Network
Application on Individual Core - Entropy Network and Hybrid DT-NN Classifier
Method Structure Error (%)
DT - 6.0
EN 3-3-4-2 5.4
HDTNNC 3-3-4-1 4.3
Quality Improvement of Transformer
Second Industrial Application :
Application on Transformer - Attributes
Symbol Attribute Name
ATTR1 Ratio of actual over theoretical total iron losses ofthe four individual cores
ATTR2 Ratio of actual over theoretical total weight of thefour individual cores
ATTR3 Magnetic material average specific losses of thefour individual cores (Watt/Kg at 15000 Gauss)
ATTR4 Rated magnetic inductionATTR5 Thickness of core legATTR6 Width of core legATTR7 Height of core windowATTR8 Width of core windowATTR9 Transformer volts per turn
Application on Transformer - DT
Avoid node 5
One way is to reduce ATTR1 Reduce the actual total iron losses of individual cores by removing from the transformer cores set one or more cores with high iron losses and add cores with lower losses
Lead the production to node 7, 8, or 4
Application on Transformer - Entropy Network
Method Structure Error (%)
DT - 4.0
EN 3-4-5-2 3.3
HDTNNC 3-4-5-1 2.2
MLP (9 attributes) 9-5-2 1.4
MLP (3 DT attributes) 3-7-2 3.2
Application on Transformer - Entropy Network and Hybrid DT-NN Classifier
Minimisation of Transformer Iron Losses
Third Industrial Application :
Current grouping process
Pre-measure and assign a grade (quality category) to each individual core and then combine higher and lower graded individual cores to achieve an “average” value for the entire transformer.
Objective of grouping process
To reduce the variation in iron losses of assembled transformers.
L = 6Sm all C ores
O ne p ossib le a rrangem en t o f co resL = 6
L arge C ores
1 0 9 75 2 61 2 8 111 3 4
1 T /Fst 2 T /Fnd 3 T /Frd
“ 11 ” “ 1 2 ” “ 1 3 ” “ 1 4 ”
7 8 9
1 2
54
1 0 11 1 2
3
6
G A In d ividua l
792 611 2 8 111 0 3 45
GA Representation
Two-point crossover operator
55
Paren ts O ffsp ring
C rossover P oin ts
5
4 4
10 8
8 10
12 12
12 12
1 3
3 1
2 2
2 2
9 9
10 8
8 10
11 11
3 1
6 6
6 6
1 3
7 7
7 7
11 11
9 9
4 4
5 5
1. Input transformer design data and measurements on L small cores and L large cores.
2. Assign integer numbers 1 through L to the small cores, and L+1 through 2*L to the large cores.
3. Initialize the population, set the probabilities of crossover and mutation, and define the GA termination criterion.
4. Estimate the fitness (i.e. the total predicted, by the MLP, iron losses).
Optimal Solution Using GAs (1)
Optimal Solution Using GAs (2)
5. Parent selection, crossover, mutation, and creation of the next generation.
6. Estimate the fitness.
7. If the termination criterion is satisfied, then proceed to step 8, else go to step 5.
8. Output the optimal arrangement of cores for all the L/2 transformers (optimal solution), the predicted iron losses of each of the L/2 transformers, and the total predicted iron losses.
9. Assemble the L/2 transformers using the optimal arrangement of cores.
Results (1)
Grouping 100 small and 100 large cores of the same production batch of 50 transformers, 160 kVA, 50 Hz.
Minimum total iron losses of 15664 W, at generation 85.
15500
15750
16000
16250
0 20 40 60 80 100
GA Cycle (Generation)
Tot
al Ir
on L
osse
s (W
)
Results (2)
The evaluation of the genetic algorithm is based on the AARE.
For the specific example, the AARE is 0.57%.
In general, the AARE is reduced by 50% (3.15% 1.6%).
305
310
315
320
325
0 10 20 30 40 50
Transformer
Iron
Los
ses
(W)
Actual
Predicted
Advantages for the transformer customer
Significant contribution in the economic operation of the electric network, due to:
Iron loss reduction, andThe fact that iron losses exist 24 hours per day, 365 days per year, even if the transformer is operating without load.
Customer satisfaction, due to lower losses, and higher economy.
Advantages for the transformer manufacturer
Reduction in iron loss variation, leading to the increase of the reliability of the manufacturer.
Reduction of the cost of materials:Magnetic materials : 2.5% (2001: 2.0 Millions USD)
Copper : 1% (2001: 1.7 Millions USD)Insulating materials: 1% (2001: 0.4 Millions USD)Oil : 2% (2001: 0.4 Millions USD)
Avoid paying loss penalties.
Increase of market share.
Conclusions
The application of the DT method provides rules useful for the quality improvement of the industrial processes
A hybrid DT-NN classifier is proposed for the quality improvement of industrial processes
The method was applied to the quality improvement of indivudual core and transformer and was compared with other methods
A new GA based grouping process was proposed, in order to reduce iron losses