and C4 revealed the following tendencies:
- FIT3A (C4=0) results in large values of C3;
FIT3B (C3=0) also led to high values of C4.
- In the four-parameter learning (FIT4), calculated C3
and C4 were not stable. The cases for C3 and C4
pairs were:
- Low C3 plus high C4. C3
was much lower than the medium value and C4 much higher than the
medium value.
- High C3 plus low C4. The high
value of C3 was much higher than the medium value and the low
value of C4 much lower than the medium.
- Medium C3 plus medium C4. This
was the perfect condition with both C3 and C4 in the
similar range of the medium values. However, the occurence of this case was
low; C3 and C4 were more or less away from the medium
value.
Further studies indicated that for most grades, C3
and C4 may be roughly described in a linear relationship:
C3
= m * C4 + n
(2)
In which m and n are empirical factors derived from the
large number of data, varied from grade to grade. This finding indicated C3
and C4 have great influence on each other. As long as a C3
and a C4 satisfy equation (2), a relatively good fit for learning
should be achieved, even though C3 and C4 might spread in
a large range, from negative value, zero and a very high value. The problem is,
even though each [C3, C4] pair may have a good fit, the C3ALL
produced from combining so many widely scattered C3 values, and C4ALL
from so many totally different C4 values, may not fit each other
well. This leads to the limitation of the long-term learning. As long as each of
qualified C3 and C4 in a pair depends on each other and
thus scatters widely, the long-term learning by combining them would hardly
reach high accuracy. Therefore, a blind long-term learning can reach only a
limited level of accuracy, but cannot go further. Human intervention is
necessary to determine the right C3, C4 value pair.
Discussion from the Data
Analysis
In this project, quite a portion of the
time was spent in an attempt to understand the data in the existing Level 2
system. There were many confusing questions, for example:
Why were the values of C3 and C4
scattered in so big a range? It did not look like a measurement problem.
Why were the C3 from the log
data consistently much higher than the theoretical value?
Different researchers published quite
different results on C3. Who is right?
Further studies were also needed to
identify which temperature region should have higher C3 and C4,
the higher-temperature region or the lower-temperature. The lower-temperature
region did have higher retained strain.
Without an understanding to the problems,
any change to the system would put it at risk. After all those questions were
answered and the problems understood, repairing the system actually was not very
difficult.
Answer for (1) was due to the dependence
of C3 and C4 on each other as discussed in the
Limitation of
Adaptive Learning. Reason for (2) was
considered to be the effect of the retained strain. Both the German and Japanese
colleagues are right and their data are consistent. If the value from the
Japanese colleagues (C3=0.21) is used, the pass strain should be
used; while if the German colleagues?data (C3=0.18) is accepted, the
actual strain (the pass strain plus the retained strain) should be applied.
Besides retained strain, there are other factors that affect the C3
and C4.
<To
Be Continued>
<Previous>
|
<Next>
|
Part1
|
Part2
|
Part3
|
Part4
|
Part5
|
Part6
|
Part7
|
Part8
|
Part9
|